{ "schema_version": 1, "requested_window": { "start": "2026-10-01", "end": "2026-10-07" }, "review_finalized_on": "2026-10-10", "unique_papers": 1022, "counts": { "rejected_or_out_of_scope": 367, "accepted": 494, "reviewed_deferred": 113, "known": 48 }, "candidate_review_complete": true, "full_source_coverage_complete": false, "full_review_watermark": "2026-10-02", "watermark_advanced": false, "retrieval_receipts": [ { "track": "recommendation", "mode": "submission-window", "candidate_count": 35, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 22, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "recommender system", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "6a188c6d36fab92579072a8df0d7d0e334f76d4c52bfc5311195ce505d0b31d4", "next_offset": 18, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "recommendation ranker", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "81c1e34e8c1622e462ff1327259bebe1e11c49e4601ad7e29105c1bb8d952d25", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "generative recommendation", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "b1559fcdbafe1cc409d476a047c8204203095514b886d22f6022141963427035", "next_offset": 25, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "LLM recommendation", "categories": [ "cs.IR", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "93361f5ca669662c809e26eb8bc273f67f6a5030cb6ceae94d96cfcbc8491e65", "next_offset": 8, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "industrial ranking recommendation", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "a3bc57cfc5207d2c3de93e4923412c05081f1b55c7d3c4d3c12b4efeadc2f66e", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "recommendation online A/B", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "09724183f5fd9ecb5ca5338e9bf1809592c60d45b493feb45e9047f9b33e8fa1", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "recommendation deployed production", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "897f6e9e077c85ca2f190cf924e86694da9980ab04de1680c7f2cfa269e1e532", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "search ranking online A/B", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "c0cd774997fef851e44e0f24c0e18cb0180598533ca068d3adf2269b00d8042d", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "advertising ranking online A/B", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "6dd4adf1a80d77d9f29d13badebf33ea368a951e6a757533b03d0329c1f247d4", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Google recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "01fb7ac189d2006ff75df5cc57e795de2946094491096e946cf012f3e833452a", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Google DeepMind recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "b3388157e2554226e78a7335bfeaa8f9cc513dd8a171f3f4482fe1b0d3da872c", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Meta recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "22d79b2b96bb8916a71794cbab738cd7ad21f32777962bf23cdd5d0b8da682d7", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Netflix recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "d53e9bcf90aaab30ef00bb566c5ae40bd33503f61d7ca80d5a3536e2561356fa", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "ByteDance recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "4fbac943940a0eec81727cfd32d80ee1bd8ee2e09b8411289537a2f5585af7f3", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "TikTok recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "996c880f9e080732f0ff93ad125d3d4f4927ca1b72cd065b7b7c329a7090c11c", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Alibaba recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "dd3aefb7a0c451ebd8cd4da7dd7c9247e0b26c7dc692c9054365672f6555c8ce", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Taobao recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "ece5dbf58af664e2b97f66aad2a042ccbdf715e75861b4eb2d481d9eb88a7dee", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Kuaishou recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "55822f391fb05184b505bd5661ed200852f05333378be72a27981ce0b8d6dbc1", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Pinterest recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "f9e22262c53b01f8b481048cbbc45c21fff8dc49f74350ae7915d8b4b93af6ad", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Amazon recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "1fc3fb909c2a0e66cff2ba41b75a927fa6907cb6109300af605b85e9c45a1a21", "next_offset": 4, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Microsoft recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "87d2c21fadf739eefdcf7f5b1d2035e778d370aa74cb80f5d219ee37804c59ce", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Tencent recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "f2099115a15d4ca2bbf21c1fd818de8f", "fingerprint": "0f07e8e67c3302df2cab6814103216d0332af7495e2ab9a65fb96d88c3ca8277", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } }, { "track": "recommendation", "mode": "-month", "candidate_count": 39, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 22, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "recommender system", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "8a100ea9d85db7535548e915b51729a5c70823629b4d136f61962570547a427b", "next_offset": 19, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "recommendation ranker", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "696b74b1e7f71001181fba9e0502a58a37e2bffff56a85538b915b98a12f170a", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "generative recommendation", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "f606491d1ad9cf3efd82e5912c852a4d42c43738f85fe44f713dce54fb2fccf9", "next_offset": 27, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "LLM recommendation", "categories": [ "cs.IR", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "22ea4bd5d5c0e662c80d1ac531dc87d6e3def88225b4f358d9d55a8da33f1c57", "next_offset": 10, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "industrial ranking recommendation", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "e1b1a4e4289b0dae72894f327efa9ce2c53ce93c69792b261982b71dbdb4021e", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "recommendation online A/B", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "7130644679af861f316bf60447d4ff42b64d28c927e1b6dcff308c63f5bcd767", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "recommendation deployed production", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "296abd92f43a860299dfb23a7407250561f3b1b336c335c54455b413460bfc2d", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "search ranking online A/B", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "771276aa89e0256c2cca8f6998351ebac9a3ccd5b014c9a8195d325e4020cd43", "next_offset": 1, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "advertising ranking online A/B", "categories": [ "cs.IR", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "287a7c7ac5b55c33648e4add9bdb2a53ce7e1f8710d1a37f7b3c9194c151a5dc", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Google recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "71c5e84d013c307d1071a0f4cf09e3b8eb5934a764adfa50df7f0e671f1efbd5", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Google DeepMind recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "572110fac31ffa6d4271bb430a866892d28d048d3e4c090f283c8f2c51b8aab9", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Meta recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "974f0f14db4469d0b6f98e00298ba17606fbdc3c03aec1bdd0eeefa238b0ff25", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Netflix recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "a09ff3983330b67fcab7793337ef597c8620959a58a47f05efd6dc452b7825aa", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "ByteDance recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "4638389ed45845f32c4461923563ce800788af99619f536e8884f13df9fb08b5", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "TikTok recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "6e025a645efadc128fa0b77071d3641064b81ecd664247bf0b862fca07690770", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Alibaba recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "0fd6f7b5a0f4e7a0895ffc36e9de53bd37df7c837265fa948e571ed1e197c830", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Taobao recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "a99a6a415b14ce9b8996456c6edace79ecb31be134c85f04fedad24238289ba0", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Kuaishou recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "7b690b48f35a167eba34b4b6f445c987bc5623ddbfd241422504c3684ae08cf6", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Pinterest recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "efdbed911fb74ab4440ee7ad544ca4a4783b82e1718bc1cbe51e1ab4f5db335d", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Amazon recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "abf4c8e9a6daac43742643614964de424276d6906122d1dc59cdfcf96594288e", "next_offset": 4, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Microsoft recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "5625efe1219f8e712faf98fca8fc74e2870528f93c227f9dc0591e7390513f30", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "Tencent recommendation", "categories": [ "cs.IR", "cs.LG", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "e0a8928dc7574f9986add0c547e86a0f", "fingerprint": "86dff282b45857b540f2e66c57af178e77391e3c5a078bba82e23e3431915aab", "next_offset": 0, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } }, { "track": "foundation-model", "mode": "submission-window", "candidate_count": 411, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 8, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "language model architecture", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "7a2a1250b6ee4b08f6d48df06222ddd84b77278dd0d178d4541adf172c97b466", "next_offset": 69, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "transformer architecture", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "4d975faa2614791d2549108f617cacf6baa78f6f94366fce858294bf729b06c5", "next_offset": 54, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "efficient LLM inference", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "151d0f6586cb7279ccda074d8b1c7ca9be582017c0650c2e488e1a065a9ec725", "next_offset": 21, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model long context", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "cca98f1620dfca36c749274d4c30f22085bf79b3ce473e3a7691dbcd76c9a3c9", "next_offset": 38, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model pretraining data", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "e4a15d751c72e375c89b7a4408af6d8822c8550021ad349b5774c7aba11d747b", "next_offset": 29, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "multimodal language model", "categories": [ "cs.CV", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "0ec5800932d852de76bb826a3c97807cabecc10f0c746121cdcb27265c2daa17", "next_offset": 67, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "vision language model", "categories": [ "cs.CV", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "16a25f1c418641e05455fae956535b421f5bc0e3d2bb9fb3899376bdd593ad06", "next_offset": 198, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 2, "resumed": false, "error": null }, { "query": "language model compression", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "e7220e8756a1499493aa8bf48752ca17", "fingerprint": "818564fefd154664c01ba359fd4729d3e5fa68c106ef3d960b14c02bde9a86af", "next_offset": 45, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } }, { "track": "foundation-model", "mode": "-month", "candidate_count": 459, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 8, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "language model architecture", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "87d55f5aee9ec473bb7391537cfda62caf98bbf5bbbb22971d018719835081ee", "next_offset": 79, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "transformer architecture", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "9914353c120a10921c884ba1b440f5db7bee123875e1bd4c4e1281ef065f83dc", "next_offset": 62, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "efficient LLM inference", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "65e118e262e13a01c9c81a83407adfb776cef6dbd30ee1d0b9e67cf27f1e9cab", "next_offset": 27, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model long context", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "79d664a3ac7e666056d2aecc38a0dc74f5643f8f2d0860ca7a06a6491494955c", "next_offset": 41, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model pretraining data", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "3ee9c43b152998b1f91e44aa4d2930eaf35f566a9d628f7f8403361acc5164c3", "next_offset": 32, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "multimodal language model", "categories": [ "cs.CV", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "bf7be4c59fa6cfc1e206f385af788c97837546f97f23db4df7b94bae851e3aae", "next_offset": 77, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "vision language model", "categories": [ "cs.CV", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "4732160e858b3e2f7a14b5ec261aed56681f8d29be0b55966769bdfb392e72c4", "next_offset": 219, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 3, "resumed": false, "error": null }, { "query": "language model compression", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "b9bb5356604b4ee59b69bab3aea69427", "fingerprint": "519e5c15144ed2dfdbb2cb46ddbd9f637512b43e3be11b023c6b78aa7c547283", "next_offset": 50, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } }, { "track": "post-training", "mode": "submission-window", "candidate_count": 177, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 8, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "language model post training", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "35bf12348d3ea785e1a7f335f284310edfde9d20d5b70c9a947e40abd898d861", "next_offset": 74, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model preference optimization", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "2784207c6676e9da36bd43f586f5a29954c516a82568ba887c3965c24129694d", "next_offset": 9, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "reinforcement learning language model", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "c045c5e27bba1d2234504b194144adc727aa94d36829dd6c63e904b36914a0b8", "next_offset": 86, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "on policy distillation language model", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "33e318e801d3aae44387e314f3455f1ea1c5ca9defbadb7d9ba5835c3fc7fa10", "next_offset": 25, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model reward model", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "769b497279081dbe51fd1afa7f3a5bd93e8efe33581555f1caa17f3ddacdb390", "next_offset": 74, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model process reward", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "a7d2fd799a9393888719e5bf3951fc00472eeb8c0661f92871497cedd9bbe9ca", "next_offset": 3, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "test time reinforcement learning", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "ef584d2c32d9ad7ea4850750569533bc16293ce15e79dfbd7deedafe50abab40", "next_offset": 19, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "online policy distillation", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "8b577b9b125f4bb6a72381a34bc675a2", "fingerprint": "89a78e59d1e5f7c19c027d06a0982c37ec2ed0a9d091cfec16f83c0568661013", "next_offset": 3, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } }, { "track": "post-training", "mode": "-month", "candidate_count": 201, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 8, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "language model post training", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "fcae38dcd0473d87145e9683f3fdf712ab98c8637d1e63d62684c53a1f80fa12", "next_offset": 84, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model preference optimization", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "aeca04c783ff7f7ff6dc7843f7922de6a5b9ddcfdc5e1ffc8a7cf7a60e4dbd42", "next_offset": 11, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "reinforcement learning language model", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "f7bdec5a5526b2df13b3a282abbab7fb3d0621722216f49ff174deb12aadf6ae", "next_offset": 96, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "on policy distillation language model", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "b91efae1ca44b48b70f48ba2ef54e68e55c0ba5d5987c38931f2837be65d1394", "next_offset": 29, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model reward model", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "7296516545d92694c859985bf4f32049a8e44094023f3c9ec529f4ce86db4bff", "next_offset": 83, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model process reward", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "24b9e60582b955e3c40dfa2e9d066b7bc06a55466de009ad17de9615198dd582", "next_offset": 4, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "test time reinforcement learning", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "1d861c71b14ee471d72ad78bcc2bbefd2737aaf974ce8f8da6eabf215bce6d80", "next_offset": 20, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "online policy distillation", "categories": [ "cs.CL", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "9c0838b1e33f49c4867899ea0f0c50eb", "fingerprint": "afd3bb0cf2057e408c6ccd3cadf1b0eca35833c846d9832ceac2bb0e446efbd9", "next_offset": 4, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } }, { "track": "agent", "mode": "submission-window", "candidate_count": 396, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 8, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "LLM agent", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "cf9ef62a8a8ae20064821cf1b94b73b8fb8a3a114e5e04f2b4dea4f0fa043faf", "next_offset": 234, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 3, "resumed": false, "error": null }, { "query": "agentic reinforcement learning", "categories": [ "cs.AI", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "75f558331c73e0c3d6d79df50fbffdcbbc91acb8760e118856ffbbc6d282b72b", "next_offset": 80, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model tool use agent", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "6c5367420698aad35dd5eef05d00ecd63290ceabcde3da6c360fe738bbc19618", "next_offset": 57, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "web agent language model", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "f2cd15f19fb196e77c837e477376e639f2c990edcde8c3f1a1de776928ab0c50", "next_offset": 16, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "software engineering agent", "categories": [ "cs.SE", "cs.AI" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "fe95274c2db5bafb0d51dbd1da2698f5c233728d567976d2c0dcb7a188edb417", "next_offset": 71, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model agent memory", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "1150d855a4cd711f07c7563fa861fe12f98f1e20a8f592a4c45e50424c042601", "next_offset": 42, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model agent planning", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "88649bde17dbde6da31f1cd80084d246ed3bf9c696ebe6486804659f122fb338", "next_offset": 27, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "large language model multi agent", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "2026-10-01", "date_to": "2026-10-07", "scan_id": "3adcb53d958c4324ad39a8cabca119fc", "fingerprint": "e82769c76a1c8bfc845f269d6b5e9085eba151c73e523c13f099b3ceedd7b24b", "next_offset": 50, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } }, { "track": "agent", "mode": "-month", "candidate_count": 437, "transport": { "coverage_complete": true, "transport_complete": true, "query_count": 8, "capped_queries": 0, "failed_queries": 0, "cache_fallback_pages": 0, "queries": [ { "query": "LLM agent", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "a3a183a8eab40f2c4d6c1409e14e7e1ea0ddaedb5dcabde0aa61332cae2079e1", "next_offset": 260, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 3, "resumed": false, "error": null }, { "query": "agentic reinforcement learning", "categories": [ "cs.AI", "cs.LG" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "590f48d0af09d834044cbc8facefea392f9d5dbc8bdd5b1c3754e1bdf3119917", "next_offset": 91, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model tool use agent", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "7232b49947b02d20d3371b6f433a46b59c92cf252a4f9161f545470b7586a41b", "next_offset": 62, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "web agent language model", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "14f92e0f7ae271931e3e830de5c1a06f6c01aadc0b1988322e4dc4c6495660af", "next_offset": 17, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "software engineering agent", "categories": [ "cs.SE", "cs.AI" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "6c234e59e22fdef9bf667c478c544267dd495e32d9599d18ced55948bd87ca56", "next_offset": 77, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model agent memory", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "ef49d1cf983e7cf38fc6498aafc0d910af170f2d25db70bded71c26505d9b7b4", "next_offset": 46, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "language model agent planning", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "8b29a834fa61cb5baf8f40ad11c4812afc6a9dbad275627dec569a180fd566e1", "next_offset": 31, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null }, { "query": "large language model multi agent", "categories": [ "cs.AI", "cs.CL" ], "match": "all", "page_size": 100, "maximum_results": 2000, "date_from": "None", "date_to": "None", "identifier_months": [ "2610" ], "scan_id": "bcb6885e3ca44365b3104b897585472e", "fingerprint": "9d9bba905a040c3edd4ccbd9383fb6e38ce40eed10e0d5bd268f47aac3398d81", "next_offset": 54, "state": "exhausted", "coverage_complete": true, "transport_complete": true, "cache_fallback_pages": 0, "pages": 1, "resumed": false, "error": null } ], "scope": "configured arXiv API query matrix only; candidate date/identifier-month filtering is separate", "review_complete": false, "watermark_eligible": false, "note": "API retrieval coverage does not certify full-text review or all-source coverage; no review watermark is advanced automatically", "cache_dir": ".cache/paper-discovery/arxiv" } } ], "papers": [ { "id": "2610.08791", "title": "World Models' Last Exam in Physics", "published": "2026-10-06T17:59:56Z", "reviews": [ { "track": "foundation-model", "id": "2610.08791", "title": "World Models' Last Exam in Physics", "published": "2026-10-06T17:59:56Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "40项物理关系测量任务覆盖九类现象，官方公开任务图像/提示/评估器及外部视频批评估入口，可纳入evolve公共评测候选；不是新世界模型算法，也不是已接入。", "review_level": "fulltext", "kind": "evaluation_infrastructure", "official_url": "https://arxiv.org/html/2610.08791v1", "code_url": "https://github.com/Einsia/phys-last-exam", "locations": [ "§2.1–2.4/Eq1–7", "§3.1–3.3/Table2", "official README", "docs/evaluation.md", "easy/P21/evaluator/_shared/unified_evaluators/runtime.py:339–355", "evaluate.py:summarize" ], "evidence": [ { "url": "https://arxiv.org/abs/2610.08791", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08791v1", "location": "§2.1–2.4/Eq1–7; §3.1–3.3/Table2; official README; docs/evaluation.md; easy/P21/evaluator/_shared/unified_evaluators/runtime.py:339–355; evaluate.py:summarize" }, { "url": "https://arxiv.org/html/2610.08791v1", "location": "§2.1–2.4/Eq1–7; §3.1–3.3/Table2; official README; docs/evaluation.md; easy/P21/evaluator/_shared/unified_evaluators/runtime.py:339–355; evaluate.py:summarize", "notes": "预设物理残差评分，VLM仅负责可观测性门；40task×8model×4sample，human排名51.58%对43.40% VLM，pairwise53.12%对45.62%。" } ], "limitations": "官方网页404但GitHub可访问。已确认论文与代码差异：论文门失败仍尝试物理测量，公开runtime gate失败直接return/physics_attempted=False；接入须保留原论文测量协议并分别报告coverage，不可把现有脚本直接宣称保真实现。难度标签基于被评模型均分事后划分；human一致性仅约53%，不是物理真值oracle。", "implementation_status": "not_implemented", "abstract_review_completed": true, "fulltext_review_completed": true, "fulltext_notes": "预设物理残差评分，VLM仅负责可观测性门；40task×8model×4sample，human排名51.58%对43.40% VLM，pairwise53.12%对45.62%。", "review_status": "reviewed" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08781", "title": "IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas", "published": null, "reviews": [ { "track": "post-training", "id": "2610.08781", "title": "IdeaAnchor: Teaching LLMs to Turn Literature into Research Ideas", "repository_status": "new", "decision": "out_of_scope", "reason": "正文为anchor数据/特权监督范式，优化器仍标准SFT、self-training和GRPO；可作为Agent科研数据方法，不纳入核心优化器实现。 该决定仅限后训练核心track，Agent方法/数据公共入口由Agent track独立判断，不能据此全局拒绝。", "review_level": "fulltext", "evidence": [ { "url": "https://arxiv.org/html/2610.08781v1", "location": "§3.1式1–3；§4.1–4.2" } ], "fulltext_review_completed": true, "fulltext_scope": "针对方法、实验协议与主要结果的正文定向核查；非逐字通读全部附录", "pool": "main", "review_status": "reviewed" } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08779", "title": "ALIVE: Interaction-Aligned Object Insertion for First-Frame-Guided Video Editing", "published": "2026-10-06T17:58:51Z", "reviews": [ { "track": "foundation-model", "id": "2610.08779", "title": "ALIVE: Interaction-Aligned Object Insertion for First-Frame-Guided Video Editing", "published": "2026-10-06T17:58:51Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "ALIVE以source+noisy target通道拼接、干净首帧和按时间overlap/描述长度归一的cross-attention；Qwen VLM从同输入预测P4而不是读target。35,800对训练与两benchmark；部分基线额外P2提示不完全同条件。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08779", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08779v1", "location": "§3；§4.2式2–4、§4.3式5；§5.1" }, { "url": "https://arxiv.org/html/2610.08779v1", "location": "§3；§4.2式2–4、§4.3式5；§5.1", "notes": "ALIVE以source+noisy target通道拼接、干净首帧和按时间overlap/描述长度归一的cross-attention；Qwen VLM从同输入预测P4而不是读target。35,800对训练与两benchmark；部分基线额外P2提示不完全同条件。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08779v1", "locations": [ "§3；§4.2式2–4、§4.3式5；§5.1" ], "implementation_status": "not_implemented", "fulltext_notes": "ALIVE以source+noisy target通道拼接、干净首帧和按时间overlap/描述长度归一的cross-attention；Qwen VLM从同输入预测P4而不是读target。35,800对训练与两benchmark；部分基线额外P2提示不完全同条件。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08778", "title": "Sherpa: Teaching LLMs to Teach Adaptively", "published": null, "reviews": [ { "track": "post-training", "id": "2610.08778", "title": "Sherpa: Teaching LLMs to Teach Adaptively", "repository_status": "new", "decision": "accepted", "reason": "逐回合gate掩码return并按到达同回合的轨迹leave-one-out归一，训练teacher适应学生偏好；有MATH分离训练/评测及教学benchmark。", "review_level": "fulltext", "evidence": [ { "url": "https://arxiv.org/html/2610.08778v1", "location": "§2.3式3；§3.1、§3.4" } ], "fulltext_review_completed": true, "fulltext_scope": "针对方法、实验协议与主要结果的正文定向核查；非逐字通读全部附录", "pool": "main", "review_status": "reviewed" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08775", "title": "Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?", "published": "2026-10-06T17:57:19Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08775", "title": "Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?", "published": "2026-10-06T17:57:19Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08775v1", "location": "§3.1–3.5 Eq1–6；§4.1；Ethics/Reproducibility；官方README与src/bottled/datasets/esci.py", "decision": "accepted", "reason": "BOTTLED作为evolve可执行评测基础设施接受：Agent在预算内生成可复用artifact并完成大批无标签任务，明确隐藏gold与预算约束，兼容模型/脚本/路由器候选。", "reviewer": "agent_review", "evidence": "MAVE/ESCI/RAID三任务；5M加权tokens、10h、A10040G；官方代码存在agent_eval.py/预算MCP/proxy/数据与评分器；ESCI源码拆questions/gold并确定fallback后计算macroF1。", "limitations": "原实验每模型每任务两runs，零样本只抽1000；GPU成本未统一入token预算；MIT代码，MAVE CC-BY-NC4.0/ESCI Apache2.0/RAID MIT；contamination judge及Jev评测未公开，不能宣称完整污染审计复现。", "code": "https://github.com/aktsonthalia/bottled", "license": "MIT code; dataset-specific licenses", "candidate_kind": "evaluation_infrastructure", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08773", "title": "AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model", "published": null, "reviews": [ { "track": "post-training", "id": "2610.08773", "title": "AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model", "repository_status": "new", "decision": "accepted", "reason": "冻结world model与judge，分阶段课程—执行者、对手—执行者训练；三seed rollout与150任务对照，不等于真实生产A/B。", "review_level": "fulltext", "evidence": [ { "url": "https://arxiv.org/html/2610.08773v1", "location": "§4 Algorithm 1；§5 Table 1" } ], "fulltext_review_completed": true, "fulltext_scope": "针对方法、实验协议与主要结果的正文定向核查；非逐字通读全部附录", "pool": "main", "review_status": "reviewed" }, { "track": "agent", "arxiv_id": "2610.08773", "title": "AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model", "published": "2026-10-06", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08773v1", "location": "§3 Environment/Threat Model/Metrics；§4 Algorithm1/Eq2–5；§5 Tables1–3；§6 Limitations", "decision": "accepted", "reason": "AdvSim2Real的三策略/两阶段训练有明确算法：课程以p≈0.5任务奖励，adversary仅获clean-success到failure flip奖励，executor fresh-rollout训练且重用历史attack inputs；可作为Agent RL核心候选。", "reviewer": "agent_review", "evidence": "表1 base→Robust3 clean 74.89→81.33%、learned-adversary mean 48.07→57.48%；表3 unseen Kimi-K3 23.00→30.72%；表2 Stage1 clean Chromium strict 25.56→44.44%。", "limitations": "攻击结果只world-model+LLM judge；Stage2未在真实浏览器测试，不能声称可执行安全；150自建form任务，3 rollout seeds非3训练seeds；攻击后可行性未经核验；去Stage1实验不compute matched，鲁棒性收益仅0.44pp。", "institution": "MBZUAI; Amazon; MIT", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08720", "title": "WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation?", "published": "2026-10-06T17:26:26Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08720", "title": "WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation?", "published": "2026-10-06T17:26:26Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08720v1", "location": "§2.1–2.3 Eq1–10；§3.1–3.3/Table3；Limitations；官方仓库/API访问2026-10-07", "decision": "pending", "reason": "WorldSolver的168物理求解器任务/固定scaffold/执行+VLM视觉+物理规则评分适合evolve，但官方链接github.com/sirujiang/WorldSolver网页与API均404，暂不能确认公开任务/评分器/许可，等待可执行发布。", "reviewer": "agent_review", "evidence": "7物理域；生成1800秒/solver600秒；执行失败记0，成功取(V+P)/2；GPT5.6Sol48.7、ClaudeOpus546.7。", "limitations": "60分pass阈值仅演示非验证标准；评分规则对Agent隐藏；不同模型用不同harness；VLM judge也参与被测模型，需独立审核；代码不可达不是算法无价值。", "code": "https://github.com/sirujiang/WorldSolver", "pending_reason": "public_artifacts_unavailable", "candidate_kind": "evaluation_infrastructure", "review_status": "reviewed_deferred", "fulltext_review_completed": true, "implementation_status": "not_implemented_in_this_review" } ], "status": "reviewed_deferred", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08718", "title": "When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting", "published": "2026-10-06T17:25:48Z", "reviews": [ { "track": "foundation-model", "id": "2610.08718", "title": "When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting", "published": "2026-10-06T17:25:48Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "新事实微调造成共享表征偏移与逐事实侵蚀分离；对更新矩阵删首奇异分量的干预在OLMo2-1B有同范数随机rank1对照，可登记机制诊断候选，不能宣传为通用抗遗忘能力算法。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08718", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08718v1", "location": "§4.1–4.4; §5.1–5.2/Figure5; §6/Limitations; Appendix D.2" }, { "url": "https://arxiv.org/html/2610.08718v1", "location": "§4.1–4.4; §5.1–5.2/Figure5; §6/Limitations; Appendix D.2", "notes": "三seed、CounterFact初始已知旧事实；合成新实体训练恢复但真实EntityQuestions制造商不恢复；删主方向可还原旧事实但损新事实，随机方向不能。" } ], "kind": "mechanism_diagnostic", "official_url": "https://arxiv.org/html/2610.08718v1", "locations": [ "§4.1–4.4", "§5.1–5.2/Figure5", "§6/Limitations", "Appendix D.2" ], "limitations": "人为新旧答案首token不重叠，单预训练模型；诊断logit均值来自已知oldfact不能在线oracle使用。只采用权重干预时须独立validation选checkpoint和rank。", "implementation_status": "not_implemented", "fulltext_notes": "三seed、CounterFact初始已知旧事实；合成新实体训练恢复但真实EntityQuestions制造商不恢复；删主方向可还原旧事实但损新事实，随机方向不能。", "review_status": "reviewed" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08713", "title": "SpaTime: Streaming Vision-Language Models for Spatio-temporal Reasoning", "published": "2026-10-06T17:22:33Z", "reviews": [ { "track": "foundation-model", "id": "2610.08713", "title": "SpaTime: Streaming Vision-Language Models for Spatio-temporal Reasoning", "published": "2026-10-06T17:22:33Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "SpaTime冻结因果几何/视觉编码器，2×2 merge投影相加，训练LoRA/state/head并建模首个Response分布作时间损失；strict只计最早回答，charitable末帧强制回答须分开。StreamVSTI训练、StreamVSI任务OOD。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08713", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08713v1", "location": "§4.1式5；§4.2式6–8；§5；§6" }, { "url": "https://arxiv.org/html/2610.08713v1", "location": "§4.1式5；§4.2式6–8；§5；§6", "notes": "SpaTime冻结因果几何/视觉编码器，2×2 merge投影相加，训练LoRA/state/head并建模首个Response分布作时间损失；strict只计最早回答，charitable末帧强制回答须分开。StreamVSTI训练、StreamVSI任务OOD。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08713v1", "locations": [ "§4.1式5；§4.2式6–8；§5；§6" ], "implementation_status": "not_implemented", "fulltext_notes": "SpaTime冻结因果几何/视觉编码器，2×2 merge投影相加，训练LoRA/state/head并建模首个Response分布作时间损失；strict只计最早回答，charitable末帧强制回答须分开。StreamVSTI训练、StreamVSI任务OOD。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08691", "title": "ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences", "published": "2026-10-06T17:08:03Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08691", "title": "ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences", "published": "2026-10-06T17:08:03Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08691v1", "location": "§3 Eq1–3；§4.1–4.2；§5.1–5.2 Eq5–13；§6 Tables1–2/Fig6；§8", "decision": "accepted", "reason": "ScienceClaw冻结模型，以typed workflow执行修复提取linked Skill–Operator bundles，boundary/source reset replay必须实际使用新版本，只有独立validation严格提升且hard constraints/预算通过才原子晋级；符合真实evolve闭环候选。", "reviewer": "agent_review", "evidence": "23学科每学科64 IID+64 OOD；7轮MacroSR OOD77.72→91.30；70/161候选晋级；OOD不参与提案/选择；去链接/去replay/独立validator均有消融。", "limitations": "主要final snapshots，作者§8承认广泛结论需重复runs和学科intervals；domain evaluator可能漏科学错误；需实际科学工具/独立verifier，不能固定轨迹fixture作为能力复现。", "code": "https://github.com/beita6969/ScienceClaw", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08678", "title": "Secure Speculative Decoding for Large Language Models", "published": "2026-10-06T16:59:37Z", "reviews": [ { "track": "foundation-model", "id": "2610.08678", "title": "Secure Speculative Decoding for Large Language Models", "published": "2026-10-06T16:59:37Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "SecureSD 按全局生成位置混合标准/宽松接受与拒绝恢复分布，早期严格验证，后续恢复快速验证；定义完整，可作为真实投机解码候选。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08678", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08678v1", "location": "§IV-B; §V Algorithm 2 / Eq7–14; §VII-A paired protocol; §VIII Tables XII–XIII" }, { "url": "https://arxiv.org/html/2610.08678v1", "location": "§IV-B; §V Algorithm 2 / Eq7–14; §VII-A paired protocol; §VIII Tables XII–XIII", "notes": "" } ], "kind": "inference_algorithm", "locations": [ "§IV-B", "§V Algorithm 2 / Eq7–14", "§VII-A paired protocol", "§VIII Tables XII–XIII" ], "limitations": "不是对任意攻击的安全保证；需目标/草稿模型、lossless/relaxed/corrected 三臂同参数对照，并把任务质量和真实延迟分开报告。原文主要 Blackwell 实测，不代表 A100 已验证。", "code_url": "https://github.com/YichiCS/Secure-Speculative-Decoding", "code_status": "official repository accessible; MIT; bench/securesd/scripts present; some judge/model checkpoints gated", "official_url": "https://arxiv.org/html/2610.08678v1", "review_provenance": "root 独立正文审查；foundation读取整合", "implementation_status": "not_implemented", "fulltext_notes": "", "review_status": "reviewed" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08674", "title": "EC-RAG: Event Chain Retrieval-Augmented Generation for Long Video Understanding", "published": "2026-10-06T16:57:46Z", "reviews": [ { "track": "foundation-model", "id": "2610.08674", "title": "EC-RAG: Event Chain Retrieval-Augmented Generation for Long Video Understanding", "published": "2026-10-06T16:57:46Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "EC-RAG训练外query模态拆解、CLIP variance取帧、ASR/OCR/DET事件描述与按时间相邻链，Contriever相关事件/邻接上下文辅助回答；三长视频benchmark，不是学习因果事件图，额外工具/token预算须计。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08674", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08674v1", "location": "§3.1–3.4式12–16；§4.1–4.2" }, { "url": "https://arxiv.org/html/2610.08674v1", "location": "§3.1–3.4式12–16；§4.1–4.2", "notes": "EC-RAG训练外query模态拆解、CLIP variance取帧、ASR/OCR/DET事件描述与按时间相邻链，Contriever相关事件/邻接上下文辅助回答；三长视频benchmark，不是学习因果事件图，额外工具/token预算须计。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08674v1", "locations": [ "§3.1–3.4式12–16；§4.1–4.2" ], "implementation_status": "not_implemented", "fulltext_notes": "EC-RAG训练外query模态拆解、CLIP variance取帧、ASR/OCR/DET事件描述与按时间相邻链，Contriever相关事件/邻接上下文辅助回答；三长视频benchmark，不是学习因果事件图，额外工具/token预算须计。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08670", "title": "Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment", "published": null, "reviews": [ { "track": "post-training", "id": "2610.08670", "title": "Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment", "repository_status": "new", "decision": "accepted", "reason": "公共评测：判断—行动差距的筛选、四级测量校准与压力对照独立于训练算法；Apache-2.0 DeepSteer仓库可访问。仅74/208主情境通过screen，报告条件分布不能外推总体。", "review_level": "fulltext", "evidence": [ { "url": "https://arxiv.org/html/2610.08670v1", "location": "§3.1–3.5；Appendix F" }, { "url": "https://github.com/deepsteer/deepsteer/", "location": "作者仓库README/文件列表已只读核查" } ], "fulltext_review_completed": true, "core_algorithm_decision": "rejected", "core_algorithm_reason": "预注册价值判断—行动差距评测，比较现有后训练配方；摘要未提出新的训练算法。", "fulltext_scope": "针对方法、实验协议、资源声明的正文定向核查；未执行作者代码", "candidate_kind": "evaluation_infrastructure", "pool": "main", "review_status": "reviewed" }, { "track": "agent", "arxiv_id": "2610.08670", "title": "Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment", "published": "2026-10-06T16:52:23Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08670v1", "location": "§3.1–3.5；Appendix F；作者仓库README/文件列表已只读核查", "decision": "accepted", "reason": "公共评测：判断—行动差距的筛选、四级测量校准与压力对照独立于训练算法；Apache-2.0 DeepSteer仓库可访问。仅74/208主情境通过screen，报告条件分布不能外推总体。", "id": "2610.08670", "repository_status": "new", "evidence": [ { "url": "https://arxiv.org/html/2610.08670v1", "location": "§3.1–3.5；Appendix F" }, { "url": "https://github.com/deepsteer/deepsteer/", "location": "作者仓库README/文件列表已只读核查" } ], "fulltext_review_completed": true, "core_algorithm_decision": "rejected", "core_algorithm_reason": "预注册价值判断—行动差距评测，比较现有后训练配方；摘要未提出新的训练算法。", "fulltext_scope": "针对方法、实验协议、资源声明的正文定向核查；未执行作者代码", "candidate_kind": "evaluation_infrastructure", "pool": "main", "reviewer": "post_training_review_reused", "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08668", "title": "Semantic Behavioral Watermarking: Paraphrase-Robust and Forgery-Resistant Provenance for LLM Agents", "published": "2026-10-06T16:50:42Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08668", "title": "Semantic Behavioral Watermarking: Paraphrase-Robust and Forgery-Resistant Provenance for LLM Agents", "published": "2026-10-06T16:50:42Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08668v1", "location": "§3 threat model；§4.1–4.3；§5；§6.1/6.3/6.4 Tables2/4/5；§7", "decision": "out_of_scope", "reason": "SBW的语义聚类/历史keyed PRF/弃权margin为Agent行为溯源水印算法，主要优化归因检测而非Agent规划、学习、工具使用或任务能力，排除本轮核心能力收录。", "reviewer": "agent_review", "evidence": "ToolBench paraphrase TPR 0.49–0.66；ALFWorld 0.92–0.97；训练无关encoder+kmeans+permutation detector。", "limitations": "§7明示ToolBench无环境执行反馈；复制/链式replay未解决；utility只是动作与argmax一致率，不是任务成功率；部分鲁棒审计artifacts未保留。", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08662", "title": "ParanoiaEval: Benchmarking Unnecessary Defensive Work in Agentic Coding", "published": "2026-10-06T16:46:14Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08662", "title": "ParanoiaEval: Benchmarking Unnecessary Defensive Work in Agentic Coding", "published": "2026-10-06T16:46:14Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08662v1", "location": "§4.1–4.3；§5；§7 Code/Data；官方README、scripts/judge.py", "decision": "accepted", "reason": "ParanoiaEval作为evolve行为约束评测接受：200同任务E+/E−证据对，任务oracle和多余风险处置judge分离，可同时评测完成度与证据响应。", "reviewer": "agent_review", "evidence": "50固定repo、400variants；oracle检验baseline失败/reference成功（28read-only例外）；冻结64人标memory、跨task检索；heldout200判决accuracy.965/kappa.93；MIT任务/runner/judge公开。", "limitations": "公开包无结果/完整轨迹；answer cards/reference补丁只能独立scorer访问；模型native harness混杂不能声称模型净能力差异；developer study限成功runs且非部署A/B；默认same-task=none，不能打开other-arm泄漏gold。", "code": "https://github.com/ZhuoningXu/ParanoiaEval_release", "license": "MIT", "candidate_kind": "evaluation_infrastructure", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08659", "title": "Selective Transfer of RL Updates for Visual Reasoning", "published": "2026-10-06T16:43:55Z", "reviews": [ { "track": "foundation-model", "id": "2610.08659", "title": "Selective Transfer of RL Updates for Visual Reasoning", "published": "2026-10-06T16:43:55Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "Selective-RL隔离R−B与B−V，逐矩阵rank1 SVD后恢复原Frobenius norm，分别重构donor插值和直接加RLdelta；3个8B VLM骨干，源验证选donor，H1验证选λ后冻结，需保留非视觉模块支持对齐。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08659", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08659v1", "location": "§3.1–3.3式1–3；§4.1；Appendix A Algorithm 1" }, { "url": "https://arxiv.org/html/2610.08659v1", "location": "§3.1–3.3式1–3；§4.1；Appendix A Algorithm 1", "notes": "Selective-RL隔离R−B与B−V，逐矩阵rank1 SVD后恢复原Frobenius norm，分别重构donor插值和直接加RLdelta；3个8B VLM骨干，源验证选donor，H1验证选λ后冻结，需保留非视觉模块支持对齐。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08659v1", "locations": [ "§3.1–3.3式1–3；§4.1；Appendix A Algorithm 1" ], "implementation_status": "not_implemented", "fulltext_notes": "Selective-RL隔离R−B与B−V，逐矩阵rank1 SVD后恢复原Frobenius norm，分别重构donor插值和直接加RLdelta；3个8B VLM骨干，源验证选donor，H1验证选λ后冻结，需保留非视觉模块支持对齐。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08651", "title": "A Case Study in Assuring AI-Written Software", "published": "2026-10-06T16:37:08Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08651", "title": "A Case Study in Assuring AI-Written Software", "published": "2026-10-06T16:37:08Z", "source_group": "original", "review_level": "abstract", "source": "https://arxiv.org/abs/2610.08651", "location": "Title and complete abstract", "decision": "out_of_scope", "reason": "医疗软件治理案例研究，无独立Agent算法贡献。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed" } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08649", "title": "Stable Scores, Unstable Answers: Frame Phase and Option Order in Video Multiple-Choice Evaluation", "published": "2026-10-06T16:36:06Z", "reviews": [ { "track": "foundation-model", "id": "2610.08649", "title": "Stable Scores, Unstable Answers: Frame Phase and Option Order in Video Multiple-Choice Evaluation", "published": "2026-10-06T16:36:06Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "PhaseFusion把3×8帧相位后验映射到内容空间作算术平均，独立于option排序控制；n442与32帧准确率差0.45pp不显著，origin flip率下降但order flip未解决；串行3.18s高于24帧2.86s，不能用attention单位冒充实测提速。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08649", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08649v1", "location": "§3.2式2；§4 Table 1；§4.3" }, { "url": "https://arxiv.org/html/2610.08649v1", "location": "§3.2式2；§4 Table 1；§4.3", "notes": "PhaseFusion把3×8帧相位后验映射到内容空间作算术平均，独立于option排序控制；n442与32帧准确率差0.45pp不显著，origin flip率下降但order flip未解决；串行3.18s高于24帧2.86s，不能用attention单位冒充实测提速。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08649v1", "locations": [ "§3.2式2；§4 Table 1；§4.3" ], "implementation_status": "not_implemented", "fulltext_notes": "PhaseFusion把3×8帧相位后验映射到内容空间作算术平均，独立于option排序控制；n442与32帧准确率差0.45pp不显著，origin flip率下降但order flip未解决；串行3.18s高于24帧2.86s，不能用attention单位冒充实测提速。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08647", "title": "SquidAgent: Parallelize Wisely, Coordinate Efficiently", "published": "2026-10-06T16:35:28Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08647", "title": "SquidAgent: Parallelize Wisely, Coordinate Efficiently", "published": "2026-10-06T16:35:28Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08647v1", "location": "§3.1–3.3 Eq1–10/Algorithm1；§4.1–4.6；Tables1–4正文讨论", "decision": "accepted", "reason": "SquidAgent一次DAG规划同时估算输出token/层对齐成本，session fork继承上下文、预写约定，以sum(tokens)/(max(tokens)+alignment)>alpha决定并行；明确可执行Agent调度机制。", "reviewer": "agent_review", "evidence": "同Sonnet4.6九任务吞吐17.2→38.1词/秒，质量rubric 97.1→98.2%；4任务去调度吞吐42.93→29.14；59子任务token预测Spearman0.77、时间预测0.16。", "limitations": "主要自建九任务、LLM rubric；token代理适用于生成主导场景不捕获tool/API时延；alpha1.9基于MathRef sweep设置，独立validation选择未建立；本地复现需真实LLM调用/成本不能fixture代替。", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08639", "title": "Forensic Reserve: Eliciting Latent Knowledge for Image Forgery Detection", "published": "2026-10-06T16:33:00Z", "reviews": [ { "track": "foundation-model", "id": "2610.08639", "title": "Forensic Reserve: Eliciting Latent Knowledge for Image Forgery Detection", "published": "2026-10-06T16:33:00Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "RGE用冻结DINOv3的site-wise ICA及real/fake discovery定位reserve、固定输出basis U，只训练零初始化A的残差UA h；classifier/backbone冻结，benchmark不参与discovery。可复用视觉adapter方法，当前证据仍image forensic专门任务。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08639", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08639v1", "location": "§3.1；§3.2式3；§4.1" }, { "url": "https://arxiv.org/html/2610.08639v1", "location": "§3.1；§3.2式3；§4.1", "notes": "RGE用冻结DINOv3的site-wise ICA及real/fake discovery定位reserve、固定输出basis U，只训练零初始化A的残差UA h；classifier/backbone冻结，benchmark不参与discovery。可复用视觉adapter方法，当前证据仍image forensic专门任务。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08639v1", "locations": [ "§3.1；§3.2式3；§4.1" ], "implementation_status": "not_implemented", "fulltext_notes": "RGE用冻结DINOv3的site-wise ICA及real/fake discovery定位reserve、固定输出basis U，只训练零初始化A的残差UA h；classifier/backbone冻结，benchmark不参与discovery。可复用视觉adapter方法，当前证据仍image forensic专门任务。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08630", "title": "Towards In-Parameter Memory Augmentation for Large Language Models", "published": "2026-10-06T16:27:47Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08630", "title": "Towards In-Parameter Memory Augmentation for Large Language Models", "published": "2026-10-06T16:27:47Z", "source_group": "original", "review_level": "abstract", "source": "https://arxiv.org/abs/2610.08630", "location": "Title and complete abstract", "decision": "out_of_scope", "reason": "参数记忆综述，无新核心算法。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed" } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08622", "title": "Agentic RCA for Internet-Scale Services Using Constrained Creativity", "published": "2026-10-06T16:21:57Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08622", "title": "Agentic RCA for Internet-Scale Services Using Constrained Creativity", "published": "2026-10-06T16:21:57Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08622v1", "location": "§3；§4.1–4.3；§6.1–6.5/Table3/Figs6–9；§8", "decision": "accepted", "reason": "E4以固定typed DAG语法约束RCA Agent，依次选playbook、组合已有operators、合成新operators；实际执行与admission形成可复现核心控制机制。", "reviewer": "agent_review", "evidence": "同Gemini3.1Pro对比LLM-only/RCA-Agent/Terminus2；DSB full0.91、starter-only0.72、无admission0.66；6生产incident pilot命中5例，其中2例需analyst hints。", "limitations": "§4.3用用户true-root-cause反馈admit，只能在训练/validation接收，不能test gold晋级；Accuracy@k,i允许多次尝试，不是单次accuracy；自建DSB有operator设计混杂；正文70 incident阶段统计与Table3总量不一致，图注≤56%与正文其他集0.80也不一致，headline不能无条件复用。", "institution": "Georgia Tech; Carnegie Mellon University; Conviva", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08621", "title": "Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness", "published": "2026-10-06T16:21:44Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08621", "title": "Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness", "published": "2026-10-06T16:21:44Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08621v1", "location": "§3.1–3.3；§4.1 Tables1–2；§4.2–4.4/Table3；§5", "decision": "accepted", "reason": "Recursive Game Creator包含designer/build、可执行多策略player rollout、盲化版本review及实际artifact retention，保留偏好证据到下一轮，属于固定模型Agent迭代闭环。", "reviewer": "agent_review", "evidence": "45 GameCraft任务同GPT6Astra baseline71.26→round3 77.89；47 GameASG任务success9/47→25/47，L2mean86.6→93.4。", "limitations": "非等预算基线，三轮与单轮混杂；用户study仅报告expert数值和playtime，未给强随机对照；97.1%coverage不是GUI速度提升；proxy player不代表人群偏好；能力复现必须实际运行游戏与policy不可gold固定轨迹。", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08586", "title": "MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory", "published": "2026-10-06T15:55:42Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08586", "title": "MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory", "published": "2026-10-06T15:55:42Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08586v1", "location": "§3.1–3.7；§4.1–4.4；§5.1–5.5；Limitations", "decision": "accepted", "reason": "MINDSET以immutable episodes和versioned schemas，按distortion/contradiction/history/fragmentation/inconsistency加权能量选择reinforce/supersede/split/new，再展开原始证据检索；定义完整可复现记忆算法。", "reviewer": "agent_review", "evidence": "LoCoMo700按conversation10簇显著提升排名；较LightMem F1+0.0256不显著p=.24023；Gemma上LightMem微领先0.0002；MAB150绝对ICL仅.024。", "limitations": "参数为开发集手调非最优理论，dev仅question-disjoint未证conversation-disjoint；LongContext按20k字节截断；基线为复刻，部分upstream artifacts缺失；最终stay margin而非switch margin决定行动，复现需保留这一实现细节。", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08561", "title": "Reinforcement Learning for Hierarchical Reasoning Rewards: Minimax-Optimal Rates with Transformers", "published": null, "reviews": [ { "track": "post-training", "id": "2610.08561", "title": "Reinforcement Learning for Hierarchical Reasoning Rewards: Minimax-Optimal Rates with Transformers", "repository_status": "new", "decision": "rejected", "reason": "定义depth curriculum/prefix自适应查询actor-critic并证明最小极大速率，但本文是形式化理论设置，未核实实际LLM训练benchmark；不列能力复现。", "review_level": "fulltext", "evidence": [ { "url": "https://arxiv.org/html/2610.08561v1", "location": "§3.2 Algorithm 1；Theorem 3及后续证明" } ], "fulltext_review_completed": true, "fulltext_scope": "针对方法、实验协议与主要结果的正文定向核查；非逐字通读全部附录", "pool": "main", "review_status": "reviewed" } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08560", "title": "Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness", "published": "2026-10-06T15:43:04Z", "reviews": [ { "track": "foundation-model", "id": "2610.08560", "title": "Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness", "published": "2026-10-06T15:43:04Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "Readiness Gating以冻结Qwen残差的嵌套CV logistic probe预测证据就绪，按训练分位阈值joint readiness/confidence首次停止或deadline最高readiness选答；视频级留出、crossed-quad消除图像/问题独立捷径。固定窗口policy不读证据时间，growing-prefix诊断采样依赖标注不能当部署延迟；§4.4给matched-video-seconds增益。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08560", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08560v1", "location": "§3.1–3.3；§4.4；Appendix C/H/I" }, { "url": "https://arxiv.org/html/2610.08560v1", "location": "§3.1–3.3；§4.4；Appendix C/H/I", "notes": "Readiness Gating以冻结Qwen残差的嵌套CV logistic probe预测证据就绪，按训练分位阈值joint readiness/confidence首次停止或deadline最高readiness选答；视频级留出、crossed-quad消除图像/问题独立捷径。固定窗口policy不读证据时间，growing-prefix诊断采样依赖标注不能当部署延迟；§4.4给matched-video-seconds增益。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08560v1", "locations": [ "§3.1–3.3；§4.4；Appendix C/H/I" ], "implementation_status": "not_implemented", "fulltext_notes": "Readiness Gating以冻结Qwen残差的嵌套CV logistic probe预测证据就绪，按训练分位阈值joint readiness/confidence首次停止或deadline最高readiness选答；视频级留出、crossed-quad消除图像/问题独立捷径。固定窗口policy不读证据时间，growing-prefix诊断采样依赖标注不能当部署延迟；§4.4给matched-video-seconds增益。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08539", "title": "RSJEV: Discriminative Remote Sensing Scene Classification with Multimodal Large Language Models", "published": "2026-10-06T15:28:36Z", "reviews": [ { "track": "foundation-model", "id": "2610.08539", "title": "RSJEV: Discriminative Remote Sensing Scene Classification with Multimodal Large Language Models", "published": "2026-10-06T15:28:36Z", "repository_status": "new", "pool": "main", "decision": "rejected", "reason": "RSJEV把已有JEV/Mapika式候选条件化单token判别读出及CE+Brier目标用于遥感分类；正文核查未发现区别于已登记LLM2Jev/既有Decider的新核心训练机制，作为专门领域复用不另立基础模型算法。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08539", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08539v1", "location": "§I–III; §IV.A–E; TableII" }, { "url": "https://arxiv.org/html/2610.08539v1", "location": "§I–III; §IV.A–E; TableII", "notes": "以候选选项token的LM-head行直接读出概率；UCM/AID/NWPU三个已有数据集。Qwen3.5-0.8B 10epoch8A40、batch128，单A40延迟102.18ms，未提供新通用公共评测设施。" } ], "official_url": "https://arxiv.org/html/2610.08539v1", "locations": [ "§I–III", "§IV.A–E", "TableII" ], "code_url": "https://github.com/Dongtcs/RSJEV", "implementation_status": "not_implemented", "fulltext_notes": "以候选选项token的LM-head行直接读出概率；UCM/AID/NWPU三个已有数据集。Qwen3.5-0.8B 10epoch8A40、batch128，单A40延迟102.18ms，未提供新通用公共评测设施。", "review_status": "reviewed" } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08528", "title": "MedCORE: Criteria-Grounded Clinical Reasoning for Interpretable Medical Image Diagnosis", "published": "2026-10-06T15:23:01Z", "reviews": [ { "track": "foundation-model", "id": "2610.08528", "title": "MedCORE: Criteria-Grounded Clinical Reasoning for Interpretable Medical Image Diagnosis", "published": "2026-10-06T15:23:01Z", "repository_status": "new", "pool": "main", "decision": "out_of_scope", "reason": "MedCORE面向皮肤、超声、视网膜病理诊断的临床标准图网络；属于特定医疗应用而非通用基础模型算法。", "review_level": "abstract", "abstract_review_completed": true, "fulltext_review_completed": false, "evidence": [ { "url": "https://arxiv.org/abs/2610.08528", "location": "Abstract（完整摘要已逐条审查）" } ] } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08526", "title": "WareFly-VLA: A Vision-Language-Action Framework for UAV Navigation and Human Tracking in Smart Warehouses", "published": "2026-10-06T15:21:14Z", "reviews": [ { "track": "foundation-model", "id": "2610.08526", "title": "WareFly-VLA: A Vision-Language-Action Framework for UAV Navigation and Human Tracking in Smart Warehouses", "published": "2026-10-06T15:21:14Z", "repository_status": "new", "pool": "main", "decision": "pending", "reason": "WareFly-VLA定义507飞行轨迹、4DoF连续动作及431/76 episode级固定split，有独立UAV评测价值；§6仅ground-truth帧上的open-loop MAE/correlation及轨迹积分，非闭环飞行成功率。§Data availability明确数据/runner仍access on request、acceptance后公开，故等待公开入口而非未读完。", "review_level": "fulltext", "kind": "evaluation_scope_recheck", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08526", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08526v1", "location": "§3；§5.2；§6.1 Tables 5–6；§7 limitations；Data availability" }, { "url": "https://arxiv.org/html/2610.08526v1", "location": "§3；§5.2；§6.1 Tables 5–6；§7 limitations；Data availability", "notes": "WareFly-VLA定义507飞行轨迹、4DoF连续动作及431/76 episode级固定split，有独立UAV评测价值；§6仅ground-truth帧上的open-loop MAE/correlation及轨迹积分，非闭环飞行成功率。§Data availability明确数据/runner仍access on request、acceptance后公开，故等待公开入口而非未读完。" } ], "review_status": "reviewed_deferred", "official_url": "https://arxiv.org/html/2610.08526v1", "locations": [ "§3；§5.2；§6.1 Tables 5–6；§7 limitations；Data availability" ], "implementation_status": "not_implemented", "fulltext_notes": "WareFly-VLA定义507飞行轨迹、4DoF连续动作及431/76 episode级固定split，有独立UAV评测价值；§6仅ground-truth帧上的open-loop MAE/correlation及轨迹积分，非闭环飞行成功率。§Data availability明确数据/runner仍access on request、acceptance后公开，故等待公开入口而非未读完。" } ], "status": "reviewed_deferred", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08514", "title": "How Much Evidence Should a Coding Agent's Self-Correction Carry? Adaptive Dirichlet Evidence for Self-Distillation", "published": "2026-10-06T15:17:56Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08514", "title": "How Much Evidence Should a Coding Agent's Self-Correction Carry? Adaptive Dirichlet Evidence for Self-Distillation", "published": "2026-10-06T15:17:56Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08514v1", "location": "§3.1–3.3 Eq1–11；§4.1–4.4；§5", "decision": "accepted", "reason": "EESD将公开执行before/after四种转移和词法相关性转为有效样本量1/sum(p²)，Dirichlet utility lower score加权CE并保留未加权forward KL；可复现Agent自修正训练机制。", "reviewer": "agent_review", "evidence": "DeepSeek/CodeARC fresh Pass@1 15.0→20.4%，配对source CI[2.8,8.0]pp；12设置范围−6.8到+5.4pp。", "limitations": "只一轮同source适应，不是source-disjoint泛化；full update vs no-update不能归因有效质量或KL；bootstrap条件于单次训练；有效质量仅relevance集中度，不自动修复执行依赖。", "fulltext_review_completed": true, "review_status": "reviewed", "implementation_status": "not_implemented_in_this_review" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08513", "title": "Wiki-Talkie: Multilingual Benchmarking of Persona-Based Agents on Real-World Discussions", "published": "2026-10-06T15:17:46Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08513", "title": "Wiki-Talkie: Multilingual Benchmarking of Persona-Based Agents on Real-World Discussions", "published": "2026-10-06T15:17:46Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08513v1", "location": "§3.1；§4；§5.2；Limitations/Ethics；official GitHub API2026-10-07", "decision": "pending", "reason": "Amazon Wiki-Talkie的五语人口行为分布评测有公共评测价值，但官方amazon-science/wikitalkie返回404，数据又明确申请访问非开放镜像；模型A–D及extractor匿名，现阶段审结待资源。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed_deferred", "reviewer": "agent_review", "evidence": "Wiki2025-10dumps，文章取首评论前revision；next-turn使用真实历史，history5–20先前评论，dialogact annotatorF1.76，JSD/情绪WD衡量群体而非个体人格；CCBYSA4数据声明。", "limitations": "behavior从用户50随机threads提炼，接入需重新核严格时间/目标排除；history与profile非信息量配平，profile长度分层不能证明因果；模型匿名阻原数值复现；saltedhash不消除文本反识别，群体研究不得个体冒充。", "code": "https://github.com/amazon-science/wikitalkie", "pending_reason": "official_repository_404_access_request_data_anonymous_models", "candidate_kind": "evaluation_infrastructure", "fulltext_review_completed": true } ], "status": "reviewed_deferred", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08482", "title": "Knee3DVLM: Dual-Sequence Full-Volume Vision-Language Modeling for Comprehensive Knee MRI Assessment", "published": "2026-10-06T14:58:53Z", "reviews": [ { "track": "foundation-model", "id": "2610.08482", "title": "Knee3DVLM: Dual-Sequence Full-Volume Vision-Language Modeling for Comprehensive Knee MRI Assessment", "published": "2026-10-06T14:58:53Z", "repository_status": "new", "pool": "main", "decision": "out_of_scope", "reason": "Knee3DVLM为膝关节MRI双序列57诊断目标建模，属特定医疗模型应用。", "review_level": "abstract", "abstract_review_completed": true, "fulltext_review_completed": false, "evidence": [ { "url": "https://arxiv.org/abs/2610.08482", "location": "Abstract（完整摘要已逐条审查）" } ] } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08463", "title": "UNREAL: Unifying Retrieval and Long-Context with a Single Model", "published": "2026-10-06T14:48:53Z", "reviews": [ { "track": "foundation-model", "id": "2610.08463", "title": "UNREAL: Unifying Retrieval and Long-Context with a Single Model", "published": "2026-10-06T14:48:53Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "UNREAL用冻结decoder中层chunk表征、训练64检索token及跨层混合做MaxSim，再重编码所选文本生成；提供multi-positive InfoNCE、公开QA训练split与long-context独立验证协议。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08463", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08463v1", "location": "§2.1–2.3 Eq1–7; §3/Table1; §4.2/Figure7; §6; Appendix B.1–B.3" }, { "url": "https://arxiv.org/html/2610.08463v1", "location": "§2.1–2.3 Eq1–7; §3/Table1; §4.2/Figure7; §6; Appendix B.1–B.3", "notes": "21Mchunk 3BtokenWiki2018，query初始top5 BM25，每chunk7vectors、query4groups、<500K训练参数；15K步batch256，500hardnegative从训练gold挖掘。H100约32K后测得TTFT优势。" } ], "official_url": "https://arxiv.org/html/2610.08463v1", "locations": [ "§2.1–2.3 Eq1–7", "§3/Table1", "§4.2/Figure7", "§6", "Appendix B.1–B.3" ], "limitations": "测试不可注入oracle，只有训练hardnegative挖掘用gold；100M场景以稀疏证据为主，dashed延迟是FLOP外推非实测。索引离线编码存储和两次模型forward须计成本；不同baseline训练预算并非完全匹配。", "implementation_status": "not_implemented", "fulltext_notes": "21Mchunk 3BtokenWiki2018，query初始top5 BM25，每chunk7vectors、query4groups、<500K训练参数；15K步batch256，500hardnegative从训练gold挖掘。H100约32K后测得TTFT优势。", "review_status": "reviewed" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08452", "title": "Agentic AutoRAG: RAG Pipeline Optimization through Reasoning-Driven Agents", "published": "2026-10-06T14:38:15Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08452", "title": "Agentic AutoRAG: RAG Pipeline Optimization through Reasoning-Driven Agents", "published": "2026-10-06T14:38:15Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08452v1", "location": "§3.1–3.3；§4；§5/Table2；Limitations", "decision": "accepted", "reason": "Agentic AutoRAG的检索/生成归因→Diagnoser→KB/成本引导Proposer构成明确搜索机制，可接入evolve候选；准确率实验有100验证/300独立heldout隔离。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed", "reviewer": "agent_review", "evidence": "冻结exam+gold-span命中仅用于验证优化；30trial×10seed，heldout judge Hotpot/MuSiQue/MultiHop .825/.330/.790；生成exam的Pareto医疗场景77%vs71.5%，成本57.7%。", "limitations": "医疗Pareto只报告反复优化的同一exam，非heldout能力；成本轴只含生成/扩展API，不含本地embedding/rerank；noKB/diag合并消融不能归因单组件；warmTPE获免费30次同库先验，且不是完整syftr；judge无人类一致性验证。", "code": "https://github.com/Agentic-Systems-Lab/Agentic-AutoRAG", "code_status": "official GitHub contents lists source/configs/knowledge_base/scripts/tests/LICENSE; not executed", "fulltext_review_completed": true } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08430", "title": "NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale", "published": "2026-10-06T14:30:06Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08430", "title": "NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale", "published": "2026-10-06T14:30:06Z", "source_group": "original", "review_level": "abstract", "source": "https://arxiv.org/abs/2610.08430", "location": "Title and complete abstract", "decision": "out_of_scope", "reason": "NeMo-DCR权重同步/精确delta传输系统，归训练基础设施而非Agent策略。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed" } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08414", "title": "Image Bitstream Fine-grained Understanding for Privacy-Friendly AIoT", "published": "2026-10-06T14:18:54Z", "reviews": [ { "track": "foundation-model", "id": "2610.08414", "title": "Image Bitstream Fine-grained Understanding for Privacy-Friendly AIoT", "published": "2026-10-06T14:18:54Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "BFG将JPEG bytes经ByteFormer分层卷积/shifted-window编码为192维，projector到512维后全局上下文融合+局部cross-attention自回归描述，另120类head；CFU-D源ID/split隔离并测试替换/删除/重复腐坏。Dogs分类58.51%采用ImageNet初始化；不解码不等价正式隐私保证，不能包装为通用大规模foundation结论。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08414", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08414v1", "location": "§III-B式2–5；§III-C；§IV-A–D Tables I/IV/V" }, { "url": "https://arxiv.org/html/2610.08414v1", "location": "§III-B式2–5；§III-C；§IV-A–D Tables I/IV/V", "notes": "BFG将JPEG bytes经ByteFormer分层卷积/shifted-window编码为192维，projector到512维后全局上下文融合+局部cross-attention自回归描述，另120类head；CFU-D源ID/split隔离并测试替换/删除/重复腐坏。Dogs分类58.51%采用ImageNet初始化；不解码不等价正式隐私保证，不能包装为通用大规模foundation结论。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08414v1", "locations": [ "§III-B式2–5；§III-C；§IV-A–D Tables I/IV/V" ], "implementation_status": "not_implemented", "fulltext_notes": "BFG将JPEG bytes经ByteFormer分层卷积/shifted-window编码为192维，projector到512维后全局上下文融合+局部cross-attention自回归描述，另120类head；CFU-D源ID/split隔离并测试替换/删除/重复腐坏。Dogs分类58.51%采用ImageNet初始化；不解码不等价正式隐私保证，不能包装为通用大规模foundation结论。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08407", "title": "Seeing the Context: Enhancing Recommender Systems with Image-Derived Contextual Signals", "published": "2026-10-06T14:15:48Z", "reviews": [ { "track": "recommendation", "id": "2610.08407", "title": "Seeing the Context: Enhancing Recommender Systems with Image-Derived Contextual Signals", "published": "2026-10-06T14:15:48Z", "repository_status": "new", "decision": "rejected", "review_level": "full-text", "fulltext_review_completed": true, "reason": "ICE-Fuse 在 TripAdvisor 图文上下文上做 RMSE/消融；未提供线上分流或部署实验，不满足工业搜广推准入。", "evidence": [ { "url": "https://arxiv.org/abs/2610.08407", "location": "§4.1–4.3 Tables 2–3; §5" } ], "fulltext_receipt": { "id": "2610.08407", "url": "https://arxiv.org/html/2610.08407v1", "sha256": "2223d22274d906bfa43f36d184feed98ec023a0af2ebe0b83ce2c0f97262ed78", "status": "downloaded", "characters": 28755, "errors": [] } } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08403", "title": "SSR: Sparse Segment Reduction for Ternary GEMM Acceleration", "published": "2026-10-06T14:15:12Z", "reviews": [ { "track": "foundation-model", "id": "2610.08403", "title": "SSR: Sparse Segment Reduction for Ternary GEMM Acceleration", "published": "2026-10-06T14:15:12Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "SSR 将三值矩阵分成正负二值矩阵，按列块模式排序并剔除全零模式，以分段求和与树规约代替乘法；公开 C++/AVX2 实现可核验。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08403", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08403v1", "location": "§III-A/B Algorithms1–2; §IV-A/B; §IV-C TableIII; §V" }, { "url": "https://arxiv.org/html/2610.08403v1", "location": "§III-A/B Algorithms1–2; §IV-A/B; §IV-C TableIII; §V", "notes": "" } ], "kind": "cpu_inference_kernel", "locations": [ "§III-A/B Algorithms1–2", "§IV-A/B", "§IV-C TableIII", "§V" ], "limitations": "报告的是 CPU，部分高稀疏设置慢于 CSC/Eigen；原文未验证 GPU，不能用仓库其他项目的 GPU 目录冒充 SSR。模型级主要是改造三值 MLP 的延迟，没有公开质量保真证据。作者 README 引用 DATE2026，应标记本次 arXiv 收录而非确认首次发表。", "code_url": "https://github.com/fpgasystems/ternaryLLM/tree/main/SSR", "code_status": "author-linked repository accessible; MIT; SSR is separate from ternaryLLM_GPU", "official_url": "https://arxiv.org/html/2610.08403v1", "review_provenance": "root 独立正文审查；foundation读取整合", "implementation_status": "not_implemented", "fulltext_notes": "", "review_status": "reviewed" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08401", "title": "GeoPID: Decomposing and Steering Visual Information in Vision-Language Models", "published": "2026-10-06T14:14:25Z", "reviews": [ { "track": "foundation-model", "id": "2610.08401", "title": "GeoPID: Decomposing and Steering Visual Information in Vision-Language Models", "published": "2026-10-06T14:14:25Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "GeoPID从无标签校准表征估计模态共享/独占子空间，推理按I+αUUᵀ放大vision-unique；22×14组合每组100道独立校准题用标签选层/α后冻结，最终评测不读gold。任务Rényi信息分解为离线诊断，平均相对7.63%需保留308组异质/部分负收益。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08401", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08401v1", "location": "§2.1–2.3式11；§3.1；§3.2 Table 5" }, { "url": "https://arxiv.org/html/2610.08401v1", "location": "§2.1–2.3式11；§3.1；§3.2 Table 5", "notes": "GeoPID从无标签校准表征估计模态共享/独占子空间，推理按I+αUUᵀ放大vision-unique；22×14组合每组100道独立校准题用标签选层/α后冻结，最终评测不读gold。任务Rényi信息分解为离线诊断，平均相对7.63%需保留308组异质/部分负收益。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08401v1", "locations": [ "§2.1–2.3式11；§3.1；§3.2 Table 5" ], "implementation_status": "not_implemented", "fulltext_notes": "GeoPID从无标签校准表征估计模态共享/独占子空间，推理按I+αUUᵀ放大vision-unique；22×14组合每组100道独立校准题用标签选层/α后冻结，最终评测不读gold。任务Rényi信息分解为离线诊断，平均相对7.63%需保留308组异质/部分负收益。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08400", "title": "Atom-JEPA: Joint-Embedding Predictive Architecture for 3D Atomistic Systems", "published": "2026-10-06T14:13:51Z", "reviews": [ { "track": "foundation-model", "id": "2610.08400", "title": "Atom-JEPA: Joint-Embedding Predictive Architecture for 3D Atomistic Systems", "published": "2026-10-06T14:13:51Z", "repository_status": "new", "pool": "main", "decision": "pending", "reason": "Atom-JEPA具备真实SE(3)等变原子/子结构预测算法与公开迁移结果，但严格等预算新目标归因尚未满足：§G.1同Uni-Mol atom-only与atom+pool的批量384/128、学习率2.42e-4/1.4e-4不同；§B未做同架构其他SSL比较。需明确可接受的匹配控制或补齐预算证据，不能把scratch对比当等预训练预算。", "review_level": "fulltext", "official_url": "https://arxiv.org/html/2610.08400v1", "locations": [ "§3", "§4.2/Table4", "Appendix B", "Appendix G.1/Table26", "Appendix H.1.1" ], "evidence": [ { "url": "https://arxiv.org/abs/2610.08400", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08400v1", "location": "§3; §4.2/Table4; Appendix B; Appendix G.1/Table26; Appendix H.1.1" }, { "url": "https://arxiv.org/html/2610.08400v1", "location": "§3; §4.2/Table4; Appendix B; Appendix G.1/Table26; Appendix H.1.1", "notes": "19M Uni-Mol与1.7M Alexandria预训练，双目标EMA EquiformerV3；Matbench40任务fold全部优于scratch，ChEMBL-MT仍有-0.5%负结果。Uni-Mol2475 GPU小时、Alexandria2147 GPU小时；主目标对照训练配置不相同。" } ], "code_url": "https://github.com/khelverskovp/atom-jepa", "institution": "Technical University of Denmark", "implementation_status": "not_implemented", "abstract_review_completed": true, "fulltext_review_completed": true, "fulltext_notes": "19M Uni-Mol与1.7M Alexandria预训练，双目标EMA EquiformerV3；Matbench40任务fold全部优于scratch，ChEMBL-MT仍有-0.5%负结果。Uni-Mol2475 GPU小时、Alexandria2147 GPU小时；主目标对照训练配置不相同。", "review_status": "reviewed_deferred" } ], "status": "reviewed_deferred", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08364", "title": "Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations", "published": "2026-10-06T13:52:15Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08364", "title": "Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations", "published": "2026-10-06T13:52:15Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08364v1", "location": "§2.1–2.2；§3.1/3.3–3.5；§4；official README/API", "decision": "accepted", "reason": "Transect为公共可执行评测观测工具：family Spec固定词表，structural/judged scanners→provenance tables→源链接报告；可用于evolve审计但不是任务能力评分器。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed", "reviewer": "agent_review", "evidence": "UK AISI/Meridian MIT公开repo，Python3.12+，structural worked example无judge调用；solo/kroll/cohort记录member votes/verifier修改。单CRUX案例580orchestrator outputs/131sessions922fileaccess，来源/覆盖明确。", "limitations": "仅单案例描述，agreement/repeatability不是正确率；读文件不证明采纳、sourceorder非walltime、compaction邻近非因果；Inspect machine-global缓存可能伪造重复稳定性，独立judge重复须freshcache；扩展扫描器可能自行调用模型。", "code": "https://github.com/AI-Safety-Institute/transect", "code_status": "MIT; executable structural example and judged entrypoints verified in README, not executed", "candidate_kind": "evaluation_observability_infrastructure", "fulltext_review_completed": true } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08341", "title": "DIPrune: Task-Aware Token Pruning with Dual Importance for Efficient Multimodal Language Models", "published": "2026-10-06T13:38:27Z", "reviews": [ { "track": "foundation-model", "id": "2610.08341", "title": "DIPrune: Task-Aware Token Pruning with Dual Importance for Efficient Multimodal Language Models", "published": "2026-10-06T13:38:27Z", "repository_status": "new", "pool": "main", "decision": "accepted", "reason": "DIPrune结合文本到视觉注意力倒置rank归一与跨层rank上升的截断归一，保留早层低但后层新出现语义token并移除对应KV；多VLM及视频骨干评测、层对齐消融。LLaVA64tokens平均保留97.7%，POPE TTFT73.23→32.50ms、额外0.24ms，需以实际TTFT而非FLOPs主张速度。", "review_level": "fulltext", "abstract_review_completed": true, "fulltext_review_completed": true, "evidence": [ { "url": "https://arxiv.org/abs/2610.08341", "location": "完整摘要逐条审查" }, { "url": "https://arxiv.org/html/2610.08341v1", "location": "§5.1式6–7；§5.2式8；§6.1；§6.5 Table 4" }, { "url": "https://arxiv.org/html/2610.08341v1", "location": "§5.1式6–7；§5.2式8；§6.1；§6.5 Table 4", "notes": "DIPrune结合文本到视觉注意力倒置rank归一与跨层rank上升的截断归一，保留早层低但后层新出现语义token并移除对应KV；多VLM及视频骨干评测、层对齐消融。LLaVA64tokens平均保留97.7%，POPE TTFT73.23→32.50ms、额外0.24ms，需以实际TTFT而非FLOPs主张速度。" } ], "review_status": "reviewed", "official_url": "https://arxiv.org/html/2610.08341v1", "locations": [ "§5.1式6–7；§5.2式8；§6.1；§6.5 Table 4" ], "implementation_status": "not_implemented", "fulltext_notes": "DIPrune结合文本到视觉注意力倒置rank归一与跨层rank上升的截断归一，保留早层低但后层新出现语义token并移除对应KV；多VLM及视频骨干评测、层对齐消融。LLaVA64tokens平均保留97.7%，POPE TTFT73.23→32.50ms、额外0.24ms，需以实际TTFT而非FLOPs主张速度。" } ], "status": "accepted", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08337", "title": "Machine Learning for German Redispatch Forecasting under Data Delays and Temporal Distribution Shift", "published": "2026-10-06T13:35:50Z", "reviews": [ { "track": "foundation-model", "id": "2610.08337", "title": "Machine Learning for German Redispatch Forecasting under Data Delays and Temporal Distribution Shift", "published": "2026-10-06T13:35:50Z", "repository_status": "new", "pool": "main", "decision": "out_of_scope", "reason": "德国电网再调度概率预测比较，属于能源时序应用而非基础模型核心算法。", "review_level": "abstract", "abstract_review_completed": true, "fulltext_review_completed": false, "evidence": [ { "url": "https://arxiv.org/abs/2610.08337", "location": "Abstract（完整摘要已逐条审查）" } ] } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08331", "title": "Transferable Spatial Temporal Coherence Adversarial Attack on Black-Box Vision Language Models for Autonomous Driving", "published": "2026-10-06T13:30:06Z", "reviews": [ { "track": "foundation-model", "id": "2610.08331", "title": "Transferable Spatial Temporal Coherence Adversarial Attack on Black-Box Vision Language Models for Autonomous Driving", "published": "2026-10-06T13:30:06Z", "repository_status": "new", "pool": "main", "decision": "out_of_scope", "reason": "自动驾驶视频VLM迁移攻击研究，主贡献为攻击构造而非基础模型能力训练/推理方法。", "review_level": "abstract", "abstract_review_completed": true, "fulltext_review_completed": false, "evidence": [ { "url": "https://arxiv.org/abs/2610.08331", "location": "Abstract（完整摘要已逐条审查）" } ] } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08327", "title": "MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation", "published": "2026-10-06T13:25:55Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08327", "title": "MedZERO: Self-Evolving Agents for Open-Ended Medical Reasoning Through Controlled Knowledge Accumulation", "published": "2026-10-06T13:25:55Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08327v1", "location": "§3 Eq1–21；§4.1–4.2；§5；AppendixB；Open access checklist", "decision": "pending", "reason": "MedZERO交替训练Examiner/Reasoner与临时→持久KG准入是合格Agent机制；正文明确投稿时不发布代码、计划出版后发布，grounding scorer/映射阈值等关键实现不足，记审结待发布。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed_deferred", "reviewer": "agent_review", "evidence": "4响应option映射多数一致性作为弱frontier，curation保留[.3,.8]；临时triplet仅可读前批且最终丢弃，独立重生成去重持久化；8B六分项均值53.5vsAgent043.7。", "limitations": "多数一致性不是医学正确率；持久准入缺独立事实验证，仍可能自我强化错误；MCQ选项只给后置judge不应给被测Reasoner；临床效用未建立，KG/web/judge依赖明显，正文未充分规定grounding评分函数。", "pending_reason": "author_code_not_released_and_core_details_incomplete", "fulltext_review_completed": true } ], "status": "reviewed_deferred", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08300", "title": "Memory Depth and Reconstructed Context Width: A Controlled Evaluation of Hierarchical Retrieval", "published": "2026-10-06T13:04:43Z", "reviews": [ { "track": "foundation-model", "id": "2610.08300", "title": "Memory Depth and Reconstructed Context Width: A Controlled Evaluation of Hierarchical Retrieval", "published": "2026-10-06T13:04:43Z", "repository_status": "new", "pool": "main", "decision": "out_of_scope", "reason": "分层会话检索深度/上下文宽度的控制实验，主要为系统诊断而非新的基础模型核心算法。", "review_level": "abstract", "abstract_review_completed": true, "fulltext_review_completed": false, "evidence": [ { "url": "https://arxiv.org/abs/2610.08300", "location": "Abstract（完整摘要已逐条审查）" } ] } ], "status": "rejected_or_out_of_scope", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08250", "title": "MASC: A Multi-Agent Self-Calibration Framework with Latent Construct Alignment for Consistent Client Role-Playing in Psychological Counseling", "published": "2026-10-06T12:31:05Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08250", "title": "MASC: A Multi-Agent Self-Calibration Framework with Latent Construct Alignment for Consistent Client Role-Playing in Psychological Counseling", "published": "2026-10-06T12:31:05Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08250v1", "location": "Multi-Agent Self-Calibration §方法 Eq1–16/Alg1；Experiment Tables2–3；官方GitHub API", "decision": "pending", "reason": "MASC以预测情绪/状态/动作、三Agent五步辩论、投票一致性优先和两轮失败记忆重试维持角色；方法与CRPC评测有关联价值，但官方仓库description明确code and data coming soon且无license。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed_deferred", "reviewer": "agent_review", "evidence": "38profiles；只读过去对话不读未来reference；最多3尝试、每次3×5generation；独立盲评而非复用内部投票标签；Hetero action/emotion/state KL .06/.27/.06。", "limitations": "38profiles及同套构念/提示限制外推；Hetero与Homo骨干不同不能隔离多样性原因；73.97%最终匹配，44.19%是以最终收敛为条件的重试占比非修复概率；KL仅群体标签频率不是治疗真实性。", "code": "https://github.com/DeSun99/MASC-CRPC", "pending_reason": "author_code_and_data_coming_soon", "fulltext_review_completed": true } ], "status": "reviewed_deferred", "implementation_status": "not_implemented_in_this_review" }, { "id": "2610.08246", "title": "LeanPlan: Optimal Planning with LLM-Generated Heuristics and Admissibility Proofs", "published": "2026-10-06T12:27:34Z", "reviews": [ { "track": "agent", "arxiv_id": "2610.08246", "title": "LeanPlan: Optimal Planning with LLM-Generated Heuristics and Admissibility Proofs", "published": "2026-10-06T12:27:34Z", "source_group": "original", "review_level": "fulltext", "source": "https://arxiv.org/html/2610.08246v1", "location": "§3–5；§7 Limitations；Reproducibility", "decision": "pending", "reason": "LeanPlan生成Lean启发式+可采纳性证明，controller拒绝sorry/admit/native_decide/新公理且独立重编译与训练任务验收，符合可验证Agent/evolve机制；作者明确代码/benchmark/log接受后才发布，审结待发布。", "implementation_status": "not_implemented_in_this_review", "review_status": "reviewed_deferred", "reviewer": "agent_review", "evidence": "13domain1170test；sandbox隔离test，按domain单次搜索均69.8分钟17.30美元；595超训练对象规模test解53vsSCP13；每任务300CPU秒8GiB。", "limitations": "SCP预算曾用test suite调优，原对照非严格完整test隔离；每domain单次Agent生成，无多seed；证明仅对certificate通过且成本