Agent: accepted in review but not delivered (206)¶
These entries passed an earlier review and lack a current catalog entry or an explicit cost closure. This is not user authorization to implement every item, nor does it mean each item passed the current cost gate. Do not use this as an automatic implementation queue. Sorted by paper ID descending.
Confirmed decisions not to reproduce full experiments are marked separately. Other items first need cost and scope reduction, not automatic execution.
| ID / paper | Current disposition | Priority organization verification | Prior review rationale |
|---|---|---|---|
| 2610.20067 Which Knowledge Should LLMs Trust? Uncovering a Generalizable Steering Direction Across Knowledge Conflicts | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | The knowledge-conflict direction is a linear intervention in residual space, using 7B–9B inference on one A100. Accept real-model direction extraction on independent development pairs and isolated test. Accuracy conditional on clean correctness is not unconditional capability; no answers in inference. |
| 2610.19818 One Monitor Does Not Fit All: Formalism-Agnostic Runtime Verification for Agentic AI | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | AgenticRV's canonical events, monitor-specific projection, unified verdicts, and pre-effect blocking are executable on Mac with Python and controlled workflows. Accept only the mechanism; instant-return monitor microbenchmarks are not semantic monitoring or agent capability. |
| 2610.19801 Delta Tokenizer: Reference-Identical Tokenization of Session Deltas and Segment-Parallelism for Long Prompts on Multicore CPUs | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Delta Tokenizer needs provably safe incremental seams under tokenizer rules, exact segmentation, and chunk scheduling, not token concatenation. Verify with real tokenizer/text on Mac CPU. No 41.8B-document or multi-server rerun and no TTFT measurement claim. |
| 2610.19659 Measuring and Reducing Unusable Replies of LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Reply usability is bounded by what a reader actually saw. One rewrite or explicit reply instruction needs no training. Accept reader checks and real rewriting on a few replies, not gold substitutes. The paper's 480 replies across eight models are not local results. |
| 2610.19474 Beyond Retrieval: Evidence Access and Event Aggregation in Agent Memory | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | TallyMem keeps raw traces, generates event specifications, and executes aggregates rather than substituting gold events. Accept raw logs→ledger→real compilation/execution. USD 1.05 covers only the paper's primary arm, not all AMA/readers. AMA includes development traces; formal comparisons should use 25 development and 99 isolated repositories. |
| 2610.19345 When Is Smaller Greener? Right-Sizing Language Models for Agentic AI | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | The contribution measures energy, money, and time per success and Pareto-filters choices. Accept CPU calculations and crossover lengths using public parameters. The 15-model GPU pilot was not reproduced, and synthetic reliability is not agent capability. |
| 2610.19316 Truthful and Expected-Payment-Optimal Model Routing in Graph-Based Agentic Workflows | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | STAR's allocation, workload monotonicity, and integral payments can run on small synthetic graphs with an exact solver, without LLM calls. Accept the auction/routing mechanism, not synthetic costs or incentive guarantees as agent gains. |
| 2610.19131 Plant Memory: Technical Language Processing and a Verified Maintenance Ontology for Agent-Accessible Failure Knowledge from SAP PM Work Orders | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Plant Memory's rule-first TLP, evidence/review-aware fact graph, and offline SQLite are affordable; a second LLM reader is optional. Accept rules and provenance only. Results on 400 synthetic work orders are not factory deployment, general agent ability, or Evolve gains. |
| 2610.18780 BLADE: Auditing Information Use for Failure Attribution in LLM Multi-Agent Systems | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | BLADE traces source use through decisions, distinguishing introduction, propagation, and repair without reference answers. The paper estimates roughly USD 1.9 for Jev inference over 1,228 traces. Accept affordable real-trace review, requiring a real decision model rather than gold blame-step fixtures. |
| 2610.18684 Knowing the User Is Not Knowing What We Established: Grounding Awareness in Personal Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Grounding Layer separates user facts from accepted, unconfirmed, revised, and scoped agreements. Keep the ledger, pre-action gate, and raw-dialogue lookup. Verify extraction from real dialogues and bounded actions, not oracle +Status/+Policy or gold actions. No full benchmark gain claimed. |
| 2610.18157 Fan Out or Relay? Ownership-Aware Routing for LLM Agents Querying Siloed Tables | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | OAR drafts SQL and computes owner cover using only public schemas, choosing fan-out or relay before reading private rows. Test on a small real database with an actual drafter, not gold SQL or owner labels. The 13 configurations were not reproduced. |
| 2610.08775 Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. BOTTLED is accepted as executable Evolve infrastructure: an agent produces reusable artifacts under budget and tackles many unlabeled tasks with hidden gold, compatible with models, scripts, and routers. |
| 2610.08691 ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ScienceClaw freezes models and repairs typed workflows into linked Skill–Operator bundles. Boundary/source-reset replay must use the new version; promotion is atomic only after strict independent validation gains and hard constraints/budget pass. A genuine Evolve candidate. |
| 2610.08662 ParanoiaEval: Benchmarking Unnecessary Defensive Work in Agentic Coding | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ParanoiaEval offers 200 same-task E+/E− evidence pairs, with task oracles separate from judges of unnecessary risk work. It measures completion and evidence response for Evolve behavior constraints. |
| 2610.08647 SquidAgent: Parallelize Wisely, Coordinate Efficiently | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SquidAgent estimates output tokens and alignment cost in one DAG plan. Session forks inherit context and prewritten contracts; parallelize when sum(tokens)/(max(tokens)+alignment)>alpha. Executable agent scheduling. |
| 2610.08622 Agentic RCA for Internet-Scale Services Using Constrained Creativity | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. E4 constrains RCA agents with a typed DAG grammar, selecting playbooks, composing known operators, then synthesizing new ones. Actual execution and admission define the mechanism. |
| 2610.08621 Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Recursive Game Creator includes designer/build, executable multi-policy player rollouts, blinded version review, and real artifact retention. Preference evidence carries forward under a fixed-model loop. |
| 2610.08586 MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MINDSET uses immutable episodes and versioned schemas. An energy over distortion, contradiction, history, fragmentation, and inconsistency selects reinforcement, supersession, split, or new schema, then retrieves raw evidence. |
| 2610.08514 How Much Evidence Should a Coding Agent's Self-Correction Carry? Adaptive Dirichlet Evidence for Self-Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. EESD maps four before/after public execution transitions and lexical relevance to effective sample size 1/sum(p²). Dirichlet lower utility weights CE; forward KL remains unweighted. A reproducible self-correction objective. |
| 2610.08452 Agentic AutoRAG: RAG Pipeline Optimization through Reasoning-Driven Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Agentic AutoRAG attributes retrieval/generation errors, then uses Diagnoser and a KB/cost-guided Proposer. Candidate Evolve search; accuracy uses 100 validation and 300 independent holdout examples. |
| 2610.08364 Transect: Retaining Observability for Long-Horizon LLM Agent Evaluations | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Transect is executable observability infrastructure: fixed family vocabulary, structural/judged scans, provenance tables, and linked reports. Suitable for Evolve audit, not task scoring. |
| 2610.08155 Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. S1-MAS uses a frozen Laya discrete-decision encoder and Qwen 0.6B evidence localization. A lightweight controller handles Solve/Decompose/Reverse/Review/Repair/Contrast and stop, for collaborative compute allocation. |
| 2610.08106 ChartBmkAgent: Harness-Governed Multi-Agent Construction of Chart QA Benchmarks from Sparse Error-Taxonomy Specifications | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ChartBmkAgent treats taxonomy/difficulty as immutable roots. The harness approves data, charts, mechanism anchors and QA; failures trigger local repair or rollback. Candidate for auditable benchmark construction. |
| 2610.08101 Beyond Corrected Memory: Execution Consistency in Multi-Agent Systems | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CAVERT separates obligation activation, observation coverage, and valid fulfillment, returning U for insufficient evidence. LLM extraction, deterministic rules, and a precommit read/refresh gate define execution consistency. |
| 2610.08076 SpeedrunBench: Challenging LLM Agents with Video Game Speedrunning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SpeedrunBench optimizes replayable actions under fixed rounds and scoring calls. Three open-source games, pinned revisions, and independent scoring containers support public Evolve evaluation. |
| 2610.08048 DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DAEDALUS generates tasks Explorer actually completes, extracts failure heuristics, verifies three consecutive successes after reset, adjusts difficulty, and injects a frozen library. Real memory construction and verification. |
| 2610.07972 Can Agents Work for Everyone? Cross-User Reliability for Mobile GUI Agents in Personalized User Interfaces | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. PAIR offers executable personalized GUI states; RePAIR normalizes advantages from four rollouts per cross-user UC subgoal. Accept exploratory mechanism; reselect independently on validation before integration. |
| 2610.07935 SIGMA: Self-Improving Alignment Generalization from a Model Spec | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Apple/JHU SIGMA uses a same-model task designer, one-fact paired changes, self-validation of specifications/rubrics, and training supervision generation. Keep real judges and GRPO, not rule fixtures. |
| 2610.07862 A self-learning scientific agent for X-ray diffraction | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Gan Jiang converts XRD tool residuals/failures into versioned skills, selects with development pairs, and freezes before holdout. Accept scientific-agent mechanism and DeltaXRDbench evaluation, not full closed-toolchain reproduction. |
| 2610.07860 WorkflowOps: Learning Agent Collaboration Priors for Multi-Agent Workflow Orchestration | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. WorkflowOps uses MiniLM capability matching, creates agents when coverage is low, and adds DAG soft edges and layered execution from synthetic-workflow transition priors. Orchestration candidate, not proven continual evolution. |
| 2610.07851 RA-MoWE: Workflow-Affinity Embeddings for Query Clustering and Agentic Workflow Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RA-MoWE routes through offline outcome-affinity probing, cluster-expert search, and BGE+MLP. Accept text-only Guided-BB; isolate measured test affinity as diagnostics. |
| 2610.07832 Harness Engineering for Software Engineering via Modular Executable Dev-Primitives | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HERMES wraps persisted files as Dev-Primitives, activates them by dependency, directs messages, diagnoses execution, and reactivates modifications for up to three rounds. Candidate for file-level orchestration. |
| 2610.07816 Do I Need the Cloud? Uncertainty-Aware Step-Level Handoff for Small Language Model Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. StepGate trains calibrated routing on local-action counterfactual failures, with six feature groups, thresholds/budgets, and optional consistency sampling to choose local, verification, or stronger models. |
| 2610.07657 Where Rules End and Judges Begin: Measuring the Judgment Boundary in Multi-Agent Systems Security | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DEFER runs 27 deterministic/similarity checks followed by an LLM panel for residual cases in host tool/memory security. Four-domain testbed, outcome oracle, and public MIT code/attacks/benchmarks/logs/scripts/tests support security evaluation. |
| 2610.07645 SkillPoison: Progressive Skill Poisoning via Successful Experiences | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SkillPoison shows governance failures when successful traces become skills stripped of applicability conditions. Candidate for defensive promotion regression or diagnostics, not capability improvement. |
| 2610.07639 HarnessSecurity-Bench: Do Security Mechanisms Really Protect Coding Agent Harnesses? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HarnessSecurity-Bench pairs native security ON/OFF under fixed legitimate tasks and separates deterministic utility from attack-effect oracles. Suitable for isolated offline harness regression. |
| 2610.07587 Large Language Model Orchestration under Heterogeneous Preferences via Explicit Persona Inference | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HARP updates each agent's Bayesian persona posterior with an offline-calibrated response scorer and uses posterior sampling for centralized planning. HARP+ adds reward-disagreement exploration. Mechanism only, no open-dialogue gain claim. |
| 2610.07557 CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CheckerBench/CheckerLab builds static checkers from vulnerable/patched versions, then freezes artifacts before independent rebuild, differential, false-positive, and semantic checks. Public 300 tasks and fixed Linux runtime support code-agent evaluation. |
| 2610.07376 MemCo: Memory-Centric Collaboration for Generalizing LLM Agents to Unseen Environments | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MemCo extracts conditional rules from local trajectory graphs, aggregates episode-reduced evidence by exact role keys, promotes/revokes with Wilson scores, and retrieves grounded state/phase/goal matches. |
| 2610.07335 Selective Critique for Cost-Aware LLM Agents in Long-Horizon Decision Making | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SAG gates by entropy and top-two margin over length-normalized executable-action likelihoods, without gold-critic correction. Every ten successes it restarts LoRA-SFT from the same base on accumulated action pairs. |
| 2610.07311 Understanding and Mitigating Inference-Time Overreliance Using Agentic Memory | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MemTrim uses LLM-normalized atomic evidence, task signatures, answer support, and a separate key–value trie. Removes duplicate/conflicting evidence, suppresses stale answers when support changes, and uses typed reuse across tasks. |
| 2610.07289 Catching Developers in the Flow: Low-Latency Agentic Program Repair at Google Scale | Full experiment not reproduced; review DONE; affordable components considered separately | Google/Meta confirmed; see priority list | Accepted in review; core implementation and public verification pending. Google FlowAgent applies pre/post-execution filtering, ReAct generate-and-validate, and final tests for low-latency presubmit fixes. Core candidate, not implemented. |
| 2610.07261 Verifying Coordination in Parallel Coding Agents: NP-Bench and a Scheduling Planner | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Nerveplane observes Git state passively, colors conflicting file scopes into batches, and orders merges by contracts. NP-Bench includes real Git integration and live coding agents for parallel collaboration evaluation. |
| 2610.07258 Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Lineage-Aware Memory Governance uses sensitive-column lineage tags, latest safe retrieval, and definition-hash conflicts to govern shared memory. Diagnostic candidate, not production security. |
| 2610.07191 Agentic Design Space Exploration for Joint Hardware Configuration Selection and Mapping of AI Inference Workloads on Heterogeneous Edge SoCs | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TraceDSE closes an actual hardware-search loop with execution traces, a tool-assisted critic, proposer, and Pareto/exploration state. Candidate Evolve mechanism. |
| 2610.06966 APEX: Active Protection at Execution Boundaries for LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. APEX freezes user authorization, checks effects at execution with WRAP, uses PLANT deception receipts for unauthorized information use, and continues safely on failures. Public AgentDojo/ASB/MCPTox/MSB/Skills support validation. |
| 2610.06964 Principles that Guide, Actions that Inform: Agent Evolution via Knowledge Abstraction | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SAGA abstracts trajectories into hierarchical experiences/principles, retrieves and adapts them, and may optimize the actor with SFT/GRPO. ScienceWorld/ALFWorld grounding and utility ledger rules are executable. |
| 2610.06843 Recursive Video In-Context Learning for Agentic Robot | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RV-ICL's hierarchical demo indexes and subgoal reentry define context access. Accept demonstration-conditioned diagnostics only; isolate stage goals/simulator state from gold-free capability evaluation. |
| 2610.06748 BazaarBench: Delegation Safety in Decentralized C2C Marketplaces Run by LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. BazaarBench continues reproducible 30-day market state for seven days with 20 fixed agents. Event/inventory truth is separate from transaction-safety stages, supporting multi-agent delegation evaluation. |
| 2610.06650 Wikidata Search Traces: A Dataset for Training Knowledge Graph Search Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Wikidata constraint graphs yield unique verifiable questions. A persistent Python REPL combines 13 graph-query families and preserves successful traces for training data and tool harnesses. |
| 2610.06454 AgentPrivArena: Evaluating and Auditing Real-world AI Agent Privacy | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AgentPrivArena/AgentPrivAudit runs 389 executable privacy tasks across six container services, extracts source/subject/recipient flows after reads, and applies Pass/Abstract/Block before writes. Candidate real-state evaluation and defense. |
| 2610.06406 What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. EBG deterministically maps evidence to behavior and scope, expands one hop on demand, and rechecks source text. AgentMonBench separates omitted rules, semantics-preserving substitutions, and pre-feedback prefixes. Candidate monitoring and conditional public evaluation. |
| 2610.06269 Evolving in Thought Space: Training a Small Model at Test Time Unlocks Better Discoveries | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Guidance-TTT trains a small guidance LoRA from archived parent summaries; a frozen executor sees the full parent and independent verifier rewards it. Actual RL uses adaptive entropic LOO advantage and centered base correction. High-priority Evolve candidate. |
| 2610.06193 Correct Code, Broken Contributions? SWE-CC: Benchmarking Repository Policy Compliance for Coding Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SWE-CC compiles 823 contribution policies across 12 SWE-bench repositories into deterministic conditional checks on trajectories and patches. Candidate coding-agent contract benchmark. |
| 2610.06192 Copies or Sources? Measuring How LLM Aggregators Count Restated Evidence in Multi-Agent Systems | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Uses exact Bayesian tasks with countable sources to test copied-evidence weights, comparing deduplication notices, unique readings, and citation-only forwarding. Multi-agent evidence aggregation diagnostic. |
| 2610.06191 Judged Useless, Queried Anyway: Tool-Using Agents Rarely Turn Their Own Evidence Judgments into Stopping Decisions | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Accumulates retrieval failures from the agent's USELESS judgments and consolidates/stops after five consecutive misses. Fixed-step comparison and 300 new questions support budget-control evaluation. |
| 2610.06100 From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Trace2Env builds frozen worldbooks from traces and separates cross-episode knowledge, current state, and per-step memory. Applicability, schema/constraints, staged commits, and real W2R replay support environment/world-model evaluation. |
| 2610.06064 TrustMI: Causally controlling how assistants trust their users | Full experiment not reproduced; review DONE; affordable components considered separately | Google/Meta confirmed; see priority list | Accepted in review; core implementation and public verification pending. TrustMI adds all-layer BiPO steering to a frozen LM with positive trust NLL against likelihood displacement. User/assistant spans train separately; main agent experiment applies user-trained matrices to user/tool outputs. Priority Meta safety and public evaluation candidate. |
| 2610.06001 AgentSpy: Making AI Agent Behavior Observable | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AgentSpy records process-tree syscalls/network outside the microVM and constructs normalized command/file/host star graphs and weighted Jaccard. Independent observability candidate. |
| 2610.05981 Mechanizing the User's Eye: Pre-Registered Deployment of a Sabotage-Validated Fail-Plausible Observer in a Production LLM Agent Runtime | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Small reliability regression: five deterministic signals feed an LLM judge gated by verbatim evidence. Fourteen public offline cases, per-detector sabotage, fixed manifest/scorer; Evolve regression diagnostic, not general capability. |
| 2610.05833 Request Order Matters: Cache-History Sensitivity in Selective KV-Cache Reuse for Rolling Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Long rolling-context KV recomputation uses 5% token budget, document mean deviation to choose spans, and contiguous remainder. Compare token top-k and contiguous controls; evidence favors contiguity rather than document boundaries or scoring itself. |
| 2610.05778 MedicalHarness: A Controlled Evaluation of LLMs and Agent Harnesses on Medical Tasks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MedicalHarness MH-Lab varies summaries, explicit plans, and tool discovery under one execution loop, separating response rate, accuracy, traces, and cost. Public code/MedTool support harness ablations. |
| 2610.05709 Nexus: An Execution Fabric for AI Agents Across Cloud, Edge, and Devices | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Nexus persists invocations as tasks with scoped delegation, journals, device receipts, and cursor resume. It reuses committed results and coordinates uncertain side effects at the resource. Execution infrastructure candidate. |
| 2610.05697 From Token-Max to Outcome-Max: How You Use AI Determines Its Productivity | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TokenEcom supervises simulated policies by verified output per cost rather than tokens, carries reusable playbooks across generations, and distills feedback into both sides' skills. Public SWE/HumanEval/GSM8K/MATH support cost-aware Evolve evaluation. |
| 2610.05622 UndoBench: Separating Task Competence from Recovery Capability in Tool-Using AI Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. UndoBench pairs control/fault runs at identical seeds. Independent mutation logs and state oracles separate capability, recovery, duplicates, and omissions. Tool-agent fault tolerance candidate. |
| 2610.05615 DREAM: Dynamic Resolution Assignment For Multimodal Multi-agent Debate | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DREAM uses multiresolution probes and mean ANLL minus deviation for confidence, with five payload branches. It discards probe history before debate, rolls back to lower-uncertainty answers using each agent's ANLL trace, then judges. Multimodal reasoning candidate. |
| 2610.05572 Disentangling Task Difficulty from Run-Level Failure in Agent Failure Prediction | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Decomposes pooled AUROC into within- and across-task comparisons. Task-grouped CV and absolute turn prefixes separate difficulty from current-run failure evidence for budget control. |
| 2610.05563 G-CARB: Graph-Localized Conformal Agent Risk Budget for Compositional Harm | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CARB calibrates conformal risk with finite-grid correction on whole-episode harm ledgers for stopped prefixes. G-CARB exposes only prefix-visible source-to-sink paths to fixed scoring and preserves empty proposals and threshold nesting. Safety gate candidate. |
| 2610.05540 SALUS: Automated Auditing of NL-to-SQL Benchmarks through Weak Supervision of Multi-Agent Output | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SALUS uses weak SQL execution/structure/intent labels in a generative label model, then balanced high-confidence pseudo-labels train SQL-feature-conditioned reliability planes. Fused verdicts detect benchmark errors. |
| 2610.05532 DelegationBench: Measuring When AI Agents Should Ask Before Acting | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Separates permission and information requirements in Evolve evaluation: 156 scenarios and 48 matched pairs, independent Act/Ask/Request/Refuse scoring, and Tool-Loop/post-approval actions beyond judgment questions. |
| 2610.05473 Hierarchical Reinforcement Learning with Stable Temporal Abstraction for Language Model Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. STAC adjusts normalized HiPER boundary advantages for sampled KEEP/SWITCH by premature replan and stale-keep costs. Only PPO boundary tokens change; host reward, critic, subgoal, and action stay. Hierarchical agent RL candidate. |
| 2610.05398 MMPostTrainBench: Benchmarking Autonomous Research for Multimodal Post-Training | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MMPostTrainBench packages multimodal training tasks with independent Harbor verifiers. MMResearch connects media evidence, hypotheses, and interventions in a graph with global/core/archive memory. Multimodal Evolve candidate. |
| 2610.05383 Sibyl: An Efficient Small-large Model Collaboration Framework for Long-horizon Tasks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Sibyl improves local policy through hindsight skill OPD and RL, then replays failed prefixes to find cloud actions that decisively reverse outcomes. Joint trajectory, matched-state, and consultation counterfactual credit trains call_cloud and internalizes skills. |
| 2610.05342 IRSTD-Agent: Agentic Infrared Small Target Detection via Zoom-Guided Interaction Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. IRSTD-Agent uses proposal/zoom/detect/drop/refine tools, constructs target-preserving, offset-recovery, and negative trajectories from train masks, and SFT predicts tools/coordinates. Test reads native images. Active visual agent candidate. |
| 2610.05334 AgentDiscover: Autonomous Discovery with Minimal Search Scaffolding | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AgentDiscover lets a sandboxed coding agent query idea/candidate/session/lesson graphs and submit to isolated scoring. Fresh sessions retain tools/skills; a meta-agent distills traces into search guidance. |
| 2610.05300 MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MESH-Harness fixes slot interfaces and capacity, caches embeddings, and chooses configurations with full-covariance LinUCB plus mixed-start coordinate search. Every five evaluations, traces drive two-slot mutation and eviction while retaining surrogate statistics. |
| 2610.05295 Readable Before Actionable: Causal Tracing of Indirect Prompt Injection | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Readable Before Actionable uses content/format counterfactual role probes, activation patching, and a separate hijacked/resisted suite to test fixed pre-action/tool-span edits on AgentDojo. Safety steering mechanism candidate. |
| 2610.05284 Fusion is the New Mutation: Bandit-Guided Evolution on Workflow Graphs | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DAGO tracks lineage DAGs, fixes five parent combinations, caches concatenated embeddings, and chooses fusion with shared diagonal LinUCB from annealed score-softmax proposals. The LLM begins from the strongest parent; absolute child validation updates the bandit. |
| 2610.05247 templar: agentic induction and evolution of standardized radiology reporting templates from large-scale clinical corpora | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TEMPLAR persists templates and anatomical/diagnostic provenance. Span-Triple BM25/embedding graphs and Leiden induce slots; a structural reducer constrains revision; external grounding and structured training feedback calibrate granularity. Medical multimodal mechanism candidate. |
| 2610.05176 AECG: Asymmetric Experience Consolidation and Governance In Multi-Agent Systems | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AECG scopes records by team/event/agent, treats successes and attributable failures asymmetrically, and combines slow accumulation with fast discounted Beta-style belief. Budgeted reviews depend on downstream exposure; narrow/repair versions activate only after paired failure and disjoint success replay. |
| 2610.05162 Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MemAdapter builds counterfactual tasks from frozen retrieved memories to learn when to use them, then reflects on their inferential role and answers from evidence. Upstream memory/retrieval unchanged. |
| 2610.05132 SpecAgent: Empowering Program Verification with Agentic Synthesis of Formal Program Specifications | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SpecAgent plans function/loop dependencies and uses ACSL RAG for annotations. Frama-C/WP feedback repairs upstream contracts, while two same-input calls test return equality and detect weak specs. Public C/SV-COMP support formal-method evaluation. |
| 2610.05124 Memory Canonicalization: A Framework and Benchmark for Cross-Model Drift in Persistent LLM Memory | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Disambiguates structure, extracts emotion fields, and verifies memory writes independently. CMSC-E uses a common natural-language rendering to separate semantic edits from JSON formatting. Memory drift diagnostic. |
| 2610.05107 SearchJev: A Fast and Calibrated System-1 Model for Search Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SearchJev reads legal-option token logits directly. LoRA learns with CE+Brier soft labels weighted by source, and independent validation calibrates temperatures per output type. Uncertain cases route to System 2. |
| 2610.05099 Small Agents with Semantic Search: Efficient Multilingual Code Localization | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Compact code localization uses a CPU colgrep late-interaction index, train-only gold turn-weighted SFT, then asynchronous DrGRPO/DAPO with dynamic filtering. Public SWE-Lite/MultiSWE-Flash uses base commits for localization. |
| 2610.05047 Beyond Task Completion: Measuring Interaction Cost in Terminal User Interfaces | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TUINaut exposes real TUI input/output with Docker/Tmux and independently scores state. Agent-KLM separates reasoning tokens and atomic keystrokes for interaction cost diagnostics. |
| 2610.05039 Causal Improvement Graph for Agentic Harness Optimization | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CIG persists Evidence–Hypothesis–Intervention–Outcome graphs. Observe/Plan/PlanSelect/Implement/Evaluate/Update invoke local proposers; external validation revises hypotheses before portfolio choice. Harness RSI/Evolve candidate. |
| 2610.05030 EVISKILL: Grounding Skill Evolution in Replayable Evidence | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. EviSkill creates Replayable Evidence Cards linking training spans to local edits, restores prefixes, reexecutes, and reflects. Global validation updates ValidatedSkill; after rejection, replay-supported edits stay in WorkingSkill for later checking. |
| 2610.04991 CIPO: Counterfactual Imagination Policy Optimization for Adaptive Tool Granularity Selection | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CIPO mines successful train tool chains with budgeted BPE and instantiates bindable, traceable, interruptible skills. G4 replaces atomic/skill at the first eligible state in two bases, continues the same policy, and adds bounded outcome differences before GRPO. |
| 2610.04961 Building LLM Agent Systems the Deep Learning Way: From Modular Design to Architecture Search | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LLM building blocks compose single-call, recursive history, ToT, GoT, and multi-query RAG harnesses. Reflection updates the meta-prompt. Optuna searches token/head/retrieval settings on 30% train, freezes for 70% test; public GSM8K/StrategyQA/TSP support evaluation. |
| 2610.04940 Software World Models: From Consequence Prediction to Decision Value | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SWM restores training worlds for counterfactual changes, learns downstream blast sets with SFT warmup and GRPO, and uses sampled probabilities to decide transfers and paid checks. Open repositories and independent synthetic worlds support verification. |
| 2610.04927 Assembling Insights for Agentic Machine Learning Engineering Systems | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MLE-InsightForge extracts competition/domain/improvement insights from source gold solutions and injects trigger-action conditions into ReAct trees. Verify source/target separation on 160 Kaggle tasks across 12 domains. |
| 2610.04921 Complex Agents, Shallow Tests: Demystifying and Enhancing Test Adequacy of Agent Harness in the Wild | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HarnessTester locates uncovered LLM-dependent code by coverage, retrieves exact constructor/fixture/mock bindings by AST, and generates branch tests preserving calls. |
| 2610.04916 PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. PreAct-Nav retrieves midterm subgoals from Navigation Memory and on-demand future evidence from a Predictive Sandbox. Frozen VLM reviews before acting and falls back when evidence is unresolved or revisions fail. UrbanNav/CityWalker support evaluation. |
| 2610.04915 Are We Measuring Scientific Intelligence? Rethinking the Evaluation of AI Scientists | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Uses null/withdrawal/flip counterfactuals verified by statistics, with EGA distinguishing old-key correctness from responsiveness to new evidence. Candidate research integrity benchmark. |
| 2610.04899 Rewrite What Matters: Adaptive Multilingual Query Rewriting for Reasoning via Agentic Reinforcement Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. mRewriter-R1 runs bounded language, structural, and semantic operators in sequence, using GRPO against a frozen downstream model. MGSM/BELEBELE/XCOPA support actual training and standard evaluation. |
| 2610.04889 ForkPilot: Self-Evolving Policy for Retrospective Search in Long-Horizon Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ForkPilot trains a task-bootstrap ensemble from search/skip completions restored from identical states. Cost, support, and calibrated margins choose commit/probe/skip. Freeze within tasks and update from completed source tasks. Six public benchmarks. |
| 2610.04868 From Memory to Guide: Spatio-Temporal Composer for Procedural Coding Memory | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Composer selects a skill with whyNow provenance, binds rules to current entities as a Runtime Guide, and reenters task requirements after completion or invalidation. Fixed-bank procedural adaptation candidate. |
| 2610.04824 Agent Behavior as Code: Efficient and Robust LLM Agents with Programmatic Specifications | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ABCAgent specifies actions as programs with unbound inputs; an executor manages scope/types while the LM designs or modifies programs. Public GAIA, WorkArena, GSM-Symbolic, and tau2 support validation. |
| 2610.04794 Knowing the Store: What a Memory Backend Must Write Down Before an Agent Can Read It | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Memory governance distinguishes active/superseded/retracted/expired state and current/past/change intent, comparing valueless state summaries, record tags, and name lookup over 444 counterfactual store versions. Public code/data support faithful monitoring. |
| 2610.04779 Asymmetric Repository Lineage Modeling and Verifier-Guided Coordination in Concurrent AI Coding Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MergeGym separates overlap of authored files at open time, actual Git merge-tree conflicts, and scheduler replay. T1 uses 229 held-out repositories, T2 has 5,936 replayable cases, and scheduling covers 97 repositories. |
| 2610.04699 Not Self-Decidable: LLMs Cannot Draw the Boundary of What an Agent Verifier Can Check | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CoVer treats model corroboration only as nomination. Fact-slot rules pass structure, grounding, invariance, and sensitivity gates; infrastructure-attested fields are required and anomalies fail closed. |
| 2610.04687 Understanding Errors in LLM-Based Question Answering over Imperfect Tables | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. GBDI permutes table rows into five independent discovery contexts, maps rows back into a union ledger, and chooses Edit/Drop/No-op from verifiable evidence. RADAR-T: 313 examples, 53 tasks, 27 tables. |
| 2610.04658 RETRACE: From Entangled Repair Histories to Reusable Experience for CI Repair | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RETRACE combines backward attribution from finished CI fixes with forward development history to extract three levels of problem experience. CI-Repair-Bench has 565 PRs across 101 repositories, earliest 30% history and 408 later targets. |
| 2610.04646 SEIS: Self-Evolving Inference Systems | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SEIS gives agents private code worktrees, fixed correctness/throughput scorers, and budgets, carrying forward verified inference engines/kernels. Public Qwen and actual GPU workloads support code evolution verification. |
| 2610.04641 Back to the Future: Regressing Readability Features from LLMs | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. BTTF combines AST-tagged embedding geometry/clusters, frozen code-LM construct surprisal and 32/256-token context differences, and standard features. Cross-model consensus selects 11 features; equal-dataset Ridge (alpha 200) predicts quantile readability and signed contributions. |
| 2610.04611 Practical Runtime Enforcement of First-Order Temporal Requirements | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. EnfFlash compiles executable MFOTL fragments to tabular EF causal/suppressive enforcement. Compare reference interpreter and actual traces across 40 formulas and agent environments. |
| 2610.04604 Anticipating the Consequences of Curriculum Decisions with Large Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LLM curriculum builds feasibility and cross-task relationships from all Craftax256 goal text and combines them with learning-progress history. Public environment and complete goal-success checks enable training validation. |
| 2610.04575 From Probe Scores to Alarm Policies: Operational Validity of Activation Monitors for Language-Model Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Aggregates any-alarm maxima by semantic task/case and isolates fit/calibration/test. Independent negative-unit order statistics and a Beta upper bound select maximal certified rank, using strict exceedance and abstention for inadequate support. Compare observable/activation/joint at one horizon; hashes bind model, threshold, and one-time test. Fail-closed claims prevent AUROC becoming a deployment guarantee. |
| 2610.04555 \(\mathrm{TRIZ}^{a}\): Guiding Agent Evolution from Pattern Recognition to Solution Invention | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TRIZ-a freezes contradictions/comparison protocols, declares 40 typed mutation contracts, executes candidates, and selects on useful/harmful functional information against fixed references. Three public cybersecurity datasets. |
| 2610.04517 EvoCast: Reliable Autonomous Research Agents for Iterative Forecasting Architecture Evolution | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. EvoCast separates LLM hypotheses/code edits from deterministic evaluation and promotion. Start with baseline and executable ablations, then bounded source mutations. Public TFB forecasting supports real runs. |
| 2610.04488 Can LLM Agents Automate Reinforcement Learning for Text-to-Speech? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TTS agent searches LM-carrier DPO and flow-carrier GRPO over CosyVoice2, then combines complementary strategies. AISHELL3/LibriTTS with fixed 5,000 conditions and Seed-TTS-Eval support training and external evaluation. |
| 2610.04432 Video2World: Benchmarking Coding Agents for Interactive World Modeling from Embodied Videos | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Video2World treats video-to-interactive-simulation as agent coding and separates executability, geometry, dynamics, and function oracles. Public 222 examples and scoring interface support evaluation. |
| 2610.04429 Asking Earns Nothing: Scoring the Decision to Act in BFCL Multi-Turn | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Pairs BFCL base with miss_func/miss_param and locates one empty-gold anchor mechanically. Structured rollouts distinguish world-changing attempts from lookup, scoring balanced act/hold accuracy. Filters invalid keys, multiple anchors, and mismatches, then human-certifies 223 pairs; strict and separate authorization-credit scores remain distinct. |
| 2610.04409 Understanding and Mitigating Hallucination Escape in Tool-Using LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. EscapeGuard learns a conflict probe from frozen LM states and biases tool name, description, and schema attention by conflict strength during selection/arguments. Public BFCL/Seal/Glaive support evaluation. |
| 2610.04379 AgentPersonaBench: Benchmarking Persona-Driven User Simulation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AgentPersonaBench uses 2,460 survey/chat/web/app tasks to check persona behavior, with leak tests, independent human audit, and separate deterministic and LLM criteria. |
| 2610.04375 Do Tool Calls Execute as Intended? Measuring and Repairing Intent-Execution Correspondence in LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. IEC uses the receiver parser at every hop to detect the first intent-to-execution mutation and confirms via single-hop repair replay. IntAct builds mutation-safe wrapper/process transport or rejects. Public synthetic payload/CLI tests. |
| 2610.04371 Functionally Equivalent or Not? Graph-Grounded Differential Surrogate Execution for Code Equivalence | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. FEAgent proposes executable counterexamples from repository obligations and blinded two-program surrogate traces. Deterministic reconciliation returns Equivalent/Inequivalent/Unclear; EquiBench/SWE-bench Verified support secondary audit. |
| 2610.04292 LMBuild: Evaluating LLM Agents for Generating Buildable and Functional Structures | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LMBuild generates 3D structures using part retrieval/generation tools and evaluates structure, function, design, and physical construction separately. |
| 2610.04261 Playing social deduction games with reinforcement fine-tuned large language models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Social-reasoning RFT shares policies by team in Werewolf/Avalon and prompts only private-visible roles and public history. Reward hidden-role prediction and later vote influence; final win is external. Use an executable game engine. |
| 2610.04196 Asynchronous Is Nearly Free for Evolution Strategies on Long-Horizon Agentic Tasks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AsyncES records rollout lag by parameter version, accepts bounded staleness, normalizes rewards within accepted cohorts, and updates central ES weights. Workers restore unperturbed weights and replay updates before perturbing. |
| 2610.04178 SHarP: Saliency-based Pruning of Agent Harnesses | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SHarP disables hand-defined harness modules one at a time and uses relative development median score/cost and paired t-tests for salience. Protect significant performance contributors, prune by efficiency, and evaluate independently. |
| 2610.04134 Agentic Resource Allocation for Batch Multi-Objective Bayesian Optimization in Autonomous Materials Discovery | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MultiTaskGP allocates batches among six qEHVI/diversified qUCB ratios. LLM chooses from recent HV/MI, history, and budget events; Adaptive also selects β. |
| 2610.04008 SkillScriptBench: Benchmarking Self-Evolution of Executable Agent Skill Packages Beyond Markdown | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SkillScriptBench checks SKILL.md/script consistency. AST-guided revision maps requirements to unique nodes, makes local edits, and preserves correct behavior. Executable skill/RSI candidate. |
| 2610.03994 Solving VeriContest with a Lean-Backed Rust Verifier | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Rust-Prover converts Verus to Rust specifications and then Lean via Charon/Aeneas. The proof agent iterates under fixed code/theorem, and clean clones check byte identity and Lean axioms. |
| 2610.03938 MLLMs Fail to Refuse when Using Tools Agentically | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Pairs one MLLM with no tools or four ReAct visual tools to measure refusal failures. Tool counts diagnose dilution; injecting the original request each turn is a mitigation comparison. |
| 2610.03894 Proxy Confidence: Auditing Black-Box LLM Agents with a Surrogate's Log-Probabilities | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ProxyConfidence uses independent open surrogates with teacher-forced/PMI and YES-NO/tool competition log probabilities. Calibrate directions per regime and combine scores for audit/correction. |
| 2610.03820 From Requirements to Attack Trees: Grounded LLM Agents for Design-Time Security Review | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Builds entity graphs with trust boundaries from requirements and diagrams, assigns attack-tree workers by target, jointly verifies, and rebuilds only affected branches while preserving IDs for up to three attempts. Public Microsoft IoT Edge is a reference safety task. |
| 2610.03796 ProsaBuddy: Assisting Mechanized Real-Time Schedulability Analysis with LLM-based Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ProsaBuddy decomposes lemmas with domain guidance and isolated Rocq subagents. Separate OS AST checks allow only ProofStep/Query in proof bodies and trusted imports outside, with read-only libraries and fixed theorems/axioms. 17 successes, 12 without skills, 14 without decomposition, 15 without paper hints. Headline human-hint runs are hint-assisted, not gold-free. Audit TCB beyond the kernel is unverified and weak models omit some hard cases. |
| 2610.03780 OceanMind: A multi-agent AI system for ocean diagnosis | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. OceanMind routes queries through six roles, skill planning, deterministic tools, and evidence synthesis. A restricted CodingAgent fills gaps; artifacts move by references and role-specific memories. |
| 2610.03631 NeutronGym: Physics-Graded Neutron Instrument Design for LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. NeutronGym uses 22 MCP tools for actual McStas instruments and caches compiled parameter simulations. Syntax/runtime/structure/science rewards and programmatic shortcut admission precede GRPO. |
| 2610.03591 HazardWeaver: Scientific Route Selection for Hazard Analysis Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HazardWeaver compiles literature into Satisfied/Violated/Unresolved conditions and finds executable routes in a typed capability graph. The LLM chooses/repairs at their intersection; hidden route, timing, provenance, and budgets score separately. |
| 2610.03585 Threat-Preserving Representation Sensitivity in Agent-Security Benchmarks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TPRS changes only tool representation under fixed tasks, harm objectives, policy, scorer, and environment; report ASR and benign utility separately on ASB/MCPTox/AgentDojo. |
| 2610.03574 HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HyperBrowseComp has 423 human-authored multilingual/multimodal evidence-chain short answers filtered by no-internet baselines. Match 25-step ReAct across provider-native/Exa/OWL; retain real browsing traces and short-answer judgments. |
| 2610.03524 From Benchmarks to Production: A Text-to-SQL System for Complex Financial Data | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. FLINT resolves opaque IDs from reference tables, retrieves expert query structures, links foreign-key chains, produces SQL, and reflects on real execution errors, with parallel substeps. |
| 2610.03448 Passing the Test You Trained On: Re-evaluating Prompt-Injection Detectors for LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Reevaluates 15 injection detectors using clean/injected differential replay, default and low-FPR TPR, training-overlap audit, and task-aware yes/no log-probability judgments for missing context. |
| 2610.03315 Lightweight, Rubric-Guided Trajectory Evaluation for Production AI Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LiteTrajEval uses offline domain rule profiles, online trace normalization and weak failure hints, compaction and duplicate references, tiered budgets and MMR evidence, with one rubric judge for diagnosis. |
| 2610.03226 D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. D2K-Bench pairs 26 tasks and 85 workloads with/without expert L1–3 guidance. Agents edit, compile, profile, and submit Triton. Only candidates passing numeric, state-update, and validity checks compete by geometric mean speedup. |
| 2610.03213 Toward SLM-based agentic task-tool intent matching | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Gemma task–tool relevance follows Base→GEPA→LoRA SFT→DrGRPO with deterministic format, label, and length rewards and per-tool Boolean relevance. |
| 2610.03102 Ask, Relax, or Act? Evaluating Actionable Indeterminacy in LLM Preference Reasoning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Ask/Relax/Act intersects accepted actions over admissible preferences and objectives to choose act versus ask. Infeasible cases get minimum allowed repair. Exact solvers generate paired interventions across four finite structures. |
| 2610.03099 Beyond Single Videos: Benchmarking and Active Evidence Seeking for E-Commerce Cross-Video Reasoning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AdsCVR/AdSeek selects video frames/audio iteratively. GRPO uses correctness plus 0.1 format, filters no-tool/zero-variance groups, and resamples up to three times. Teacher diagnoses evidence/reasoning errors before targeted repair and assistant-masked SFT. |
| 2610.03089 Securing Computer-Use Agents Against Branch Steering Attacks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. COBRA preplans and commits programs under trusted inputs. BRH parses observations within fixed branch/field rules, while HTTP/MCP proxies enforce operations, arguments, and approved schema hashes. Fusion of five plans keeps legal alternatives. |
| 2610.03036 WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. WebFovea hardens visual action round trips with special-token parsing, coordinate mapping, actual select/fill readback, feedback states, five screenshots of history, bounded DOM text, and executor action budgets. |
| 2610.03020 DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DyadMem separates Capture/Update/Recall and the full pipeline. Six memory types include agent–user URAM; chronological delete/add/dedup precede terminal valid-bank retrieval and grounded QA. |
| 2610.03014 Beyond Predefined Sinks: Security-Aware Dependency Analysis for LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AgentSecGraph anchors potentially sensitive operations, normalizes frameworks, and builds typed Security-ADGs. Python AST backward tracing detects overwritten constants; JavaScript uses lexical tracing. Preserve guard and trust-boundary provenance. |
| 2610.02970 A Guideline-Augmented Multi-Agent Framework for Schema-as-Code Biomedical Named Entity Recognition | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. GAMA learns annotation rules from training gold and verifies support on train. Top-N/type guideline memories produce ranked span/type plans, schema-as-code, and two-stage grounding/type/structure repair. |
| 2610.02952 GTDD: Generative Test-Driven Development for AI Coding Agents with Adversarial Testing | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. GTDD commits candidates/config first. Informed or blind testers generate valid inputs, not expected outputs. A trusted reference supports differential checks and shortest prefix/operation reduction, with at most k feedback rounds. Fresh audit is separate. |
| 2610.02928 Discriminating Fixture Coverage in Agent-Infrastructure Verification Suites | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Discriminating Fixture Coverage freezes projection suite/hash, checks kills on external mutants, and instruments activation versus invisible pollution. Add fixtures along seven predicted input dimensions and rerun unchanged oracles. |
| 2610.02925 Positive-Unlabeled Learning for Agent Safety False Alarm Auditing | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Safe-alarm PU ranking transfers OOF role-boundary true negatives, soft-protects limited mass against false-alarm negative gradients, distills frozen reference median ranks with reliability gates, and uses hierarchical true negatives and constrained KL refinement. |
| 2610.02885 PsyEvo: A Personalized Counseling Agent That Self-Evolves at Test Time | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. PsyEvo HBSP samples from a fixed population skill prior and client Gaussian posterior. SOCA uses a dual-order agreement DAG for replies and mean-preserving ordinal delayed credit. Across sessions, normalized weighted DPO updates shared LoRA. |
| 2610.02814 VeriPy Source-Preserving Verification and Compatibility Checking for Python Components | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. VeriPy preserves executable Python AST and lowers explicit admission/environment specs to Dafny or Lean. Proof sidecars, axiom gates, hash-bound guards, differential checks, and scoped compatibility preserve evidence. |
| 2610.02740 Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction-Reality Gaps | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. PH uses shared-policy self-evaluation from state/action without privileged future. If ternary majority disagrees with outcome, stopped-gradient (1+alpha) weights RL/OPD/Combined; unknown is neutral. OLMo3-7B uses one-turn verifiers and multi-turn PRM/user simulation with matched-quality random weighting. Calibration is not guaranteed; alpha 2 can collapse to predicting failure. Count M extra forwards; no test gold in self-evaluation. |
| 2610.02710 Self-Supervised Scaling of Terminal Environments for Scientific Domains | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SWR turns 500 real workflows into 3,000 tasks with public IO/hidden judges. Typed scenarios, replay counterexamples, and exact hidden agreement filter SFT; duplicate sampling only reweights. SWR100 has new tasks on known workflows. Separate 35k development checks and 200 frozen audit cases; 2% false accept/reject. Public-output oracle selection is diagnostic only, not agent input. No locatable complete author release. |
| 2610.02687 Decoupling Memory from Context: Structured Memory for Token-Efficient Test-Time Continual Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. GraphMemory adds compact entry selection and directed evidence to ACE, with BFS bounded to two hops, 12 nodes, and 2,500 tokens. Rank by Beta(1+help,1+harm) mean; attribution only uses loaded nodes and Curator edits edges explicitly. Index grows, so not end-to-end O(1). FiNER/Formula select on validation and isolate test. Fewer tokens but queries about 1.8× slower and some accuracy lower. |
| 2610.02670 LEAP: Learning Efficient Action Proposals For LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LEAP trains a tiny draft with canonical target-action LoRA NLL, not target logits. Propose in a cloned environment, verify on the same GPU in parallel, commit only matching prefix, and hand off clone state without duplicate tool calls. OpenAGI/TaskBench do not execute tools; tau2/BFCL use local backends. Whole-task 1.10–1.63× speedups use paired bootstrap. Unsafe external side effects cannot be cloned; target authorship does not ensure identical concurrent trajectories. |
| 2610.02638 Batched Speech Decisions Without Decoding: Single-Token Supervision Lets a Frozen LLM Hear Beyond the Transcript | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DuplexJev freezes ASR/LLM, trains only connector with transcript KL and single-letter decision CE, and randomizes typed options. Shared prefix KV, block-diagonal suffix masks, and restarted positions give exact batched logits. Multilayer A does not beat last-layer B; speaker-disjoint speech and emotion training still lose 6–17 ZJU content points. 160–200ms streaming was proposed, not measured; readout component only. |
| 2610.02617 WebUIProof: Benchmarking WebUI Code Generators with UI-Agent Execution Harness | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. WebUIProof runs DOM/vision UI tests, counts build failures zero, and aggregates Full+0.5Partial. Grounding and step budget affect verdicts; roughly 80% agreement over 100 human audits is not a deterministic oracle. VisRL uses screenshot/interaction/analyzer rewards and PPO. Public 149 general tasks plus build/eval/compute_acc; original 219 include 70 3D tasks absent from the repository. Scope reproduction to 149; RL/3D release not assumed. |
| 2610.02616 VERSE: Verified Self-Evolving Optimizer for Agent Harnesses | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. VERSE adds trace minimization, dual draft verification, replay/perturbation, failure audits, and optimizer prompt/tool/hook updates to fixed-weight harness search. Split 110 train, 50 validation, 108 March test, and 107 July OOD. One evolution per row with three test repeats is not three independent searches; compute differs. Preserve gates, budget, and validation; one flip is not a fix. |
| 2610.02588 Open-Endedness Bench: Measuring Epistemic Process from Agent Records | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. OEB creates act/proposition cards, validates quotations by substring, votes three times on closed questions, and aggregates an epistemic graph deterministically. Abstain without opportunities; fewer than five not scored. Four axes stay separate from persona. 90 human audits and 119 existing runs on 12 tasks are correlational; median 652 judge calls are not cheap online rewards. Verified MIT pipeline/converters and HF scored runs. |
| 2610.02569 Pincer: Resource Authorization for Agents using a Digital Twin | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Pincer trusts full OS resource mediation and isolated twins. Trusted user state and PRF per path remove semantic names; access pattern, content, and precommitted policy all vote. PRF does not imply system-wide immunity, and history is not user authorization. Two synthetic personas with one model each; conditional ASR/BGR measure resource gates, not end-task success. Author implementation/data forthcoming, not verified runnable. |
| 2610.02542 How To Train Your World Model: Fine-tuning vs RAG for LM-based World Modeling | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LM world-model study compares FT and RAG, with HWM retrieving past transitions outside the same episode for conditional NTP and optional hierarchical observation/action lookup. Frozen policy standardizes planning. Authors report 17 wins in 20 combinations; real-environment oracle and top-20 gold reranking are diagnostics only. Nonconstructed WMs win 9/20, limiting extrapolation; separate transition fidelity and utility. |
| 2610.02525 Learning What to Investigate Next: Meta-Reasoning for Long-Horizon Research Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MIRA fixes curation/execution and trains outer workorder/terminate decisions. Shared actor–critic uses discrete return-token KL, decision GAE, and cross-version bootstrap bridging; off-range ratios mask policy gradient but keep squared log-ratio penalty. Separate buffers/optimizer steps. Four environments with proxy/gold separation. Critic initialization confounds sharing ablation; inner tool/experiment tokens are untrained. |
| 2610.02523 Hypothesis-guided discovery of cognitive algorithms via program refinement | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Starts from human FlipPy: Game proposes revisions, Coding implements, Audit repairs up to five times, and Inference evaluates likelihood. 846 candidates, 733 executable, 24 semantic-cluster representatives; not autonomous theory discovery. Held-out trials were not used for selection; 72/137 improved. Synthetic recovery regained 75% of gap only 8.4% of the time. A bounded program-improvement operator is relevant, but anonymous OSF has no locatable link. |
| 2610.02480 MEA: A Reward-Driven Multi-Agent System for Faithful Model Explanations | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MEA shares a backbone across Proposer tool policy and Actor explanations. Mask/requery original classifiers for faithfulness and use terminal GRPO advantage across both phases. Vision area λ=.3 and tool count λ=.05 limit gaming; text/tabular top 25%. Three modalities and unseen Q8/Q9. Occlusion diagnoses model behavior, not real causality. Preserve actual classifiers/XAI, not self-judging. |
| 2610.02478 Tropical Reinforcement Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Tropic builds observed state/segment graphs in deterministic resettable environments, freezes actor per round, and uses max-plus prefix/suffix to find verified paths. Frontier samples and top-L splices need full replay. Token NLL trains on best and low-coverage archive paths, no negative term or reward baseline. Four environments, two backbones, three seeds, held-out instances. No gold routes or non-resettable real APIs. |
| 2610.02472 APDMem: Agent-Controlled Progressive Disclosure for Query-Adaptive Long-Term Memory | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. APDMem prebuilds L0 summaries/L3 raw text and caches L1 facts/L2 turn notes on demand. Hybrid/RRF search and agent drilling stop at no new evidence or 20 steps, then draft→notes→answer. LongMemEvalS 500; baselines are published, not all rerun. Online cost only measured for this method and excludes shared answer generation; no equal-budget superiority. |
| 2610.02460 CUEing User Simulators: Calibrated User Embeddings for Multi-Turn Benchmarking | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CUE encodes turns and prior assistant attention, decodes slot commands, and samples with latent diffusion. Content suppression does not guarantee removal of task/outcome clues. Completed-human-trajectory replay is retrospective, while sampled personas without target traces simulate new users; mixtures are not population distributions. 483 tauUSI plus held-out SimulatorArena/PRISM. Apache 2 runtime/training and HF artifacts accessible; code unrun. |
| 2610.02425 Finding the Move Is Not Winning the Game: XiangqiBench for Closed-Loop Evaluation of LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. XiangqiBench has 119 searched and human-verified endgames, not formal proofs. Hide gold; fixed 40-ply/depth-18 defender, three consecutive illegal moves fail, simplified repetition. 8,568 traces, 12 models, two settings, three trials. Pass@3 and all-success reported separately; Sighted/Restricted prompts differ. Public MIT runner/data; Pikafish GPL separate and replaced NNUE requires frozen engine/network. |
| 2610.02404 Trained Agentic Context Management | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Trains context management with minimal read_chunk/spawn tools and oracle-synthesized SFT tree-reduce, sequential-fold, and boundary expansion; RL regressed. RULER five per type, OOLONG ten per family, native baselines differ in context. 8k agent budget can scale but cost not formally measured. Select temperatures on validation and hide oracle scripts at evaluation. |
| 2610.02396 Inherit-MAS: Test-Time Evolution of Multi-Agent Systems through Workflow and Execution Inheritance | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. InheritMAS prunes latest-incumbent DAG then edits once; final judge ranks lexicographically. Reuse requires exact resolved request, model/parameters/upstream/environment fingerprints, valid record hash, and read-only nodes. State-changing sinks always run. WorkBench/HotpotQA; retrieval budgets differ in some controls, inherited tokens are equivalent history not paid usage, and top judge rank is not monotonic true utility. |
| 2610.02361 SEDIMA: Cross-Run Hierarchical Insight Memory for Evolutionary Search Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SEDIMA writes raw evaluation→LLM insight→attention-centroid cluster; before mutations it retrieves clusters, insights, and evidence for three suggestions without changing the searcher. AlgoTune/ALE private scores are separate from public fitness and memory originates from disjoint tasks; cold also writes online. 100 candidates do not match tokens/calls; main table one run, ablation/transfer three seeds. Bounded context is not constant total retrieval cost. |
| 2610.02351 DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DeReAct Brain proposes intent/command; independent Critic fails closed by constraints; CM stores only environment-backed state and judges completion. Brain also sees first five and recent turns. GAIA 165/SWE 489 use native non-LLM success scoring; three attempts yield pass@1/3 with separate grounding judge. CM is model judgment, not formal verification or external-side-effect proof; no success-flag substitute. |
| 2610.02349 MIRROR: Multipath Quorum Integrity for LLM Multi-Agent Communication | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MIRROR creates canonical bytes once, then sends three payload and two digest routes for majority recovery. Unkeyed hashes do not authenticate; assumes honest endpoints, fewer than half compromised routes, and second-preimage resistance. AutoGen/CAMEL/MetaGPT and Gemma/Gemini share backends, not independent failure domains or measured latency. Protects transit, not an injected sender. |
| 2610.02330 Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CITA/CIM combines observed same-context successes, Bayesian synthetic graphs without real APIs, and LLM simulations to train value BCE, gap MSE, transitivity, and confidence. Vc ranks tools; confidence-weighted values mix with task rewards in GRPO. Toolathlon/TOUCAN/TRAJECT, Qwen 7B/Llama 8B. Value is not causal truth; mark sandbox synthetic and evaluate independently. |
| 2610.02304 SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SimuVerity covers 101 tasks in ten engineering domains, isolating public specs from hidden thresholds. Delivery, execution, and engineering qualification precede six-axis scoring; candidate failures count, infrastructure failures rerun. Six agents × three trials per task; ten strong-task ablations are selective. Apache 2 code, scenarios, scorers, references/import recipes, and results verification are public. Some tasks need commercial MATLAB/Simulink and asset licenses. |
| 2610.02204 Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RPG rebuilds MuJoCo tasks from ABC video and freezes a human-reviewed evaluator. Runtime sees deployed observations only; independent privileged episodes diagnose failures. Implementor revises skills/prompts; paired five-seed, 22-task gate requires mean gain and no task drop over .2, with post-merge retest. Fifteen rounds in one run and unequal candidate compute. Saved-round holdout seeds are not unseen tasks; three physical tasks × ten trials with frozen calibration. |
| 2610.02202 ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ScholarCatalyst uses 207 projects and 894 queries (207 Core/687 Sub), author-verified inspirations/hard negatives, leakage removal, and completion-time corpus cutoffs. Recall labels incomplete; oracle reranking diagnostic only. Public 191k documents, evaluation, and local-corpus search scripts under CC BY-NC 4; evaluation rather than new loss. |
| 2610.02161 DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DuoMind pairs per-robot VLM orchestration and VLA actions, exchanging intent, subgoal, belief, and optional uncertainty with no special joint training. Seven RoboPoly plus eight RoboTwin adapted bimanual tasks, 50 dual demonstrations per task converted into 100 single-robot examples. Global and wrist cameras; not a large physical swarm. Communication and π0 ablations support modularity, not scale claims. |
| 2610.02150 From Knowledge Access to Source Learning: Developing Source-Specific Competence | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SourceLearn maintains source-grounded entities using Inspect/Deepen/Connect and consolidation. Task-guided learning uses training answers only to diagnose missing evidence and reread source, updating representations/prompts rather than weights. Five tasks with three 30/70 guidance/test splits and official AppWorld split. Test answers cannot enter memory; separate QA judges and native execution. |
| 2610.02005 Counting Moves, Weighing Voices: Bayesian Dialectical Argumentation for Calibrated Multi-LLM Councils under Persistent Adversaries | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. BDA converts Propose/Challenge/Concede into signed endorsement counts and integrates Beta reliability in Dawid–Skene-style aggregation. About 25 labels for matched-budget calibration; K>2 pseudolikelihood needs tempering, and theory covers binary no-challenge assumptions. Stable identities can fail in positional MCQ; sleeper attacks and reliability shifts are not guaranteed. Aggregation mechanism, not universal robustness. |
| 2610.01769 CONTRA: Discovering and Qualifying Behavior-Changing Questions for Selective Clarification in LLM Code Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CONTRA identifies relevant clarifications while the original need is unresolved. Per possible answer, sample three programs and ≥7 shared inputs; require 2/3 within-group stability and a cross-group difference. Failure is not equivalence. Select questions or stop adaptively. 419 underspecified and 80 complete tasks, up to five questions, GPT-4o three-vote matching. Four-model macro F1 41.20 versus 27.32 with isolated execution. Plugin deployment is not production A/B. |
| 2610.01569 Managing Context and Communication in Distributed Agentic UAV Swarms | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. UAV gossip uses peer-tagged residual embeddings and recency novelty, with offline double thresholds for discard/compact/full. Receive queue only estimates neighbor knowledge. Three token FIFOs and ID/path/TTL dedup. Ten-aircraft SITL search/rescue with analytical 802.11n, not real networking. Proposal all complete versus flood 70–85%; later timing/error conditioned on success. Do not only report 6–7MB to <0.5MB savings. |
| 2610.01564 Chaining Skills to Hijack LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. APEX cross-skill artifacts can incorrectly promote progress into authorization; isolated execution feedback optimizes chains and defense uses taint plus original requests. 690 attempts, four targeted families, 512 successes; families overlap, utility is weighted checks, and extra workloop tokens alone are not unauthorized. MIT execution/trace/variants public, SkillsBench separate. Defensive sandbox only, no external attacks. |
| 2610.01508 OverAct: Measuring and Mitigating Proactive Over-Authorization in LLM Tool-Calling Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. OverAct uses conservative literal-minimal tool sets, not universal authorization law. Eight domains, 48 tools, 240 double-annotated originals expanded to 720 paraphrases, set scoring without judges. SelfAudit combines request-grounded justification and pre-execution filtering; filtering drives gains, justification alone regresses; TCR .94→.89. Logs mentioned but no public download verified. |
| 2610.01506 MCRI: A Four-Dimensional Framework for Analyzing and Evaluating Agent Skills | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MCRI rates Metadata/Constraints/Resources/Instructions 1–5 then aggregates by task, selecting skills by mean score × mean confidence. Three public benchmarks, 50 subtasks, 345 Mind2Web steps, 58,275 calls. Correlation/popularity not causal quality. Appendix D supports prescreening, but Appendix F explicitly omits Resources and no full repo/data reproduction is claimed. |
| 2610.01023 Groundability, Not Scale Alone: When Weak Reviewers Can Audit Strong Coding Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Reviewability separates testimony, structured, and independently grounded evidence. Frozen cascade checks empty patches and new static errors, generated tests that fail on base, then 8B review/abstention. Hidden tests are an upper bound, not deployable. 32 Django design cases, 121/122 GPT and 59 Gemini one-shot tests; risk .33/.26, over-reject .66/.67. Generated passes are not proof; no blanket weak-to-strong or cost claim. |
| 2610.01017 Pay for the Fault, Not the Flow: Label-Free In-Flow Multi-Agent Workflow Optimization | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. InFlowOp sets cost to negative log rubric match, decomposes top-down and merges bottom-up, then detects decomposition/assignment errors via declared IO contracts. Swap agent before redecomposing, rerun affected descendants, cache rest. Semantic match is not an oracle; rollback requires trustworthy checks/reversible effects. Braid has eight domains, three closed models only five; tool-less single-LLM baseline weak. Match turns and isolate profiling from test. |
| 2610.00979 RISED: RubrIcs for agentic multi-environment Selection and sElf-Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RISED uses fixed rubric vocabularies (82 across four environments, 59 in main three) and asynchronous Qwen 32B labels. Round-robin quotas and MMR select 64/128 groups; use positive/negative rubrics to select, but teacher only top-three positive union, then k3 OPD with importance weights plus GRPO. Negative rubrics steer. Qwen3/4B on three environments gives only +1.2/.6 pp versus strong baseline. Half optimizer steps is not half cost; OPD signal persists in 62.9% zero-variance groups; no-distill also removes negative steering. |
| 2610.00977 ABSENTIA: Detecting Broken Access Control Vulnerabilities in Web Applications | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ABSENTIA profiles apps, unions two LLM pattern extractions into routes, checks invariants/enforcers and source-grounded counterexamples, and composes across shared files. Byte-identical dependency closures cache across commits. BAC30 covers 30 disclosed bugs in 25 projects; advisories/fixes oracle-only and fixed negatives not whole-repo safety. Independent human precision audit of 73 did not confirm judge gains. Defensive static mechanism; code forthcoming. |
| 2610.00885 FORALL-LEAN-AGENT for Auditable Reasoning in Formal Mathematics and Software Verification | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Forall-Lean-Agent isolates actor and fresh read-only reviewer, requiring dual approval on the same digest. Proof-bodyless retrieval, process-tree sandbox/provider-only proxy, syntax/shadowing/axiom audit, then serial clean checks. 100 VSB/672 Putnam and two LeanEval suites; denominator differences and legacy reference leaks/file drift matter. Unequal cost/compile/review budgets preclude single-component attribution. |
| 2610.00710 ReLiveGym: Evaluating Long-Lived Agents over Weeks of Replayed Reality | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ReLiveGym replays weeks of real streams over eight tasks. Only waits advance clock, not model latency; timestamp-filter observations, deterministic timed commits, hard wallet, sleep/watcher/cron and daily hindsight memory/BoN pruning. BoN uses settled past feedback only. Public code, eight scorers/config but no bundled data; DATA explains public rebuilding and bundle link is placeholder. Needs 33GB, inferred timing can err, Linux shell disabled without OS fencing. |
| 2610.00675 LabBook: Harnessing Experimental History for Efficient LLM-Driven Discovery | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LabBook maintains one free-text semantic memory and lossless log. Auto-inject latest code/result plus best/stagnation; up to five read-only History/Code/Result queries, then jointly emit full program and memory. No forced best parent. 49 FrontierCS plus nine public tasks, matched model/prompt/evaluator, ablations of memory/retrieval/full-history/top-score. Normalize per task only for aggregation and select on isolated validation. Code forthcoming. |
| 2610.00651 Agent Evaluation Reliability: More Tasks Won't (Always) Fix An Agent Leaderboard | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Bernoulli-logit Bayesian crossed variance separates models from model–scaffold deployments; π²/3 residual cannot be decomposed without repeated cells. 29,923 HAL, 54 models, 13 scaffolds. D-study posterior shows more tasks cannot remove model×scaffold/benchmark variance. Breadth conclusion assumes balanced exchangeability/setup cost; not arbitrary tasks. Public MIT R/Stan tests/outcomes; dollar costs are assumed prices. |
| 2610.00648 Incident-Arena: Getting agents to the last nine of reliability | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. IncidentArena has 20 K8s incidents and three substrates, hidden deterministic outcome SLIs and restart/reattack gates, frozen healthy thresholds, oracle/no-op/adversarial QA. 3,000 native trials, 13 Slack cases cause unequal coverage, wide reasoning CIs, and post-hoc hacking judges differ from hidden scorer. Apache 2 tasks/Harbor fork public. Linux Docker, eight CPUs, 16GB RAM, 40–77GB storage; isolate clusters/secrets, no production rights. |
| 2610.00613 Spatial Strategies, Not Actions: Vector-Quantized Geodesics as Tools for LLM-Driven Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Offline 250 geodesic five-step fragments are origin-normalized and VQ/K-medoids compress to five prototypes plus five primitives/wait. LLM describes once and selects five-step open-loop skills; collision filtering uses only observed obstacles. One Qwen-VL grid-world is not high-dimensional generalization; signed-measure skill additions only proposed. Keep actual geodesic sampling/VQ, not handwritten substitutes. |
| 2610.00435 How AI Agents Discover Scientific Equations: From Hydrotope Rediscovery to New Water-Wave Amplitudes | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. PI assigns analytical and numerical exploration to two students sharing notes, with independent recursive BG counterexample search to reconstruct global amplitude formulas across frequency chambers. |
| 2610.00400 Representation Transitions Reveal Emerging Safety Risks in Multi-Turn LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DART derives a harmful-minus-benign transition mean, removes benign mean and leading covariance directions, accumulates projections at segment boundaries, and on threshold locates the largest-contributing segment for reminders and regeneration. |
| 2610.00372 When Harnesses Lose the Signal: Causal Evaluation of Recovery in LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CIR pairs refresh/no-refresh from identical states, distinguishing rescue and harm. Three logistic models estimate anomaly and both outcomes; OOF thresholds respect clean-harm constraints, with at most one timely refresh. |
| 2610.00354 Proof-Gated Signing: Solver-Checked Transaction Guards that Hold Under State Drift for Onchain AI Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Proof-Gated Signing simulates transactions, verifies policies over price intervals with Z3, compiles balance/recipient/allowance/owner postconditions, and atomically executes or rolls back to prevent check-to-inclusion drift. |
| 2610.00313 Rules to Tools: Executable Checks for LLM Agents in Scientific Computing | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Rules to Tools compiles public equations, boundaries, and output constraints into callable checkers. Repair sees actual residuals and hidden grading is independent. Match written rules, probes, and budgets across text/tool arms. |
| 2610.00182 Localizing Post-Wire Semantic Changes in MCP Agent Frameworks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MCP post-wire differential tests use allowlisted projectors to preserve semantic atoms and compartments, separating protocol legality, typed retention, and consumer access at each framework's public carrier. |
| 2610.00097 DramaAgent: Agentic Storytelling Video Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DramaAgent uses hierarchical storyboards and fixed character references, generates per-scene video/audio candidates, scores identity/semantics/timing/audiovisual match, and revises the weakest axis with bounded regeneration. |
| 2610.00025 Measuring the Microtask Eligibility Gap: When Is an Off-the-Shelf SLM Enough for an Agent Harness? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Four microtasks use non-LLM baselines for prespecified thresholds and eligibility intervals. Select log-probability operating points on calibration, with BM25 memory shortlist then SLM reranking and tool-baseline gating. |
| 2610.00012 When Do Causal World Models Help Modular LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. FedCausalCompose matches local interventional responses, discovers edges, and imposes composition constraints. Diagnostic agent world-model candidate; learn edges rather than use oracle graphs. |
| 2610.00010 Heavy-Tailed Memory Traces in Long-Horizon Language Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CTWM allocates core/tail memory by power-law access rank. Executable compression candidate; token savings with near-zero task success are not capability gains. |
| 2510.15831 VISTA: A Test-Time Self-Improving Video Generation Agent | Full experiment not reproduced; review DONE; affordable components considered separately | Google/Meta confirmed; see priority list | Accepted in review; core implementation and public verification pending. Priority Google historical discovery: structured prompt planning, bidirectional champion tournament, normal/adversarial/meta critiques, DTPA revision and actual regeneration/reselection. Executable agent self-improvement, not a video-backbone innovation. |
See reconciliation.json for source pointers, titles, and current status.