跳转至

Seven backfilled papers from the October 2 announcements: 2026-10-04

简体中文

Three user-approved P0 and four P1 items from the October 2 announced window, not seven papers published within hours. Official arXiv full text and author links, no DeepXiv. No new full scan or complete Google/Meta rerun; coverage_complete=false, watermark 2026-10-02, overlap from 2026-10-01.

Priority Method Executable part Boundary
P0 Jev lookahead Cloned simulator, newly available actions, System One provider Diagnostic provider, not ALFWorld success
P0 T2SPO Real TabPFN, support standardization/SVD, credit and clipping Official CPU checkpoint, synthetic states
P0 Weakest-Link All prefixes, absorbing penalty, immediate full-vocabulary gradient Real small-policy update, not LLM distillation
P1 CMP Randomized retrieval exposure, Hájek and risk decisions Intervention measurement, not long-run deletion
P1 TESS Sequential two-model training, PVM selector and independent pool scoring Synthetic classifiers, not LM lift
P1 RLCPR Entropy buckets and concentrated-posterior length penalty Formula/gradient checks
P1 Mingbird Portable M1/M3/M4/M5 and real tool acceptance loop Scripted policy, not agent accuracy

Callable L1 APIs are not end-to-end reproduce adapters or registry-only Evolve integrations.

PYTHONPATH=src python scripts/run_oct04_seven_papers.py
python -m pytest tests/test_oct04_seven_papers.py

T2SPO pins tabpfn==2.2.1 in a separate environment; see its page for checkpoint revision. Three-seed MAEs 0.00290, 0.00138 and 0.00234 on synthetic learnable remaining distance are not real environment prediction or the paper's encoder/RL. Weakest-Link mean expensive-action probability 0.3546→0.01036; TESS PVM 0.0803→0.0129; CMP 100 exposures per memory. All diagnostic_only=true regardless of seeds.

Full benchmarks, LLM/hosted Jev, cross-task memory utility and large training remain on the roadmap. No new CUDA claim/A100/A30 verification. CMRec Amazon-M2 and other earlier boundaries remain.