Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 39 · 2026
Sep 21 – Sep 27
20 papers, repos and datasets selected this week.
Papers
Picked Sep 25
- PaperPACT: From Credit Assignment to Critic Alignment→
PACT gets 72.87% avg on 4 math benchmarks, +8.80 over GRPO and +13.16 over PPO; 67.4% on SWE-bench Verified, +2.0 over GRPO.
- PaperRecursive self-improvement of AI research agents→
In 8 days it found 7 accepted rewrites, raised grade 0.703→0.778, matched/exceeded a human-built agent on 4 held-out benchmarks, and cut reward hacking 55%→32%.
Picked Sep 24
- PaperThe Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks→
Best model gets only 59.7% on 502 questions; longer reasoning does not help, and distilled students improve decisions on unseen tasks and SWE-bench Pro success.
- PaperGAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation→
With the same generator/training, FVD drops 12.7% on RealEstate10K and 23.1% on DL3DV; camera-trajectory error is roughly halved on RealEstate10K.
Picked Sep 23
- PaperRRSI: Regularized Recursive Self-Improvement of Agent Harnesses→
Across 8 benchmarks, RRSI gains up to +14.1 on evolve splits, +4.7 OOD, and uses 30% fewer policy tokens than unregularized evolution.
- PaperGrounded Action Model: 3D Grounding as a Foundation for Robotics→
55.3% on RoboTwin 2.0 vs 52.0% Spatial Forcing; 61% on LIBERO-PRO vs 53% π0.5; 17/20 real-robot visual-shift successes vs 4/20.
Picked Sep 22
- PaperCodeMidas: Scaling Agentic Coding RL Environments from Code Itself→
Unlike issue- or test-derived coding RL datasets, it extracts 5,545 executable tasks directly from source code.
- PaperRecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents→
It uniquely benchmarks hybrid agents by requiring live app inspection, code recreation, and hidden behavior-plus-visual verification across five platforms.
- PaperConfidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents→
Beats or matches 10-sample self-consistency in AUROC on 23/24 comparisons, with lower ECE at 1/10 the generation cost.
Picked Sep 21
- PaperJEPA-Anything: Learning Predictive Models across Different Worlds→
Unlike monolithic JEPAs, it factorizes latent targets into orthogonal predictors across seven domains.
- PaperWhen EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation→
Across Qwen3, Llama 3.2, and Gemma 3, semantic EOS handling substantially mitigates length inflation; decoding-stop-set alignment alone does not.
- PaperLimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence→
Beats dataset-specific models and tabular FMs on TabArena, TALENT, and BCCO; reaches Elo 1935, 1506, and 1432.
Repos
Picked Sep 25
- RepoTencentARC/GameHorizon→
Covers 5,000 gameplay hours from 21 AAA titles, 5,000 offline MCQs, 20 online tasks, and benchmarks 47 models in one suite.
Picked Sep 24
- Repojaredpalmer/kev→
Unlike recent typed-decision repos, it packs many question types into one calibrated, decode-free pass with 2.0× speedup and 4e-6 output parity.
Picked Sep 23
- Repomizorewww/laya-mlx→
Ports typed decision inference to Apple Silicon MLX, matching 63/63 outputs at 7.4-13.4 ms locally.
Picked Sep 21
- RepoNandhaKishorM/laya→
Unlike similar single-pass classifiers, it routes across English and multilingual checkpoints, covering 45 languages at 7.2 ms/question batched.
Datasets
Picked Sep 25
- Datasetmicrosoft/ProgramDistill→
Unlike GUI-agent benchmarks, it evaluates coding agents by restoring masked web-app features with replayable browser traces and executable verifiers across 4,063 tasks.
Picked Sep 24
- DatasetHarland/OmniVChat→
Makes native audio-visual dialogue measurable without text queries, ASR, or captioning, with fine-grained rubric scores shared across eval and RL.
Picked Sep 23
- DatasetEgoSteer/EgoSteer-RealWorld→
54,454 real-robot egocentric bimanual episodes across 193 tasks pair RGB-D, 74D actions, and language.
Picked Sep 22
- Datasetsecemp9/arxiv-complete→
Unlike paper-text corpora, it combines 3.15M arXiv papers with source files, resolved TeX, and 5.03M-version history in one auditable 16.08 TB snapshot.