Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 40 · 2026
Sep 28 – Oct 04
20 papers, repos and datasets selected this week.
Papers
Picked Oct 2
- PaperFalse Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents→
Across 7 search benchmarks, CrossFit beats coupled self-evolution by 8.8/8.4 pts and Search-R1 by 8.7/7.8 at 4B/9B.
- PaperContext Language Models→
Zero-shot CLMs beat SOTA context management: +11.4% acc and -21.5% FLOPs on BrowseComp-Plus; +5% and -59% FLOPs on EdgeBench.
Picked Oct 1
- PaperTabFM: A Zero-Shot Foundation Model for Tabular Data→
A 400M model trained only on synthetic SCM tables beats tuned AutoGluon zero-shot.
- PaperWhat Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling→
Simple-WAM gets 79.5 perturbation success vs 67.7 explicit and 53.8 latent, while running 3.8× faster than explicit and within 1.2× latent latency.
Picked Sep 30
- PaperYuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality→
Unlike end-to-end audio models, it explicitly plans readable scores before full-song rendering.
- PaperCoWindow Attention: Full Causal Coverage Is a Collective Property→
Unlike routed sparse attention, it guarantees full causal coverage collectively across heads, preserving 8K recall while cutting 128K train latency 7.4×/8.6× and decode 3.0×.
Picked Sep 29
- PaperFuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders→
On ImageNet-256, decoder swap alone cuts unguided gFID 3.01→2.21; joint regularization cuts DiT-Base gFID 13.96→9.93.
- PaperInternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data→
Best reported success on LIBERO-Plus 92.8%, RoboTwin C2R 71.9%, C2C 90.0, EBench 66.0, and RoboDojo 23.9%.
Picked Sep 28
- PaperTraining Object Permanence in World Models→
On a 300-question human-rated exam, PWM-WROP ranks 1st among continuation models and 3rd overall; at matched resolution it also gets the best LPIPS and MS-SSIM.
- PaperAll-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts→
On TextMuSS-Bench it reaches 82.06% accuracy, +1.31% over the strongest STR baseline. In CC-OCR, F1 rises 65.71%→80.89%, beating the best VLM's 80.73%.
Repos
Picked Oct 2
- RepoPostHog/jeeves→
Unlike single-pass decision models, it adds reasoning to a Jev-compatible 9B classifier, raising JevBench from 0.866 to 0.935 while keeping calibrated outputs and optional 0.3-second no-think inference.
Picked Oct 1
- Reponinjahawk/livenerf→
Unlike static benchmarks, it append-only tracks post-launch model drift daily with pinned prompts and pre-registered statistics.
Picked Sep 30
- Repolyuyiqi/open-jev-fast→
Cuts JevBench mean latency from 258 ms to 42.0 ms on 1× B300 with near-identical accuracy (197/231 vs 198/231).
Picked Sep 29
- Repojev-chat/jev-chat-jarvis→
Works without root, hooks, or app APIs; supports app-specific adapters plus offline OCR fallback, and never auto-sends messages.
Picked Sep 28
- RepoContrastive-LM/CLM→
Unlike typed-decision classifiers, it scores state-action pairs for routing, ranking, and verifiers at up to 9× lower latency than Jev.
Datasets
Picked Oct 2
- Datasetdmitriigluzdov/casmi26-attributed-v4g-reproduction-assets→
Makes the published CASMI26 pipeline reproducible at file level by pinning exact inputs, paths, sizes, and SHA-256 hashes.
Picked Oct 1
- DatasetUW-Madison-Dept-Radiology/PETARSeg-11K→
Unlike typical radiology datasets, it links 11K PET/CT lesion localizations directly to aligned free-text findings.
Picked Sep 30
- Datasetmxlin043/EmbRACE→
Unlike navigation datasets, it pairs 48,264 egocentric steps with verified first-person rationales and closed-loop state-based evaluation in seven unseen environments.
Picked Sep 29
- DatasetMykolaL/FBIS-73M→
It pairs 74.0M field-instance masks with a 100-country zero-shot split for cross-country generalization.
Picked Sep 28
- Datasetallratestoday/central-bank-exchange-rates→
Makes institution-issued FX rates measurable across countries and decades for compliance-grade analysis, without mixing in inferred or market rates.