Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 41 · 2026
Oct 05 – Oct 11
20 papers, repos and datasets selected this week.
Papers
Picked Oct 9
- PaperLong-WAM: Scaling the Context of World-Action Models→
Unlike bidirectionally adapted WAMs, it preserves causal video pretraining, enabling 19.2-second histories to improve control.
- PaperUniWAM: Unified World-Action Model→
Authors report SOTA across in-distribution, OOD, instruction-following, and long-horizon robot evaluations, plus a log-linear scaling law for human-robot co-training.
Picked Oct 8
- PaperUNREAL: Unifying Retrieval and Long-Context with a Single Model→
Best model lifts recall@10 from 49.1% to 73.2% on HotpotQA and NoLiMa accuracy from 1.0% to 24.83% at 128K.
- PaperMultimodal Flow: Unified Flow Modeling of Language and Vision in Embedding Spaces→
With 150B tokens, MF-1 gets 0.821 GenEval, 83.44 DPG-Bench, and 75.3 avg on VQAv2/MMBench/POPE; it also beats matched hybrid and discrete baselines.
Picked Oct 7
- PaperOptimizing the Optimizer: Language Models Discover Faster Molecular Relaxation→
Best variant uses 40.2-77.2% of Sella's force calls at r2SCAN-3c DFT with equal or better energy recovery.
- PaperALoDLM: Adaptively Looped Diffusion Language Models→
It allocates denoising compute per token, letting an 8B DLM beat AR baselines while decoding 2.7× faster on GSM8K.
- PaperLanguage Models that Play Chess and Explain Their Moves→
Gains 900+ Elo from 1782 to 2697, beating GPT-5.6-Sol (2071) and Gemini-3.1-Pro (2201); puzzle no-mistake rate tops baselines.
Picked Oct 6
- PaperInterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation→
Shows unseen-task humanoid behaviors, long-horizon compositions, and autonomous transfer to a real Unitree G1 without retraining.
- PaperHierarchical Continuous Diffusion Language Models→
At 6M params, HC-DLM beats CCDD on Sudoku Hard 72.41% vs 70.73% and Countdown CD5 37.52% vs 25.35%.
- PaperDoes Learning Protein Folding Generalize to Broader Reasoning?→
Raises General-10 macro from 45.09% to 48.33% (+3.23 pp) with gains on all 10 benchmarks; folding scores are 2.7-3.5x Qwen3.5-9B.
Picked Oct 5
- PaperOneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction→
It unifies proactive response timing and reusable caption memory in one causal decoder.
- PaperThe Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation→
Across 4 base/RL-teacher pairs, RIDE is the only method whose mean matches or exceeds the RL teacher on all pairs, beating OPRD by 0.97-4.06 Avg@16 points.
Repos
Picked Oct 9
- RepoBQ-Wang0511/TalkLikeYou-Streaming-RealTime-Demo→
Runs at 30+ FPS on RTX 3090 with under 12 GB VRAM, and supports live mic, video dubbing, and streaming chat-driven animation in one demo.
Picked Oct 8
- RepoTier-Flow/LoopVL→
Unlike most model repos, it releases reproducible recurrent VLM evaluation and visual-state analyses across 16 benchmarks.
Picked Oct 6
- RepoPSRben/VisionHOPE→
Unlike most vision backbones, it uses self-modifying memory updates and ships checkpoints across classification, detection, and segmentation.
Picked Oct 5
- Reponokia-applied-research/AnyJev→
Unlike recent typed-decision repos, it retrofits open LLMs into calibrated classifiers from one prefill, boosting BANKING77 auto-decidable coverage 7.7%→52.0% at ≤5% error.
Datasets
Picked Oct 9
- Datasetgoodpjw2008/rsna-knee-2-5d-convnext-reader→
Releases end-to-end DICOM inference code and weights for a 0.929 public-LB knee MRI baseline.
Picked Oct 8
- Datasetilluin/daedalus-traces→
Makes agent-memory pipelines auditable and reproducible end to end, with traces linking bank generation to downstream evaluation.
Picked Oct 7
- DatasetScholarCatalyst/ScholarCatalyst→
Unlike topical paper corpora, it benchmarks inspiration retrieval with 894 author-verified positives, hard negatives, and rationales.
Picked Oct 5
- Datasetallenai/fetchman-data→
352,696 humanoid pickup episodes jointly align language, proprioception, actions, and head+wrist egocentric video.