Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 44 · 2026
Oct 26 – Nov 01
20 papers, repos and datasets selected this week.
Papers
Picked Oct 30
- PaperCan Discrete Tokens Serve as Semantically Lossless Representations of Visual Signals?→
FVQ uses symmetric parallel full-feature codebooks to avoid high-bit saturation, slightly surpassing continuous QwenViT.
- PaperHy-Nomas: Towards Free-form Interactive World Generation→
Unlike other world models, it supports multi-turn free-form video interaction by separating persistent scene memory from local action commands.
- PaperRollback: Experience Backtracking in World Models for Policy Improvement→
Beats the strongest baseline by +10.67%, +11.11%, and +11.85% success on 3 real robot tasks; exceeds average human success on 2 tasks within 8 hours.
Picked Oct 29
- PaperThe Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents→
It clusters tasks by shared capabilities, co-evolving simulators and reusable modules for 34-point gains on 28 benchmarks and 26 points on 18 unseen tasks.
Picked Oct 28
- PaperRynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions→
A 4B policy reaches 98.8% on LIBERO, 84.7% on LIBERO-Plus, 51.9% on VLABench, and 88.6% on RoboTwin 2.0.
- PaperTrajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents→
Across five Qwen3 sizes, it improves Tau1, Tau2, and BFCL at every scale, raising the 3-benchmark mean by 16.9-29.8 points.
- PaperTrial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution→
Across 3,420 runs, harness scores span 59.8%-72.0%, and 31.1% of runs pass the task while failing at least one harness criterion.
Picked Oct 27
- PaperBuilding AI Research Teams for Autonomous Discovery of Linear Attention Architectures→
The discovered CFDN beats GDN2 at 1.3B on language modeling, common-sense reasoning, and synthetic and real-world retrieval; removing any team component weakens search.
- PaperScaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds→
Fits CPT loss with a 3-factor scaling law (R² 0.982 Minecraft, 0.995 in-corpus); mixed data gives ~2.8× OGA over Minecraft-only at 1,600h.
Picked Oct 26
- PaperLaLa-World: Scaling Latent World Models with Natural Language Actions→
Best average MPC planning among LWM baselines: 72.1% and 50.5%. Matches a generative baseline with 6x less training cost and 13.9x faster planning.
- PaperScaling Parameter-Efficient Fine-Tuning for Brain Encoding→
Best published Algonauts 2025 single-model score: 0.327 ID, 0.228 OOD; 261-model ensemble reaches 0.343/0.248.
Repos
Picked Oct 30
- RepoMUSE-AI-Lab/SurveyGraph→
Graph-aware retrieval nearly doubles recall on a core diagnostic (p=0.004), and 80.6% of generated sentences have full or partial citation support.
Picked Oct 29
- Repomuhammad-talha-ad/LOCUS→
Unlike other grounding repos, it exactly recovers greedy coordinate-token boxes in 5.6 passes, running 1.5× faster than vLLM on identical weights.
- Repogoogle-deepmind/detectorscope→
Unlike black-box AI text detectors, it exposes named feature sums and token-level evidence while preserving accuracy above 0.7 interpretability on COLING and RAID.
Picked Oct 27
- RepoTencent/WeVisOne-VL→
Beats strong open baselines across tasks; e.g. 71.3 Dense200 F1, 96.4% rejection, and 9.8 MAD on WeRefMat referring.
Picked Oct 26
- RepoCuboiLeo/SuperMimic-Reproduced→
Unlike most control repos, it fully reproduces SuperMimic on Unitree G1, matching 0.82 LAFAN1 in 1.66 hours on one RTX 5090.
Datasets
Picked Oct 29
- Datasetcrtproduction/name-that-llm-open-pool→
Unlike most attribution datasets, it provides 254,561 de-duplicated open passages labeled across 18 LLMs with provenance and leakage flags.
Picked Oct 28
- Datasetbaidu/MarathonDocBench→
Makes hundred-page parsing measurable with cross-page structure and 8 document-level metrics, exposing failures of page-wise and general VLM parsers.
Picked Oct 27
- DatasetQwen/BabelArena→
Unlike translation-only multilingual benchmarks, it fully localizes tasks, tools, and environments across 23 languages.
Picked Oct 26
- DatasetSeerRay-Lab/LensBench→
Matched masks across explicit/implicit wording and target counts isolate instruction-grounding failures across four domains.