Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 45 · 2026
Nov 02 – Nov 08
20 papers, repos and datasets selected this week.
Papers
Picked Nov 6
- PaperToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents→
Raises Qwen3.6-35B-A3B from 23.0 to 41.0 avg (43.2 with RL); frontier models drop up to 38.2 pts on Pro benchmarks.
- PaperRoboColosseum: Evaluating Physical Intelligence at Scale In-The-Wild→
Built from 5,699 comparisons at 124 sites; active matchmaking hits ~90% ranking agreement in 400-500 trials vs ~1,000 uniform.
Picked Nov 5
- PaperRoute Anything: A Study of Model and Policy Routing for General-Purpose AI Systems→
Oracle routing beats the best fixed candidate by 5.4 points; learned routers gain 0.3 overall, 2.1 on video, 1.1 on robots with 24% less policy time, and 8.82x efficiency.
Picked Nov 4
- PaperCUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements→
CUA-Master-27B scores 80.2 OSWorld Verified, 36.9 OSWorld 2.0, 43.9 WeaveBench, and 25.0 on Agents' Last Exam.
- PaperLibrary of Layers (LoL): Adaptively Ordering Transformer Layers at Runtime→
Competitive with dense, looped, and MoE baselines; safety probes on router outputs reach 0.95 AUROC, and forced routing controls sudoku/maze strategy with ~100% accuracy.
- PaperWave Forcing: Towards Speed-of-Light Streaming Video Generation→
It pipelines denoising with cache refresh to stream one video chunk per forward pass.
Picked Nov 3
- PaperLatent Reasoning in Continuous Space for Unified Multimodal Models→
Sets SOTA on GenEval at 0.88, +7%p over BAGEL; also improves DPG-Bench to 90.04 and OneIG-Bench to 0.866.
- PaperOn the SPOT: Enhancing Speculative Decoding for Diffusion Language Models via Commit-Order Prediction→
On SDAR-8B, SPOT cuts target passes by up to 6.8x and reaches up to 3.4x throughput, matching or beating baselines across tasks at T=0 and T=1.
Picked Nov 2
- PaperProving and Improving the Platonic Representation Hypothesis with JEPAs→
It proves cross-modal JEPA encoders align up to rotation and beats 72B LMs with 42x fewer samples.
- PaperYoutu-Parsing-Omni: One Encoder, One Schema, All Modalities→
SOTA 96.96 on OmniDocBench and best open-weight 75.08 on OmniParsingBench; 54.2 [email protected] on chemistry and 22.73 CER on sheet music.
- PaperVibeVideoBench: Evaluating Multimodal Agents on Code-Driven Video Generation→
It benchmarks autonomous code-written MP4 generation in 1-hour sandboxes with evidence-based grading across 239 tasks.
Repos
Picked Nov 6
- RepoGenLang-V/GenDB-V→
Combines formal end-to-end correctness with large reported gains: up to 61.3× vs DuckDB and 28.3× vs Umbra.
Picked Nov 5
- Repowangzan101/Mika→
Runs a cross-app Android agent entirely on-device, including background-screen execution and user handoff.
- RepoinclusionAI/IDM→
Uses internal EMA feature-moment matching to reach 94.25% LongText-CN in two steps.
Picked Nov 3
- Repohyz317/HairDeck→
Unlike most single-image 3D tools, it outputs hair cards, spring-bone rigs, and GLB/VRM exports together.
Datasets
Picked Nov 6
- DatasetAxion1-0/axion-training-data→
Unlike mostly evaluation benchmarks, it releases 18,363 full agent-training trajectories plus 484 RL tasks with verified tool-use.
Picked Nov 5
- Datasetdeliverbench/DeliverBench→
Makes continuous project delivery measurable beyond single tasks by scoring checkpoint outputs and final packages separately in one workspace.
Picked Nov 4
- Datasettimmzak/ReText-Captions→
Adds text annotations to standard ReID benchmarks, making caption-conditioned and text-augmented person matching measurable on existing protocols.
Picked Nov 3
- Datasetgoogle/physical-engineerings-last-exam→
Unlike recent agent benchmarks, it evaluates long-horizon engineering deliverables in realistic multi-tool CAD, BIM, and simulation workspaces.
Picked Nov 2
- DatasetInternRobotics/RoboAtom→
Aggregates 12 robot datasets into 100K-1M time-aligned atomic action annotations with a 500-trajectory held-out segmentation-and-description benchmark.