Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 27 · 2026
Jun 29 – Jul 05
20 papers, repos and datasets selected this week.
Papers
Picked Jul 3
- PaperReinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs→
RLMF beats standard RL by up to 63% and reaches 0.84 cMFG on Llama3.1-8B and 0.83 on Qwen3-8B across 10 tasks.
- PaperOSWorld2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks→
Best agent completes only 20.6% of tasks at 500 steps; GPT-5.5 plateaus near 13%, showing current agents are far from professional-level use.
Picked Jul 2
- PaperBeyond IID: How General Are Tabular Foundation Models, Really?→
Across 11 models and 142 datasets, TFMs win on tiny-to-medium IID data, but tree/deep baselines lead on non-IID, large, and high-dim tasks.
- PaperOrca: The World is in Your Mind→
It learns one next-state latent from video and language, then probes frozen backbones for text, images, and robot actions.
Picked Jul 1
- PaperScaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent→
It matches 1T-class agent results with 35B parameters by scaling 45K-token horizons and merging six specialists via domain-routed on-policy distillation.
- PaperAgentic Abstention: Do Agents Know When to Stop Instead of Act?→
On WebShop, CONVOLVE lifts Llama-3.3-70B timely abstention recall from 26.7 to 57.4 and overall recall from 83.2 to 100.0.
Picked Jun 30
- PaperOPID: On-Policy Skill Distillation for Agentic Reinforcement Learning→
Beats outcome-only RL and prior skill-distillation baselines: e.g., 74.2 WebShop success vs 68.0 SDAR and 63.3 GRPO on Qwen2.5-3B.
- PaperCausal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models→
SOTA streaming video generation; a 2-step causal Wan2.1-1.3B reaches 84.63 VBench-T2V, and TF continuous-time CMs converge 10× faster than dCMs.
Picked Jun 29
- PaperImproved Large Language Diffusion Models→
Keeps masked diffusion end-to-end at 8B scale, matching Qwen2.5-7B and surpassing LLaDA by 14.5-21.6 points on reasoning benchmarks.
- PaperWan-Streamer v0.1: End-to-end Real-time Interactive Foundation Models→
Gets ~200 ms model-side latency and ~550 ms total latency with 350 ms network, enabling sub-second duplex audio-visual interaction at 25 FPS.
Repos
Picked Jul 3
- Repoxiaobiaodu/Flux-GS→
Unlike most 3DGS repos, it exports compressed scenes as JSON for mobile/WebGL decoding.
Picked Jul 2
- RepoYoonkyo/mu0→
Unlike action-token VLAs, it predicts embodiment-agnostic future 3D interaction traces pretrainable from video alone.
Picked Jul 1
- Repowanshuiyin/Anti-Autoresearch→
Uses 46 verdict-bearing integrity patterns with deterministic adjudication; eval reports 100% recall on 7 injected deterministic defects with no clean false positives.
Picked Jun 30
- Repodavnords/HardMatch→
Provides 1,000 hand-annotated hard pairs, including 900 test and 100 validation pairs, for stress-testing image matching methods.
Picked Jun 29
- Repodeepseek-ai/DeepSpec→
Unlike single-method decoding repos, it unifies data prep, multi-GPU training, and evaluation across DSpark, DFlash, Eagle3, and four target-model setups.
Datasets
Picked Jul 3
- DatasetMAIR-Lab-HUST/SciIR-82k→
It pairs 82k scientific figures with three-part reasoning annotations and full CC-BY provenance.
Picked Jul 2
- Datasetsunil123kumar/ai-agent-failure-benchmark-dataset→
Makes root-cause and severity evaluation of agent failures measurable instead of collapsing all wrong outputs into one error bucket.
Picked Jul 1
- DatasetVoxel51/SceneFun3D→
Combines 710 Faro-scanned indoor scenes with part-level affordances, motion, language, and 2D/3D correspondences.
Picked Jun 30
- Datasetalitaqishah/global-urban-heat-intensity-dataset-20152025→
Puts urban heat exposure, greenness, built cover, density, and mortality in one aligned panel for cross-city modeling without raster/API wrangling.
Picked Jun 29
- Datasetnvidia/Cosmos3-DROID→
Unlike single-lab teleoperation sets, it spans 71.9K episodes across 564 scenes and 86 tasks.