Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 33 · 2026
Aug 10 – Aug 16
20 papers, repos and datasets selected this week.
Papers
Picked Aug 14
- PaperMechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence→
Targets autonomous theory-level mechanistic discovery, grounding experiments in 13k interpretability papers and 32 tools.
Picked Aug 13
- PaperBeyond Pixels: From Video Priors to 4D Worlds→
Beats matched Wan+4RC cascades in DINO-F1 by 2.88-3.45 on Text4D-200 and 5.81 on I4D-200, with better human ratings.
- PaperRynnValue: Scaling Robotic Value Foundation Models with Temporal Distance→
Gets 0.675 Kendall's tau_a on RBM-EVAL-OOD vs 0.655 preference-SOTA and 0.292 progress-only; boosts real-world success 52.5%→72.5% online, 63.8%→82.5% offline.
- PaperSWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring→
170 curated tasks average 11.4 files and 261.6 LOC; the best model resolves only 41.2%, showing a hard, unsaturated benchmark.
Picked Aug 12
- PaperStealing Reasoning Traces from Proprietary LLM APIs→
Exploits weaker sibling models as cross-provider decryption oracles, recovering 182 credentials from 315,320 reasoning blobs.
- PaperOn-Policy Self-Distillation without Any Supervision→
On 5 math benchmarks, it beats base Qwen3 by 8.5%/10.7% in non-thinking 4B/8B and tops OPSD by 3.2%/2.3% on average.
- PaperScaling Inherently Interpretable Language Models→
Steerling-8B stays within ~10% of open peers' average benchmark performance despite 2-16x less compute; interpretability metrics improve with scale.
Picked Aug 11
- PaperStreamArena: Toward Continuous, Interactive, and Long-Horizon Agentic Streaming Video Understanding→
StreamMind leads streaming systems on all 4 capabilities, improving strongest baselines by 53.7%-228.1% and cutting query-to-answer latency by 66.2%.
- PaperSFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs→
Multi-stage SFT averages 23.1% below base, while multi-stage RL gains 24.9%; RL task-update similarity is ~1e-5 vs SFT ~1e-1.
Picked Aug 10
- PaperToward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning→
Skill-Entropy RL raises Skill2-Bench from 34.4% to 68.4% on Qwen3-4B and 14.6% to 40.1% on Qwen3-1.7B.
- PaperRecursive Synthesis for Long-Horizon Terminal Tasks→
Recursively grows sandbox-verified terminal tasks from seeds, yielding 37,484 tasks at $0.05 each.
Repos
Picked Aug 14
- Repoled829155-droid/RecSys-MLops→
Covers the full real-time path and automated rollout flow, including 10%→25%→50% sticky A/B promotion with automatic rollback gates.
Picked Aug 13
- Repoyanghaha0908/GROW→
Applies on-policy group-relative reward reweighting directly to flow-matching TTS, improving WER and speaker similarity while training 2.9× faster than DiTAR-GRPO.
Picked Aug 12
- Repozju3dv/INTACT-JEPA→
Unlike search-based world-model controllers, it learns direct goal-conditioned actions, hitting 95.33% success at 2.9-5.5 ms latency.
Picked Aug 11
- Repoxmarre/ComfyUI-Spectrum-MiniMax-H3→
Cuts H3 transformer steps by 30-35% in 20-step runs; one reported RES workflow reduced time 212.73s to 160.97s (24.33%).
Picked Aug 10
- Repojd-opensource/JoyAI-Video-Edit→
Unlike offline video editors, it performs causal instruction-guided editing on open-ended streams at 30.19 FPS.
Datasets
Picked Aug 14
- Datasetmuhammadaqeelkabir/steam-games-dataset-v2→
Combining SteamSpy rankings with official store metadata makes richer EDA, description NLP, and score or engagement prediction feasible in one table.
- Datasetstevenleehans/rsna-knee-llm-report-labels→
Provides probabilistic LLM report labels for all 4,407 RSNA knee MRIs, beating regex targets 0.8927 versus 0.8136 macro AUC.
Picked Aug 11
- Datasetpymaster/CrawlSinger-OS→
Provides 2,317 hours of singing with unified lyric-note-score alignments, far beyond typical sub-100-hour SVS datasets.
Picked Aug 10
- Datasetairoa-org/airoa-moma-5k→
Provides 1.18M real-robot mobile-manipulation primitive episodes, each independently trainable yet task-linked.