Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 38 · 2026
Sep 14 – Sep 20
20 papers, repos and datasets selected this week.
Papers
Picked Sep 18
- PaperScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments→
Builds 64 environments, 2,812 tasks, and 1,076 checks; trained models improve held-out scientific-code repair and selected general code, reasoning, and knowledge benchmarks.
- PaperAgora: Git as Shared Memory for Collective AutoResearch→
In a 12-day 13-agent run, it improved a weight-transfer task from 3.39 to 1.899 bpb, closing 62% of the gap to trained GPT-2 124M.
Picked Sep 17
- PaperPhysBrain 1.5: From Vision-Language Models to Physical Foundation Models→
72.5 avg on 28 embodied benchmarks; best open-source overall, first on 14, near GPT-6-Astra 73.3 and Gemini 3.6 Flash 73.0.
- PaperContinual Learning Mechanisms Compose for Long-Horizon Memorization→
Best method lifts average final retention from 1.2% to 34.9% after 100 tasks, a 28x gain, and ranks top 3 on all 3 datasets.
Picked Sep 16
- PaperDream-RSI: Recursive Self-Improvement through Evolving Worlds→
Across 8 tasks, it improved or matched discovery quality with far lower cost: up to 162x fewer agent calls, over 50x budget savings, and up to 2.09x better kernel performance.
- PaperZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search→
Unlike larger post-trained specialists, it trains an open 7B from scratch to 256K context, ranking first across 14 reasoning benchmarks and reaching 4.2x faster 16K pretraining time-to-loss.
Picked Sep 15
- PaperBreaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models→
Improves LIBERO-Plus OOD success by 3.87-10.70 points across 4 models, with 13.30-16.70-point real-world gains under unseen visual shifts.
- PaperSAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking→
Beats SeerAttention-R across tasks and budgets; at 1024 tokens gains are +6.0–7.7 on MATH500 and +10.6–15.5 on GPQA-Diamond.
- PaperSenseNova-U1.5: Towards Native Unified Visual Intelligence→
Authors report broad gains in image fidelity, text rendering, composition, multi-reference editing, and interleaved generation, with native support up to 4K.
Picked Sep 14
- PaperT1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks→
Raises Terminal-Bench 2.1 from 43.8% base to 64.0%; on Long-Horizon Terminal Bench reaches 27.9%, beating GPT-5.4 and GLM-5.1.
- PaperNCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction→
Adds product-quantized next-concept prediction to autoregressive decoding, matching OLMo-3-7B loss with 51.3% tokens.
- PaperAn Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics→
Reached IMO 2026 gold purely with natural-language generate-verify-refine search, without formal tools.
Repos
Picked Sep 18
- Reposuanmiao/fly-self-driving→
Completes 58/60 unseen traffic streets with no car collisions; the same measured graph beats a rewired control 58/60 vs 16/20.
Picked Sep 17
- Repoatria-asi/Atria-Dawn-Preview→
Unlike agent runtimes, it ships a 744B MoE model with leading tool-use benchmarks.
Picked Sep 16
- RepoAetherLabsAI/RSIAgent→
Unlike tool-use agents needing finetuning, it improves OSWorld 2.0 partial credit 71.97→78.98 using frozen-memory exploration across 82 tasks.
Picked Sep 14
- Repokevinzakka/mjbatch→
Unlike GPU-focused simulators, it exposes 4096-way CPU MuJoCo batching with direct NumPy state access.
Datasets
Picked Sep 18
- DatasetHokin/object-permanence→
Makes object-permanence reasoning directly trainable and measurable with controlled task structure plus a shared 300-question exam.
Picked Sep 17
- DatasetCohereLabs/tiny-aya-l2-thinker-multilingual-reasoning→
Provides 250k aligned reasoning traces across 44 non-English languages, unlike mostly English-only reasoning datasets.
Picked Sep 16
- Datasetopenbmb/UltraData-SFT-Agent-2609→
Unlike smaller or single-turn tool-use sets, it releases 484k executable multi-turn trajectories with tool schemas, observations, errors, retries, and final answers.
Picked Sep 15
- DatasetLinzhan/UniML3D→
Unlike text-to-motion datasets, it canonicalizes 7,430 heterogeneous humanoid, animal, and object skeletons with topology and orientation annotations.