Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 37 · 2026
Sep 07 – Sep 13
20 papers, repos and datasets selected this week.
Papers
Picked Sep 11
- PaperShow-Harness: Just a VLM Agent Can Play Robots→
Authors report robust cross-task, cross-embodiment gains over representative agentic and VLA baselines, plus strong sim-to-real transfer; no exact numbers shown here.
- PaperScaling Automatic Research Agents via World Models→
Uses learned world-model rewards with anchor-based bias correction, cutting RL training 3-4×.
Picked Sep 10
- PaperAuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing→
Leading on zero-shot/instruction speech generation and general editing; AuK-Flash runs 4-step CFG-free inference with 4.5× speedup.
- PaperDriveZero: End-to-End Driving Beyond Human Demonstrations→
Unlike imitation-heavy driving papers, it reaches 95.3 PDMS camera-only without human trajectory supervision.
Picked Sep 9
- PaperRandom Attention: Rethinking KV Cache Eviction for Efficient Reasoning→
Across 4 models and 6 reasoning tasks, it matches the strongest prior evictor and serves 32-43% higher throughput than TriAttention in vLLM.
- PaperBilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems→
On 500 SWE-bench instances, the full Kimi-based system resolves 72.2% vs 70.8% for public mini-SWE-agent v2.
Picked Sep 8
- PaperWorldSculpt: Generating Compositional Worlds from Grounded Videos→
Consistently outperforms prior methods on single-object, multi-object, and UE-MeshyScene tests, with larger gains as occlusion and scene complexity increase.
- PaperIris: Climbing to the Search Frontier→
Iris-pro gets 88.6 BrowseComp, 92.9 DeepSearchQA, and 56.4 HLE, leading or tying open-source agents in its size range.
- PaperPost-Training Language Models for Gold-Medal Performance in Coding Competitions→
On IOI 2026, the competition system scored 535.4/600, beating the top human's 498.27; on IOI 2025, Nano-CC rose from 130 to 468 with GenCorrect.
Picked Sep 7
- PaperTerminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments→
Builds 37.3k task-sufficient environments; SFT on this data improves Qwen3.5-27B by +11.9 on Terminal-Bench 2.1 and +13.8 on EvoCode-Bench v2 MT@4.
- PaperLLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes→
Uses mostly real image-only pretraining, then frozen dLLM conditioning and 2-4-step distillation.
- PaperLanguage Models Can Control Their Own Attention→
Across 15 zero-shot long-context tasks, DA cuts attended decode tokens by 52.0% on Gemma-4-31B and 31.1% on Qwen-3.6-27B, with only 1.27pp and 2.75pp accuracy drops.
Repos
Picked Sep 11
- RepoUMass-Embodied-AGI/SyncWorld→
Avoids per-deployment fine-tuning by inferring the action-to-visual mapping from an in-context calibration episode.
Picked Sep 10
- RepoInstituto-IA-LNCC/manaca-1b-base→
Unlike most LLM releases, it is fully reproducible end-to-end on 2×24GB GPUs.
Picked Sep 9
- RepoOpenWAM-Official/OpenWAM→
Unlike single-purpose world-model repos, it releases an open modular WAM stack and pretrained checkpoints trained on 518.5M frames.
Picked Sep 7
- Repoifm-ai/uno→
Unlike speculative decoding repos, it parallelizes exact multi-token sampling without a draft model.
Datasets
Picked Sep 11
- Datasetfacebook/WearableQA→
Makes health reasoning measurable on noisy real wearable histories, including missing days and user variability, not synthetic signals.
Picked Sep 10
- Datasettencent/OSWorkerBench→
Unlike prior GUI datasets, it pairs 100 executable office workflows with step-level demonstrations.
Picked Sep 9
- Datasetspatialverse/UniPhys-Bench→
Unlike typical 3D asset sets, it unifies human-verified kinematics, part physics, and object mass/scale on 1,927 simulation-ready articulated objects.
Picked Sep 8
- Datasetopenbmb/UltraData-Code→
Unlike typical code corpora, it pairs 400B scored algorithmic code tokens with 150B implementation-grounded exercise records across 11 languages.