Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 34 · 2026
Aug 17 – Aug 23
20 papers, repos and datasets selected this week.
Papers
Picked Aug 21
- PaperSPADE: Self-Play in Adaptive Synthetic Executable Environments→
It jointly learns to generate executable Gym environments and solve them, outperforming fixed-environment self-play by +5.3 average across eight held-out benchmarks.
- PaperZetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence→
Reaches 90.8% on LIBERO-Pro and 93.6% on RoboCasa, with 11.1x lower inference latency and 20.6x higher rollout throughput.
Picked Aug 20
- PaperEmbodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation→
SOTA on VLN-CE: 66.2% SR on R2R-CE and 65.7% on RxR-CE; uses only 90k trajectories and runs in 16.58s/task vs 37.47s and 41.46s baselines.
- PaperUI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations→
Sets open-weight SOTA at 77.0% OSWorld-Verified and 66.2% WindowsAgentArena; one demo lifts OSWorker strict success 17.2%→35.4%.
Picked Aug 19
- PaperImproving the matrix multiplication exponent with modern optimization and AlphaEvolve→
Certified a new upper bound of ω < 2.371177, improving the previous best 2.371339.
- PaperLarge Discovery Models: Empirically-grounded Model-Based Open-Ended Search→
Beats LLM-only and BO baselines: 2.4× more BPB reduction, 18.2% lower binding energy, and >60% better molecular hypervolume.
- PaperApodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence→
Beats published AAV capsid-design SOTA by 7%; biomedical environments raise GPT-5.5/5.6-sol scores by 2.5/7.6 points.
Picked Aug 18
- PaperSelf-Supervised Visual On-Policy Distillation→
Unlike teacher- or label-dependent distillation, it uses clean-versus-degraded views to self-distill, lifting Qwen3.5-4B from 70.7% to 77.4% and surpassing Qwen3-VL-235B.
Picked Aug 17
- PaperIntern-S2-Preview: Scientific Agentic Foundation Model→
Biology-Instructions rises 56.92→60.32 with the 4B Memory Decoder on a frozen 397B backbone; the new time-series encoder supports 300k steps with ~5-6× faster inference and ~20% memory.
- PaperAlaya-EVOKE: From Linear-Scaling Supervision to Endless World→
Supports open-ended generation with 2.11s per 1.5s chunk on one H200; as a 3-step world model it reaches SOTA on WBench and stays competitive on VBench-Long and VBench-2.0.
- PaperOpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution→
Targets stateful agents by evolving shared environment state, reaching 85.0% ASR across 75 setups.
Repos
Picked Aug 21
- RepoSigmanticAI/apex-inference-chip→
Measured 0.56 tok/s on FPGA, up from 0.004 tok/s host-driven baseline; every RTL block is bit-exact against the golden model.
Picked Aug 20
- RepoMirroS-Lab/HarnessEval-W→
Unlike scalar-only video benchmarks, it uses model-agnostic skill routing with auditable evidence trees.
Picked Aug 19
- RepoAlayaLab/Evoke→
Unlike recent camera-control video models, it sustains long interactive rollouts with bounded memory and mid-rollout re-prompting in three CFG-free steps.
Picked Aug 18
- RepoMaincodeHQ/mainarch→
Removes ROCm/HIP/HSA at runtime and shows live MI355X inference; beats RCCL on small all-reduce messages with ~20 µs at 2 KiB vs ~26 µs.
- RepoTencentARC/SCoPE→
Adds explicit camera-ray token coordinates to a pretrained video DiT with under 0.1% new parameters.
Datasets
Picked Aug 21
- DatasetOpenMOSS-Team/SWE-bench-Science→
Makes scientific-code agent evaluation reproducible with isolated envs, held-out verifiers, and explicit license-gated task selections.
Picked Aug 20
- DatasetAnthropic/claude-protein-binder-design→
It links 1,440 de novo binders to raw two-vendor kinetics, structures, and stepwise design provenance.
Picked Aug 18
- Datasetr0b0tlab/deepseek-v4-pro-0813-agentic→
Unlike recent agent datasets, each of 21,207 multi-turn traces includes deterministic verifiers and exact reward contracts.
Picked Aug 17
- DatasetTrNi/GenSAGA-3D→
Unlike typical 3D corpora, it pairs 9.3K text-conditioned meshes with four safety tiers and reusable views, SDF, point-cloud, and embedding artifacts.