AI

  • arXiv cs.AI — SqueezeBench — A saturation-resistant math benchmark where each accepted Lean-verified bound becomes the new record to beat: frontier-model agents given one hour per record improved all 20 initial records, but for 8 of 10 problems the best agent record still trails the published literature asymptotically.
  • arXiv cs.AI — SWE-Critix — Post-training 30B verifiers to judge coding-agent trajectories lifts held-out classification accuracy from 57% (base) to 66% (SFT) to 78% (RL); models, data, and code released.
  • arXiv cs.AI — LongHarness — A stateful evidence-management harness lifts GLM-5.3-Flash’s exact-match on OfficeQA Pro from 15.8% to 60.2% (V1) and 4.4% to 61.1% (V2) — author-reported as beating Claude Code and Codex on the same benchmark.
  • arXiv cs.AI — ProgramGym — 1,648 program-reproduction training environments across C/C++/Go/Rust; SFT on their trajectories raises a 27B model from 34.87% to 42.84% on ProgramBench, with an agent workflow reaching 55.71–56.25%.
  • arXiv cs.AI — EmbodiedCLBench — First embodied-simulation benchmark for continual learning in self-evolving agents finds experience from basic tasks transfers to unseen advanced ones, but gains don’t scale with more learning samples.
  • Hugging Face Papers — RSI-Anything — World-model-guided recursive self-improvement raises code-repair success from 26.3% to 31.2% at comparable test usage, and cuts GUI-task actions by 39.4%.

General

  • Sakana AI — MASS has one shared model act as solver, workflow optimizer, and evaluator for a team of its own subagents, then fine-tunes on the best team’s executions, reporting higher score-per-output-token than the base model on several research benchmarks.