A-EVO LAB

Focus on Recursive Self-Improvement (RSI) · Amazon

We train AI researchers.

Research skill can be learned. We train models to propose, predict and run the experiments that make AI better, and we study agents that evolve their own harness from experience.

See the research Contribute a task to OpenRSI ↗
Policy · LM2LM A trained 27B proposer beats GPT-6-sol Avg@4 gain on the held-out Qwen3 family · 10⁻³ BPB · real execution · higher is better
  1. LM2LM Qwen3.8-27B + RL53.72
  2. GPT-6-sol + 300 records21.63
  3. GPT-6-sol20.19
  4. Qwen3.8-27B base3.53
  5. Qwen3.8-27B + records2.97
vs GPT-6-sol
2.5× avg gain
Win rate vs base
87.5%
Data cost
−33%

Qwen3 family held out from proposer training and RWM groundingLM2LM ↗

Value · Research World Models Experience beats scale: 13 of 13 models improve Spearman ρ predicting intervention gains · OLMo3-100Mno records+ 320 experimental records
  1. Claude Opus 50.90
  2. Grok 4.60.89
  3. GPT-6 Astra0.88
  4. GPT-5.6 Sol0.88
  5. GPT-5.6 Luna0.87
  6. GPT-5.6 Terra0.86
  7. DeepSeek V3.20.86
  8. Kimi K2.50.86
  9. GLM-50.80
  10. Nemotron 3 Super0.80
  11. MiniMax-M2.50.78
  12. Qwen3 235B0.77
  13. Mistral Large 30.64

Best model without records: 0.62 · −78% selection regret on Qwen3 from other environmentsRWM ↗

Harness · A-Evolve-Training Autonomous 30B post-training, one point from the top human team Nemotron Reasoning Challenge · held-out score · snapshot 1 Jun 2026
  1. Top human team #10.87
  2. A-Evolve-Training #8 · autonomous0.86
Rank
8 / ~4,000
Base
30B Nemotron
Rounds
4 × 8 agents

Data mix, schedule and ablations chosen by the systemReport ↗

01 / AI Researchers Primary program

Train the researcher, not just the model it improves.

Agents can propose experiments far faster than anyone can afford to run them. So we treat research as search under a fixed compute budget, and learn each piece of it: the game, a value estimate, a policy, and the harness that runs the campaign.

Game

OpenRSI Index

Real model-development tasks with fixed verifiers, fixed GPU budgets and a human baseline. The agent can change the method, never the judge.

Up to 20,864 H100-h per run

Index ↗Tasks ↗

Policy

LM2LM

Train the proposer with RL against the world model, and keep that reward honest with periodic real runs.

33% lower data-generation cost than 300-record ICL

Project ↗arXiv:2610.23344 ↗

Harness

A-Evolve-Training

The system around the researcher for multi-week campaigns: an immutable substrate, memory-free workers, and a meta agent that rewrites only the search policy.

0.86 vs 0.87 human #1 at 30B

Report ↗arXiv:2606.20657 ↗

  1. Real runs
  2. Experimental records
  3. Sharper world model
  4. Better proposer
  5. Faster idea search

Failed experiments are not wasted budget. They are world-model training data.

Environments

Pre-training, post-training and generation are not separate directions. They are places a researcher works, with a budget attached.

StageTaskWhat the researcher changesBudget per run
Pre-trainingMarin-Scaling-LadderOptimizer, validated from 550M to 2.5B18,432 H100-h
Post-trainingQwen-122B-RL-MergeSynthesized RL tasks≈20,864 H100-h
GenerationGPIC LeaderboardRecipe for one more training pass5,815 H100-h
LM developmentLM2LM suite8 autoregressive and diffusion LM families≈10 GPU-h

Turn your own research into an environment in about an hour with the RSI-Anything agent. Contributors receive authorship. Contribute ↗

02 / Self-Evolving Agents

Agents that evolve their own harness.

Keep the model frozen and let the agent improve its prompts, skills, memory and workflow from experience. We ask when that helps, when it hurts, and how to keep it working on open-ended task streams.

A-EVO-Lab / a-evolve Open-source framework for evolving any agent, in any domain, with any evolution algorithm. Adopted or integrated by OpenRLHF, DeepSpeed, SGLang and GEPA. 844 stars97 forksGitHub ↗
A-Evolve · zero manual harness engineering Evolved harnesses on public agent benchmarks Claude Opus 4.6 solver · leaderboard rank at time of reportingbaseline harnessevolved harness
  1. MCP-Atlas #179.4+3.4
  2. SWE-bench Verified #576.8+2.6
  3. Terminal-Bench 2.0 #776.5+13.0
  4. OSWorld69.6+3.9
  5. SkillsBench #234.9+15.2

Accuracy / pass rate, %. Baseline is the same model without the evolved harness.Code ↗

  • EMNLP 2026Harness Updating Is Not Harness Benefit

    Writing harness updates and benefiting from them are separate skills. Across seven evolvers, best and worst differ by at most 3.1pp; mid-tier agents gain most.

  • EMNLP 2026Evo-Harness

    Compile one-shot execution context into reusable skills. Wins every baseline on CL-Bench, τ-bench and WebArena-Infinity.

  • 2026Adaptive Auto-Harness

    A harness tree with solve-time routing for long-running streams. 80.9% vs 50.8% next best on PolyBench prediction markets. Code ↗

  • EMNLP 2026MemMA

    Coordinate memory construction and retrieval, with verification and repair before memories are kept.

  • PositionAgentic Evolution Is the Path to Evolving LLMs

    The case for improving the system around the model as its own axis of progress.

03 / News

Recent.