Recursive self-improvement · Autonomous research
Founder & Research Lead, A-EVO Lab at Amazon
I study how AI can improve AI—from evolving agent harnesses to building autonomous research systems, evaluating their discoveries, and training models for research.
I lead A-EVO Lab at Amazon and co-lead OpenRSI Index. Previously, I worked at Microsoft Research and studied machine learning at Carnegie Mellon University (CMU).
A research lab at Amazon focused on recursive self-improvement.
Explore the lab ↗Training models for autonomous research.
arXiv · OCTOBER 2026
Language Models Learn to Improve Language Models · LM2LM
A 27B research model trained with RL achieves higher average experimental gains on a held-out Qwen3 family than GPT-6-sol—even with 300 experimental records in context. Research World Model rewards make training scalable, while real experiments keep the feedback grounded.
Held-out Qwen3 · Avg@4 gain: 53.72 vs. 21.63 for GPT-6-sol + ICL (10⁻³ BPB)
arXiv · OCTOBER 2026
AI agents can propose experiments faster than we can afford to run them. Research World Models use real experimental experience to predict an intervention’s gain before execution, directing limited compute toward the most promising experiments.
Beyond experiment selection, this predictive capability supplies rewards for Autoresearch model training in LM2LM, providing outcome-based feedback without executing every proposed experiment.
Real experimentation: 170,000+ H100 GPU-hours · 2,653 records · 9 research environments
Learning from experience through persistent improvements to skills, memory, prompts, and workflows.
EMNLP 2026
Separating the ability to produce harness updates from the ability to benefit from them.
OPEN-SOURCE FRAMEWORK
A shared framework for building evolvers that turn agent experience into durable improvements. The foundation for our work on online and adaptive harness learning.
EMNLP 2026 · MAIN
Compiling noisy, one-shot execution experience into reusable skills for online harness learning.
EMNLP 2026 · MAIN
Coordinating memory construction and retrieval, with verification and repair before memory is finalized.
arXiv · 2026
Sustaining agent performance on changing task streams through a stateful evolver, a harness tree, and task-wise routing.
Systems that organize and execute the research process: hypotheses, experiments, evaluation, and revision.
arXiv · 2026
An autonomous system conducts a multi-week post-training campaign and revises its search policy when its internal proxy becomes misleading.
Reported score: 0.86 · top human submission: 0.87
Evaluating autonomous research against real model-development problems and human baselines.
BENCHMARK · PROJECT CO-LEAD
An open benchmark for AI research agents, built around auditable tasks and fixed verifiers.
My contribution: co-led and advise the project.
Marin-Scaling-Ladder · Qwen-122B-RL-Merge · GPIC
Selected work on agent training, tool-use evaluation, and value-guided decoding.
NeurIPS 2026
ACL 2026
ICLR 2026
NeurIPS 2017