repositories

Infrastructure for continually self-improving agents. Connects live interactions, feedback, and continual learning to improve model weights and agent harnesses.

Official code for SPADE, a self-play framework that jointly learns to generate executable environments and solve them (arXiv, 2026).

Robust, lightweight infrastructure for multi-agent autonomous self-evolution, built for autoresearch (COLM 2026). Works with Claude Code, Codex, Cursor, and more.

Learning to synergize memory and reasoning for efficient long-horizon agents (ICLR 2026).

A framework for batch parallel speculative decoding (ICML 2026).

Optimizes the total throughput of batch speculative decoding by strategically selecting draft tokens for verification (ACL 2025).

Official code for "DETAIL — Task DEmonsTration Attribution for Interpretable In-context Learning" (NeurIPS 2024).

Code for the AAAI 2023 oral paper "Probably Approximate Shapley Fairness with Applications in Machine Learning".