0x960
0x960 is an OpenEnv self-improvement environment where an agent improves a Chess960 engine by editing its evaluation code, benchmarking the result, and iterating through self-play.
Core trick
Agent edits a Chess960 engine's evaluation code through five bounded tools, and reward comes from match results against champion builds.
Description
0x960 is an OpenEnv self-improvement environment where an agent improves a Chess960 engine by editing its evaluation code, benchmarking the result, and iterating through self-play. We use a teacher-student setup: a stronger coding agent generates successful edit-test-finish trajectories, a smaller open model is distilled on those traces to learn the workflow, and then the engine is further improved through automated champion-challenger search and benchmarking. The result is a reproducible environment for training and evaluating long-horizon tool-using agents on real engine optimization, with measurable gains in reward, behavior, and engine strength.
README excerpt
--- title: "0x960" emoji: "♟️" colorFrom: indigo colorTo: gray sdk: docker app_port: 8000 pinned: false --- # 0x960: Self-Improving Chess960 Engine with OpenEnv **An AI system that doesn't play chess. It engineers better chess engines.** 0x960 is an OpenEnv self-improvement environment where the policy acts like an engine engineer. Each episode gives the model a constrained workspace and five bounded actions — `read_file`, `write_file`, `run_static_eval`, `run_match`, `finish`. The reward is grounded in downstream match performance on Chess960 positions, not text quality. > **Hackathon Track:** Statement 4 — Self-Improvement > **Stack:** OpenEnv 0.2.1 · HF TRL GRPO · Qwen 3.5 · GPT-5.4 Teacher · ACP Runtime · Codex Swarm ## Why this setup matters Most RL-for-LLMs setups optimize text completions against proxy signals. 0x960 optimizes multi-step tool use with verifiable outcomes: - Code edits must be valid (invalid writes are rolled back instantly) - Changes are tested in real matches against held-out Chess960 positions - Performance is measured against Stockfish UCI anchors - Chess960 prevents reward hacking through opening memorization — the engine must generalize across 960 starting positions --- ## Results ### Training Reward Improvement The core training signal — reward from the bounded OpenEnv environment — improved dramatically across the pipeline: | Stage | Model | Reward | Behavior | |-------|-------|--------|----------| | Base model (no training) | Qwen 3