Tsinghua's Frontis-MA1 pushes 35B model to 71.21% on MLE-Bench Lite for recursive self-improvement
ceciletamura · x · 2026-09-03
A Tsinghua University team released Frontis-MA1, a paper studying recursive self-improvement (RSI) using machine learning engineering (MLE) as an executable testbed for AI4AI.
- OpenMLE stack: an open full-stack RSI research system spanning verifiable task environments with execution feedback (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo).
- Frontis-MA1 (35B): post-trained as a meta-evolution agent around four atomic program-evolution operators (Draft, Improve, Debug, Crossover), trained via execution-grounded SFT and RL on data deduplicated against all eval benchmarks, then composed into long-horizon search coupling learning and evolution in a single loop.
- Results: on MLE-Bench Lite under a 12-hour per-task budget on a single RTX 4090 capped at 12GB VRAM, it lifts Medal Average from 39.39% to 60.61% with OpenMLE-Evo, and 71.21% with OpenMLE-Evo-Max.
More from coding & agent
- Open-source coding agent Pi hits 100,000 GitHub stars, v2 teased — gklambauer · 2026-09-03
- Codebase-specific agent harnesses may beat general-purpose ones, argues secemp9 — secemp9 · 2026-09-03
- Teknium runs Hermes Agent for 15 hours with recursive subagents, shedding 375,000 lines of code — Teknium · 2026-09-03
- Building a local Gemma planner-executor-critic agent pipeline for web research — HlddenDreck · 2026-09-03
- Dev builds Arcon: a local AI assistant on Qwen3-4B with memory, mood and tools — Feathered-Beast · 2026-09-03
- After DeepSeek Harness: why agent harnesses still lack any scientific evaluation — samrauh · 2026-09-03