OPEN LINK ↗
// 1h agoRESEARCH PAPER
Self-HPO Rewrites Its Coding Harness
Microsoft Research, the University of Virginia, and Carnegie Mellon introduced Self-HPO, a coding-agent framework that analyzes failed runs, rewrites its prompts and runtime harness, then uses reinforcement learning to improve the policy inside it. Starting with Qwen3.5-35B-A3B, it raised average resolve rates from 57.8% to 64.8% across three SWE-bench evaluations. [Research announcement](https://www.cs.virginia.edu/~tqf5qb/blog/self-hpo/)
// ANALYSIS
Self-HPO makes a compelling case that agent scaffolding and model weights should evolve together, though the results remain early rather than proof of open-ended recursive self-improvement.
- –The framework improved SWE-Bench Verified from 61.0% to 69.5%, with gains also transferring to multilingual and Pro benchmarks.
- –Its strongest practical insight is targeting harness changes at the policy’s current failure modes instead of optimizing against a static task set.
- –The approach requires no external teacher model, although using Opus 4.8 as the harness proposer pushed average performance higher.
- –Results come from two optimization rounds, Python-only training, one base model, and author-run evaluations, so broader production impact is still unproven.
- –For agent builders, harness code is becoming a trainable system component—not just fixed prompt glue.
// TAGS
self-hpocoding-agentai-codingagenttrainingevaluationresearch
DISCOVERED
1h ago
2026-10-20
PUBLISHED
1h ago
2026-10-20
RELEVANCE
9/ 10
AUTHOR
mark_k