Two self-improving agent papers hit EMNLP: a 9B model's harness updates match Opus
yuyinzhou_cs · x · 2026-08-22
HenryLAI's team got two papers on self-improving agents accepted to EMNLP 2026, with code released in A-Evolve.
- "Harness Updating Is Not Harness Benefit": counter-intuitive findings — the updater's base capability barely matters; a Qwen3.5 9B model produces harness updates matching Claude Opus 4.6 (best-vs-worst evolver gap ≤3.1pp). Benefit is non-monotonic: mid-tier agents gain the most, strong models hit a ceiling, and weak models gain least (skill activation 25% vs 96% for strong) despite having the most headroom.
- "Evo-Harness: Context-to-Harness Skill Compilation": how to turn single noisy executions into reusable skill harnesses on the fly.
Both works tackle one core question: what actually makes an agent improve itself through its harness.
More from coding & agent
- Engineer: my job went from silently thinking at a computer to constantly talking to it — KevinNaughtonJr · 2026-10-16
- Vercel Goes Full Backend: Python GA, Containers, and Daemons — cramforce · 2026-10-16
- WorldMemArena Accepted at NeurIPS: Memory Storage ≠ Memory Use — xwang_lk · 2026-10-16
- Vercel Ship Demos Container Deploys for Go, Ruby, Java and More — edwin · 2026-10-16
- TRL adds day-0 support for fine-tuning Muse Spark Open's 816B MoE with LoRA on 64 H100s — QGallouedec · 2026-10-16
- Databricks engineers land impactful PRs on day 2 with AI dev tools, vs week 3 before — Yuchenj_UW · 2026-10-16