Qwen3-27B jumps from 39% to 65% on SWE-bench Pro with RL and one small LoRA
TheTuringPost · x · 2026-11-21
TheTuringPost recaps a Hugging Face recipe that flips the usual approach: keep the harness, adapt the model. With RL and a small LoRA, Qwen3.8-27B went from 39% to 65% on SWE-bench Pro. In a reply thread, rasbt notes LLMs are specifically fine-tuned for particular harnesses — the harness is part of the training setup, so the model can't absorb it, like a CPU/GPU absorbing the OS. The open question: which side is cheaper to change for your setup?
Related event: Debate: Will LLMs Absorb the Agent Harness, or Should Models Be Fine-Tuned(3 posts)→
More from coding & agent
- A project manager's working review: Pebble Index 01 vs Sandbar Stream AI ring — Scobleizer · 2026-11-21
- Founder says Squad agents have fully taken over his Gmail inbox with follow-ups and pings — tibo_maker · 2026-11-21
- Garry Tan: vibe-coded hobbyist GTA VI PC port fixing hated flaws could take 'a month or two' — garrytan · 2026-11-21
- shadcn explains uisdk: import primitives, not wrappers, one codebase across 5 frameworks — shadcn · 2026-11-21
- Microsoft's Fabric HorizonDB signals the shift to agent-driven applications — DavidLinthicum · 2026-11-21
- Multica Ships Realtime Docs for 100+ Humans and Agents, Replacing Feishu Internally — jiayuan_jy · 2026-11-21