20-Step RL on LFM3-1.7B Lifts opencode Pass@1 by 9.4 Points
yb2698 · x · 2026-11-06
After a 20-step smoke RL run on LFM3-1.7B, evaluated on 150 held-out tasks never seen in training (paired with base, 95% CI): opencode +9.4 ± 6.3 points pass@1 (69%→79%), hermes +6.7 ± 7.1, pi flat (−1.3 ± 6.7).
The author notes results are early and noisy (one sample per task), but the opencode gain clears the noise band. Full runs start on a livestream.
More from coding & agent
- Zep Agent Skills turns verified agent traces into versioned, governed procedural memory — TheTuringPost · 2026-11-06
- Sentry's next MCP evolution: one CLI login powers both MCP and full CLI access — zeeg · 2026-11-06
- Box launches Box Content Retrieval with agentic platform and MCP support at BoxWorks — shashib · 2026-11-06
- Give your agent an interpreter, not tools: 19 min to 2 min on the same model — Itz_hany · 2026-11-06
- Auditing the top 99 MCP servers: no tool poisoning, the real bugs are in HTTP mode — Trout_dev · 2026-11-06
- Notion says third-party agent MCP traffic is up 60x this year, warns MCP endpoints lack security policy — dsp_ · 2026-11-06