Bespoke Labs repo-specialized post-training lifts Inkling coding by 57pp without losing generality
simonguozirui · x · 2026-09-04
Bespoke Labs (Alex Dimakis et al.) post-trained Inkling on a single GitHub repo using SFT on teacher-model trajectories plus GRPO reinforcement learning in curated repo-specialized environments. SFT alone delivered a 52pp improvement on a held-out fontTools eval; adding RL brought total gains to 57pp over base Inkling. The specialized model also improved on out-of-distribution benchmarks (Terminal-Bench 2.1, SWE-Bench), showing specialization doesn't cost general ability while using fewer tokens per correct answer.
More from coding & agent
- OpenAI Live API tip: control voice sessions from your backend via sideband WebSocket — pbbakkum · 2026-10-21
- OpenAI's MultiLangBench: PHP scores highest (91.3%) at building a database engine from scratch — ricklamers · 2026-10-21
- AI auto-optimization speeds genomics code 22x: extract_loci from 7.8s to 0.35s — jmschreiber91 · 2026-10-21
- UserProof lets Claude Code watch your user session recordings and boost conversion for $29/mo — doooyle · 2026-10-21
- dphnAI's distributed inference network served 32B DeepSeek tokens in 24h, ~6% of Venice's daily volume — Ronangmi · 2026-10-21
- Superconductor revamps agent picker spanning 8 coding agents with search and favorites — sergeykarayev · 2026-10-21