RLVR-trained small model fixes 83.7% of LaTeX errors, 6x faster at 1/40 the cost
simonguozirui · x · 2026-09-17
Tinker (Sundial) fine-tuned thinkymachines' Inkling-Small with RLVR to fix LaTeX compile errors directly inside the editor: 83.7% fix rate on real errors, matching Claude Fable 5.1 while being 6x faster and 1/40 the cost ($0.0013 per fix), responding in under a second.
Their training emphasized getting RLVR right:
- Training on verified real fixes rather than artificially injected errors
- Reward design that explicitly disincentivizes reward hacking, with deterministic scoring
The motivation came from interviews with mathematicians, whose top daily annoyance was battling LaTeX compilation errors. Asking a chatbot works decently, but switching windows breaks writing flow, and copy-pasting snippets loses context—especially across multiple files. They scraped 272,000 questions from TeX.StackExchange and kept 3,978 with usable snippets as training data.
More from coding & agent
- Bash is no longer all you need for reliable agent tool calling — yenkel · 2026-09-17
- Dev lead running 50 deploys a day shares how to write and run tests in the vibe coding era — dotey · 2026-09-17
- YC built AI versions of its partners on GLM-5.2, cutting latency 31% vs OpenAI — ycombinator · 2026-09-17
- Dev ports bash-tool project to Claude Managed Agents, sandbox spins up independently — trq212 · 2026-09-17
- Developer Reverses Course: MCP Is Now Better Than CLIs for Most Integrations — trq212 · 2026-09-17
- Shape Claude's Tools the Way You Want Instead of Hiding Behind Indirection — trq212 · 2026-09-17