Teaching Tiny Models to Prove Hard Theorems | Lewis Tunstall | HF ML Club India EP1
- Source
- youtube.com
- Author
- Hugging Face
- Date
Context
A proof cannot be graded against an answer key the way a numeric answer can, which is the obstacle in a talk by Hugging Face researcher Lewis Tunstall on QED-Nano, a 4-billion-parameter model post-trained to prove olympiad-level theorems. For reinforcement learning, the team had Google's Gemini 3 write a rubric for each problem, a checklist that awards partial credit for steps such as setting up a particular expansion, and had a grade each attempted proof against it. That score served as the reward.
Because proofs run very long, the team added a "reasoning cache" developed by CMU collaborators: the model writes an attempt, summarizes what worked and what did not, and retries with that summary as a starting note. The paper's abstract says QED-Nano surpasses much larger open models such as GPT-OSS-120B and approaches the proprietary Gemini 3 Pro at a fraction of the cost. In the talk, Tunstall puts the single-attempt IMO-ProofBench score near 40 percent, and says harnesses spending over a million per problem get quite close to Gemini 3 Pro. Tunstall also notes that a language-model grader only works if it is roughly as capable as the model being graded. All results are the authors' own.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
- inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
video⚡ Open Model Pretraining Masterclass — Elie Bakouch, HuggingFace SmolLM 3, FineWeb, FinePDFLatent Space
videoWhat happens now that AI is good at math? — the OpenAI Podcast Ep. 17OpenAI
videoStanford CS25: Transformers United V6 I From Next-Token Prediction to Next-Generation IntelligenceStanford Online
Checking sign-in…
Loading comments…



