4B Model Hits 54% on IMO-ProofBench, Costing 3x Less Per Problem
SergioPaniego · x · 2026-08-22
A 4B parameter model achieved 54% accuracy on IMO-ProofBench, competing with models 7-50x its size while costing 3x less per problem.
Key techniques include:
- Pure post-training: No pre-training compute required.
- Distillation SFT: Supervised fine-tuning via knowledge distillation.
- GRPO: Reinforcement learning optimization with rubric rewards.
- Reasoning Cache: Test-time refinement using a caching mechanism.
More from Models
- Scale's Muse Spark Open claims strongest US open-weight model: 816B params, 1M context — bansalg_ · 2026-10-16
- Zuck announces Meta Muse Spark Open weights: 816B-param image-text model tuned for Mac Studio — bansalg_ · 2026-10-16
- Google launches revamped Gemma docs hub covering fine-tuning, inference and model internals — _philschmid · 2026-10-16
- Post-trained NVIDIA Nemotron-3.5-Super-VL hits 35% on AutoResearchExam, matching a 20× larger model — AlexGDimakis · 2026-10-16
- Nathan Lambert slams Meta's Muse Open License as 'trash' — natolambert · 2026-10-16
- Google launches Gemini 4 Argon to Ultra and Pro subscribers, touting physics and 3D skills — ShikharMurty · 2026-10-16