arxiv Moonshot AI (arXiv) · 555d ago

Simple Scaling Lab releases s1-32B with test-time scaling via budget forcing

from English

The Simple Scaling lab introduces s1-32B, a reasoning model that achieves strong performance through supervised fine-tuning on a curated dataset of 1,000 samples and a technique called budget forcing. This approach controls test-time compute by either terminating the model's thinking process or appending "Wait" to extend it, allowing the model to double-check its answers.

The s1K dataset contains 1,000 questions paired with reasoning traces distilled from Google's Gemini Thinking Experimental model, selected for difficulty, diversity, and quality.
Budget forcing allows precise control over test-time compute; suppressing the end-of-thinking token and adding "Wait" encourages further exploration of the solution path.
s1-32B exceeds OpenAI’s o1-preview on competition math questions by up to 27% on MATH and AIME24 benchmarks.
Scaling test-time compute allows the model to extrapolate its performance from 50% to 57% on AIME24.

The authors demonstrate that simple supervised fine-tuning combined with direct control over thinking duration is a highly effective and sample-efficient method for achieving test-time scaling and strong reasoning capabilities.

Importance 2/3 Beats a top-lab benchmark Moonshot AI (arXiv) DeepSeek Google DeepMind Reasoning models Research paper

Benchmarks

Benchmark	Model	Score
AIME 2024	s1-32B	57%

Related articles

arxiv arXiv cs.AI · 51d ago

Lean as Process-Verified Reward Oracle in RL for Theorem Proving

This work shows that Lean can serve as a symbolic process oracle, providing fine-grained, verified feedback during reinforcement learning. By parsing proof attempts into tactic sequences and using Lean's elaboration to mark sound steps and first failures, the system generates dense, type-theoretic reward signals. Experiments demonstrate tactic-level supervision outperforms outcome-only methods on benchmarks like MiniF2F and ProofNet, highlighting Lean's role as both evaluator and training reward source.

arxiv arXiv cs.LG · 52d ago

Spotlight: Using Spot GPUs to Accelerate DiT RL Post-Training

Spotlight enables DiT RL post-training by leveraging idle spot GPUs, reducing costs by 1.4-6.4x while achieving superior image quality. It uses stale model weights in exploration and reconfigures sequence parallelism on-the-fly, allowing efficient GPU utilization without breaking training pipelines.

arxiv Moonshot AI (arXiv) · 916d ago

DeepSeekMath 7B achieves 51.7% on MATH benchmark using GRPO and curated data

The DeepSeekMath paper introduces a 7B parameter open language model designed to enhance mathematical reasoning capabilities. The model is created by continuing the pre-training of DeepSeek-Coder-Base-v1.5 with 120B math-related tokens from Common Crawl, alongside natural language and code data.

lab Google DeepMind Blog · 3d ago · 16 views

Google DeepMind open sources WeatherNext AI model for cyclone forecasting

Google DeepMind and Google Research have open sourced the WeatherNext 2 and WeatherNext Cyclones AI models, which achieved state-of-the-art accuracy in predicting tropical cyclone tracks, intensity, and wind structure. Published in Nature, the work demonstrates that these models provide forecasters with an extra day of predictive accuracy compared to prior systems.

media Don't Worry About the Vase · 5d ago · 5 views

OpenAI's Astra solves 10 major open mathematics problems

OpenAI has released results from its unreleased internal model, Astra, which successfully solved ten significant open problems in mathematics. The solutions were generated by the model and subsequently formalized into Lean certificates by human researchers.