Research map / Large language models / Reasoning and chain of thought

OpenThoughts: Data Recipes for Reasoning Models

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematic ablation study (1,000+ controlled experiments) of reasoning SFT data pipeline choices, yielding the OpenThoughts3-1.2M dataset and OpenThinker3-7B, a state-of-the-art open-data 7B reasoning model.

Problem

Open reasoning SFT datasets are built with many heuristic design choices, and the best recipe for question sourcing, filtering, answer generation, verification, and teacher selection is unclear because proprietary recipes are not public and ablations are expensive.

Approach

Ablate each stage of the reasoning data pipeline (question sourcing, mixing, filtering, dedup and repeated sampling, answer filtering, teacher choice) independently at 31.6K scale on Qwen2.5-7B-Instruct, greedily keep the best option per stage, and scale the resulting recipe to 1.2M examples.

Key results

• OpenThinker3-7B reaches 53.3% on AIME25, 51.7% on LCB 06/24-01/25, and 53.7% on GPQA Diamond, beating DeepSeek-R1-Distill-Qwen-7B by large margins.
• Using two high-quality code question sources instead of 16 yields about a 5% average accuracy increase in the pipeline ablations.
• QwQ-32B as teacher outperforms DeepSeek-R1 by 1.9 points on code and 2.6 points on math averages despite scoring lower on benchmarks.

Surprising findings

• A weaker teacher (QwQ-32B) yields better student performance than the stronger DeepSeek-R1.
• Answer filtering (majority consensus, verification, length-based) did not beat training on all unfiltered answers.

Limitations

• No exploration of RL datasets or RL training; the work is SFT-only.
• No staged SFT or curriculum learning explored.
• Step-wise greedy selection assumes cross-domain transfer and may not be globally optimal.

Experimental setup

Datasets: OpenThoughts3-1.2M OpenMath-2-Math NuminaMath-1.5 OpenCodeReasoning StackExchange-CodeGolf StackExchange-Physics OrganicChemistry-PDF OpenR1-Math AutoMathText CodeFeedback MathPile KodCode-V1

Benchmarks: AIME24 AMC23 MATH500 CodeElo CodeForces LiveCodeBench 05/23-05/24 GPQA Diamond JEEBench AIME25 HMMT 02/25 HLE MCQ LiveCodeBench 06/24-01/25

Metrics: accuracy average accuracy across benchmarks standard error

Baselines: DeepSeek-R1-Distill-Qwen-7B OpenR1-Distill-7B Nemotron-Nano-8B AceReason-7B Skywork-7B AM-1.4M NemoNano-1M s1.1 LIMO

Builds on: Sky-T1 pipeline BespokeStratos-17K OpenThoughts-114K OpenThoughts2-1M LLaMA-Factory Evalchemy

Closest papers by meaning

• OpenMathInstruct-2: Accelerating AI for Math with Massive Open-Source Instruction Data (ICLR 2025 · similarity 0.90)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.89)
• Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.87)
• SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond (NeurIPS 2025 · similarity 0.86)
• OpenMMReasoner: Pushing the Frontiers in Multimodal Reasoning with an Open and General Recipe (CVPR 2026 · similarity 0.86)
• GRIP: A Graph-Based Reasoning Instruction Producer (NeurIPS 2025 · similarity 0.86)