Research map / Large language models / Instruction and fine-tuning
Reasoning Quality Emerges Early: Data Curation for Reasoning Models
Hongyi Henry Jin, Wenhan Yang, Meysam Ghaffari, Carlos Morato, Baharan Mirzasoleiman
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add difficulty filtering using first-100-token loss at a randomly perturbed checkpoint (ablation) | TEMP · M23k, Qwen2.5-7B-Instruct, 1k | average accuracy (%): 54.09 -> 55.02 | Improved |
| add source budget weighting toward harder sources (ablation) | TEMP · M23k, Qwen2.5-7B-Instruct, 1k | average accuracy (%): 55.02 -> 55.28 | Improved |
| add diversity sampling via clustering of perturbed loss vectors (ablation) | TEMP · M23k, Qwen2.5-7B-Instruct, 1k | average accuracy (%): 55.28 -> 56.52 | Improved |
| add brittle example selection within clusters (ablation) | TEMP · M23k, Qwen2.5-7B-Instruct, 1k | average accuracy (%): 56.52 -> 57.69 | Improved |
| use an undertrained fine-tuned checkpoint to compute the fine-tuning direction (ablation) | TEMP · M23k and OpenThoughts-Math | accuracy | Hurt |
| use an inexact but reasonable fine-tuned checkpoint for direction (ablation) | TEMP · M23k and OpenThoughts-Math | accuracy | No clear effect |
| vary number of data picked per cluster k (ablation) | TEMP · M23k | average accuracy | No clear effect |
| cluster on first 1k reasoning tokens instead of the full response (ablation) | TEMP · M23k | average accuracy | No clear effect |
| use 1 or 2 perturbed checkpoints for difficulty filtering (ablation) | TEMP · M23k | average accuracy | No clear effect |
| use TEMP instead of S2L, learnability, perplexity, embedding diversity, M1k or random | SFT data selection · M23k, Llama3.1-8B-Instruct, 1k | average accuracy (%): up to +1.7% | Improved |
| use S2L loss-based selection on Qwen2.5 | SFT data selection · M23k, Qwen2.5-7B-Instruct, 1k | average accuracy | No clear effect |
| compute losses only on initial tokens rather than full traces | TEMP · OpenThoughts-Math selection | tokens processed (million): 352.9 -> 30.3 | Improved |
Closest papers by meaning
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.90)
- The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models (NeurIPS 2025 · similarity 0.90)
- Hint Tuning: Less Data Makes Better Reasoners (NeurIPS 2026 · similarity 0.90)
- UFT: Unifying Supervised and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.90)
- Training Large Language Models To Reason In Parallel With Global Forking Tokens (ICLR 2026 · similarity 0.89)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
