Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning
Authors: Feng Chen, Allan Raventós, Nan Cheng, Surya Ganguli, Shaul Druckmann
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our algorithm demonstrates improved mathematical reasoning on MATH and Mini F2F benchmarks under several scenarios: (1) providing answers to math questions; and (2) proving theorems by searching over proof trees of varying shapes. Overall our work underscores the importance of co-designing two traditionally separate phases of LLM development: training-time protocols and test-time search and reasoning strategies. |
| Researcher Affiliation | Academia | Feng Chen Stanford University Stanford, CA 94305 EMAIL Allan Raventós* Stanford University Stanford, CA 94305 EMAIL Nan Cheng University of Michigan Ann Arbor, MI 48109 EMAIL Surya Ganguli Stanford University Stanford, CA 94305 EMAIL Shaul Druckmann Stanford University Stanford, CA 94305 EMAIL |
| Pseudocode | No | The paper describes algorithms (DCO, DCOa, DCOstep) and provides mathematical equations for them, but does not present them in a structured pseudocode block or algorithm box. |
| Open Source Code | No | Our codebase uses Py Torch [47], Accelerate [48], and deepspeed (https://github.com/ microsoft/Deep Speed) to enable efficient training with memory constraints, and vllm [49] for efficient inference. Code will be made available here (https://github.com/allanraventos/refine). |
| Open Datasets | Yes | For experiments with MATH dataset [43] (license: MIT), we fine-tune the Llama-3-8B-base [42] on the MATH [43] dataset. We use Lean Dojo benchmark[44] (license: CC-BY 2.0) and adopt the random train and test split as introduced in Yang et al. [44]. |
| Dataset Splits | Yes | We follow Lightman et al. [24] and use 12,000 problems for training and the remaining 500 for testing. ... We use Lean Dojo benchmark[44] (license: CC-BY 2.0) and adopt the random train and test split as introduced in Yang et al. [44]. The random test set includes 2,000 theorems. |
| Hardware Specification | Yes | All experiments with model size smaller than 10B are performed on machines with 8 NVIDIA H100 GPUs or 8 NVIDIA A100 GPUs. Experiments with model size larger than 10B are performed on machines with 8 NVIDIA H200 GPUs or 8 NVIDIA B200 GPUs. |
| Software Dependencies | No | Our codebase uses Py Torch [47], Accelerate [48], and deepspeed (https://github.com/ microsoft/Deep Speed) to enable efficient training with memory constraints, and vllm [49] for efficient inference. |
| Experiment Setup | Yes | In Sections 4 and 5.2, Figure 1, Figure 2 (a) and (b), we fine-tune the model for 4 epochs with a learning rate of 2e-5 and batch size 64. We adopt a linear learning rate warmup in the first 20 steps. ... For the Co T experiments in Section 5.4, we use learning rate 2e-5 and batch size 128, with the same learning rate warmup. ... We fine-tune the model Qwen2.5-Math-1.5B [28] on the training set for 3 epochs with learning rate 1e-5 and batch size 64. We adopt a linear learning rate warmup in the first 20 steps. |