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..
Preference Distillation via Value based Reinforcement Learning
Authors: Minchan Kwon, Junwon Ko, Kangil kim, Junmo Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results show that TVKD consistently improves performance across various benchmarks and model sizes. Section 4: Experiments. Section 4.1: Experiments Setting. Section 4.2: Main Results. Section 4.3: Analysis. |
| Researcher Affiliation | Academia | 1Korea Advanced Institute of Science and Technology (KAIST) 2Gwangju Institute of Science and Technology (GIST) EMAIL, EMAIL, |
| Pseudocode | No | The paper describes mathematical derivations and theoretical concepts in sections like 'Mathematical Derivation' and 'Proof of Lemma', but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | We will also release the code. We publish our code in supplementary. |
| Open Datasets | Yes | For preference distillation, we use two datasets: (1) DPO-MIX-7K2, a curated collection of high-quality pairwise preference data, and (2) Help Steer2 [27], which is designed to improve helpfulness in LLMs. 2https://huggingface.co/datasets/argilla/dpo-mix-7k |
| Dataset Splits | No | The paper mentions using a 'test set for evaluation' and a 'validation set for checkpoint selection' but does not provide specific details on the dataset split percentages, sample counts, or the methodology used to create these splits for the DPO-MIX-7K and Helpsteer2 datasets. |
| Hardware Specification | Yes | We conducted our experiments using four Nvidia RTX 4090 GPUs. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The SFT for both student and teacher models is conducted over 3 epochs, using a learning rate of 2 10 5 and a batch size of 128. The DPO teacher is trained with β = 0.05, a learning rate of 5 10 7, a batch size of 128, and for 2 epochs. ... In TVKD, we experiment with α [0.1, 0.2, 0.5, 0.7, 1.0, 1.5] and β [0.1, 0.2, 0.5, 1, 2, 5]. |