Research map / Large language models / Instruction and fine-tuning
Universal Reasoner: A Single, Composable Plug-and-Play Reasoner for Frozen LLMs
Jaemin Kim, Hangeol Chang, Hyunmin Hwang, Choonghan Kim, Jong Chul Ye
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 |
|---|---|---|---|
| train the reasoning module with GRPO and add its logits to a frozen backbone, instead of GRPO full fine-tuning | Qwen2.5-3B · Math average (T=0.6) | average pass@1: 40.3 -> 42.7 | Improved |
| same, on Llama backbone | Llama3.2-3B · Math average (T=0.6) | average pass@1: 25.4 -> 33.3 | Improved |
| add reward-trained reasoning module vs naive logit ensemble of pretrained module (ablation) | Qwen2.5-3B · Math average (T=0.6) | average pass@1: 35.2 -> 42.7 | Improved |
| use UniR instead of GRPO Full for translation | Llama3.2-3B · IWSLT2017 EN-DE | BLEU: 25.26 -> 27.03 | Improved |
| raise sampling temperature from 0 to 0.6 (ablation) | GRPO Full (Llama3.2-3B) · MATH-500 | pass@1: 42.4 -> 36.1 | Hurt |
| raise sampling temperature from 0 to 0.6 (ablation) | UniR (Llama3.2-3B) · MATH-500 | pass@1: 49.2 -> 48.8 | No clear effect |
| transfer 3B-trained reasoning module to a larger frozen backbone | Qwen2.5-14B · Math average | average pass@1: 51.1 -> 53.6 | Improved |
| add translation module to math task with German inputs | frozen backbone on MGSM German · German math problems to English | math accuracy | Improved |
| increase weight alpha of math module versus translation module (ablation) | composed UniR modules · German-to-English math task | math accuracy vs translation quality | Mixed |
| text-only trained math module guiding a VLM | Qwen2.5-VL-3B · MathVerse vision intensive and Geometry3k | accuracy | Improved |
| compose readmission and LOS modules (ablation) | UniR on MIMIC-IV · readmission accuracy | accuracy: 52.91 -> 44.88 accuracy; F1 26.77 -> 28.79 | Mixed |
| further train pretrained DeepSeek 1.5B module with UniR (ablation) | Qwen2.5-3B backbone · Math average | average pass@1: 43.7 -> 46.9 | Improved |
Closest papers by meaning
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.91)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.89)
- Tina: Tiny Reasoning Models via LoRA (ICLR 2026 · similarity 0.89)
- SABER: Switchable and Balanced Training for Efficient LLM Reasoning (AAAI 2026 · similarity 0.89)
- Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning (ICLR 2026 · 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.
