Research map / Large language models / Instruction and fine-tuning
Distributional Alignment Games for Answer-Level Fine-Tuning
Mehryar Mohri, Jon Schneider, Yifan Wu
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 |
|---|---|---|---|
| fine-tune with Pairwise-GRPO | Qwen2.5-3B-Instruct · GSM8K | accuracy (%): 75.06 -> 79.61 | Improved |
| fine-tune with Pairwise-GRPO | Llama-3.2-3B-Instruct · GSM8K | accuracy (%): 66.19 -> 72.71 | Improved |
| fine-tune with Pairwise-GRPO | Phi-3-mini-4k-instruct · GSM8K | accuracy (%): 73.69 -> 82.87 | Improved |
| fine-tune with Coherence-GRPO | Qwen2.5-3B-Instruct · GSM8K | accuracy (%): 75.06 -> 80.36 | Improved |
| fine-tune with Coherence-GRPO | Llama-3.2-3B-Instruct · GSM8K | accuracy (%): 66.19 -> 69.37 | Improved |
| fine-tune with Coherence-GRPO | Phi-3-mini-4k-instruct · GSM8K | accuracy (%): 73.69 -> 81.50 | Improved |
| fine-tune with Pairwise-GRPO | Qwen2.5-3B-Instruct · TriviaQA | exact match: 32.95 -> 35.32 | Improved |
| fine-tune with Pairwise-GRPO | Llama-3.2-3B-Instruct · TriviaQA | exact match: 39.12 -> 47.85 | Improved |
| fine-tune with Pairwise-GRPO | Phi-3-mini-4k-instruct · TriviaQA | exact match: 32.03 -> 45.50 | Improved |
| fine-tune with Coherence-GRPO | Qwen2.5-3B-Instruct · TriviaQA | exact match: 32.95 -> 32.90 | No clear effect |
| fine-tune with Coherence-GRPO | Llama-3.2-3B-Instruct · TriviaQA | exact match: 39.12 -> 40.25 | Improved |
| fine-tune with Coherence-GRPO | Phi-3-mini-4k-instruct · TriviaQA | exact match: 32.03 -> 32.00 | No clear effect |
Closest papers by meaning
- Back to Blackwell: Closing the Loop on Intransitivity in Multi-Objective Preference Fine-Tuning (NeurIPS 2026 · similarity 0.87)
- The Consensus Game: Language Model Generation via Equilibrium Search (ICLR 2024 · similarity 0.86)
- Self-Aligned Reward: Towards Effective and Efficient Reasoners (ICLR 2026 · similarity 0.86)
- Offline RL by Reward-Weighted Fine-Tuning for Conversation Optimization (NeurIPS 2025 · similarity 0.86)
- Reasoning Elicitation in Language Models via Counterfactual Feedback (ICLR 2025 · similarity 0.86)
- From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium (ICML 2025 · similarity 0.85)
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.
