Research map / Large language models / Agents, code and math
Post-Training LLMs as Better Decision-Making Agents: A Regret-Minimization Approach
Chanwoo Park, Ziyang Chen, Asuman E. Ozdaglar, Kaiqing Zhang
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 |
|---|---|---|---|
| post-train with Iterative RMFT | Gemma-2-9b-it · FOL Uniform reward (T=50) | avg(LR): 28.47 -> 12.05 | Improved |
| post-train with Iterative RMFT | Gemma-2-9b-it · FOL Gaussian reward (T=50) | avg(LR): 20.62 -> 20.45 | No clear effect |
| post-train with Iterative RMFT | Gemma-2-9b-it · FOL Gaussian reward (T=50) | max(LR): 137.19 -> 93.08 | Improved |
| post-train with Iterative RMFT | Gemma-2-9b-it · FOL Sine-trend reward (T=50) | avg(LR): 18.00 -> 13.28 | Improved |
| post-train with Iterative RMFT | Gemma-2-9b-it · FOL Uniform reward (T=50) | regret growth rate: 0.81 -> 0.57 | Improved |
| post-train with Iterative RMFT | GPT-4o mini · FOL Uniform reward (T=25) | avg(LR): 22.28 -> 17.09 | Improved |
| post-train with Iterative RMFT | GPT-4o mini · FOL Uniform reward (T=25) | max(LR): 70.82 -> 39.65 | Improved |
| post-train with Iterative RMFT | GPT-4o mini · FOL Gaussian reward (T=25) | avg(LR): 27.42 -> 25.63 | Improved |
| post-train with Iterative RMFT | GPT-4o mini · FOL Sine-trend reward (T=25) | avg(LR): 11.24 -> 9.83 | Improved |
| select low-regret trajectories (RMFT) instead of imitating UCB with algorithm distillation on the same selected-data budget (ablation) | Qwen3-8B · MAB (T=100) | average regret: 58.72 -> 22.37 | Improved |
| select low-regret trajectories (RMFT) instead of AD with matched generation budget (ablation) | Qwen3-8B · MAB (T=100) | average regret: 33.44 -> 22.37 | Improved |
| select low-regret trajectories (RMFT) instead of AD with twice the generation budget (ablation) | Qwen3-8B · MAB (T=100) | average regret: 32.97 -> 22.37 | Improved |
Closest papers by meaning
- Do LLM Agents Have Regret? A Case Study in Online Learning and Games (ICLR 2025 · similarity 0.92)
- Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning (NeurIPS 2025 · similarity 0.90)
- Enhancing Decision-Making of Large Language Models via Actor-Critic (ICML 2025 · similarity 0.90)
- Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.89)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.89)
- Efficient Reinforcement Learning with Large Language Model Priors (ICLR 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.
