Research map / Large language models / Agents, code and math
LLMs are Greedy Agents: Effects of RL Fine-tuning on Decision-Making Abilities
Thomas Schmied, Jörg Bornschein, Jordi Grau-Moya, Markus Wulfmeier, Razvan Pascanu
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 |
|---|---|---|---|
| apply RLFT on self-generated CoT rationales | Gemma2 2B/9B · Gaussian button MAB (sigma=1) | cumulative regret | Improved |
| apply RLFT | Gemma2 2B · Button MAB 10 arms | action coverage: +12% after 30K updates | Improved |
| apply RLFT | Gemma2 2B · Frequency bias, 0-10 repetitions | frequent-action fraction: 70% -> 35% | Improved |
| apply RLFT | Gemma2 2B · Frequency bias, high repetitions | frequent-action fraction | No clear effect |
| apply RLFT | Gemma2 2B · Tic-tac-toe vs random | average return: 0.15 -> 0.75 | Improved |
| apply RLFT | Gemma2 2B · Tic-tac-toe vs MCTS | average return: -0.95 -> 0.0 | Improved |
| add exploration bonus (+1 for untried actions) during RLFT (ablation) | RLFT Gemma2 2B · MAB | action coverage: 50% -> 70% | Improved |
| try all actions initially (ablation) | Gemma2 27B · MAB | cumulative regret | Improved |
| use CoT in RLFT (ablation) | Gemma2 2B · 10-arm Gaussian MAB | regret | Improved |
| use CoT in ICL (ablation) | Gemma2 · Button MAB 10 arms | action coverage: 25% -> up to 65% | Improved |
| increase generation budget G from 256 to 512 (ablation) | RLFT Gemma2 2B · MAB | performance | Improved |
| decrease generation budget G (ablation) | RLFT Gemma2 2B · MAB | performance | Hurt |
Closest papers by meaning
- Tricks or Traps? A Deep Dive into RL for LLM Reasoning (ICLR 2026 · similarity 0.91)
- Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.90)
- Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs (ICLR 2026 · similarity 0.89)
- Do LLM Agents Have Regret? A Case Study in Online Learning and Games (ICLR 2025 · similarity 0.89)
- Exploration vs Exploitation: Rethinking RLVR through Clipping, Entropy, and Spurious Reward (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.
