Research map / Reinforcement learning / RL from feedback
Reward Design with Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use LLM proxy reward (few-shot prompting) instead of SL reward model | DealOrNoDeal · DealOrNoDeal | RL agent accuracy: 46% on average over SL | Helped |
| add explanations to the single-example prompt (ablation) | LLM proxy reward, Ultimatum Game · Ultimatum Game | labeling accuracy and RL agent accuracy: avg. drop of 31.67% labeling and 28.8% RL agent accuracy when removed | Helped |
| use LLM proxy reward with 10 examples vs SL trained on same 10 examples | Ultimatum Game · Ultimatum Game | labeling accuracy | No clear effect |
| use single example with explanation in LLM prompt vs SL trained on single example | Ultimatum Game · Ultimatum Game | labeling accuracy | Helped |
| add solution concept description and zero-shot objective in prompt vs No Objective baseline (ablation) | Matrix Games zero-shot LLM reward · Matrix Games | labeling accuracy: 48% on average (regular order); 36% (scrambled) | Helped |
| scramble joint action-reward associations in matrix game prompt (ablation) | Matrix Games zero-shot LLM reward · Matrix Games | labeling accuracy | Hurt |
| use LLM reward vs SL in DealOrNoDeal for Versatile style | DealOrNoDeal · DealOrNoDeal Versatile | labeling accuracy | No clear effect |
| replace GPT-3 (175B) with GPT-2 (1.5B) as the LLM (ablation) | Ultimatum Game / DealOrNoDeal proxy reward · Ultimatum Game and DealOrNoDeal | labeling accuracy and RL agent accuracy: underperforms GPT-3 in labeling by avg 15% and RL agent accuracy by avg 49% | Hurt |
| remove task description rho_1 from single-example Ultimatum prompt (ablation) | LLM proxy reward, Ultimatum Game · Ultimatum Game | labeling accuracy: +8% labeling and +9% RL agent accuracy (single example); -4.4% labeling (10 examples) | No clear effect |
| vary keyword, example negotiations, and explanations in DealOrNoDeal prompt (ablation) | LLM proxy reward, DealOrNoDeal Stubborn · DealOrNoDeal Stubborn | labeling accuracy | No clear effect |
| remove explanation quality (vary explanations) in DealOrNoDeal prompt (ablation) | LLM proxy reward, DealOrNoDeal Stubborn · DealOrNoDeal Stubborn | labeling accuracy | Helped |
| train agents with user-provided examples in prompt (correct style) vs flipped labels (opposite style) | DealOrNoDeal pilot user study · Pilot user study (N=10) | user alignment rating (1-5): 3.72 ± 1.2 vs 1.56 ± 1.05, p < 0.001 | Helped |
Closest papers by meaning
- REvolve: Reward Evolution with Large Language Models using Human Feedback (ICLR 2025 · similarity 0.90)
- On the Modeling Capabilities of Large Language Models for Sequential Decision Making (ICLR 2025 · similarity 0.89)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.89)
- RLZero: Direct Policy Inference from Language Without In-Domain Supervision (NeurIPS 2025 · similarity 0.89)
- ToolRL: Reward is All Tool Learning Needs (NeurIPS 2025 · similarity 0.89)
- Reward-Guided Prompt Evolving in Reinforcement Learning for LLMs (ICML 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.