AI Paper Atlas

Research map / Large language models / Agents, code and math

R4: Nested Reasoning-Retrieval for Reward Modeling in Role-Playing Agents

Renzhi Wang, Chongqiang Wei, Zhisheng Wang, Piji Li

ICLR 2026 · Large language models › Agents, code and math · approach: Long-horizon LLM agents

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 changedWhereMeasuredEffect
add reasoning to reward model and agent (ablation)R4 · CharacterEval Character Consistencyaverage score: 48.23 -> 64.64Improved
add retrieval to reward model and agent (ablation)R4 · CharacterEval Character Consistencyaverage score: 51.67 -> 64.64Improved
use character-specific instead of generic retrieval (ablation)R4 · CharacterEval Character Consistencyaverage score: 47.95 -> 64.64Improved
add reasoning to the reward model (ablation)R4 · CharacterEval Character Consistencyaverage score: 49.87 -> 64.64Improved
add retrieval to the reward model (ablation)R4 · CharacterEval Character Consistencyaverage score: 58.31 -> 64.64Improved
add reasoning to the agent (ablation)R4 · CharacterEval Character Consistencyaverage score: 56.94 -> 64.64Improved
add retrieval to the agent (ablation)R4 · CharacterEval Character Consistencyaverage score: 55.47 -> 64.64Improved
add consistency reward to RM (ablation)R4 · CharacterEval Character Consistencyaverage score: 59.83 -> 64.64Improved
add format reward to RM (ablation)R4 · CharacterEval Character Consistencyaverage score: 62.41 -> 64.64Improved
use R4 reasoning-retrieval RM instead of scalar RM (ablation)R4 · CharacterEval Character Consistencyaverage score: 43.89 -> 64.64Improved
use R4 RM instead of CharacterRM (ablation)R4 · CharacterEval Character Consistencyaverage score: 46.87 -> 64.64Improved
scale agent from 7B to 32BR4 · CharacterEval Character Consistencyaverage score: 55.74 -> 64.64Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy