AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Text2Grad: Reinforcement Learning from Natural Language Feedback

Hanyang Wang, Lu Wang, Chaoyun Zhang, Tianjun Mao, Si Qin, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 15 citations (OpenAlex)

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
span-level NL-feedback gradients (Text2Grad) instead of scalar-reward PPOLlama-3.1-8B-Instruct · SLF5KROUGE-L: 0.262 -> 0.291Improved
Text2Grad instead of PPO on codeLlama-3.1-8B-Instruct · KodCode (avg pass@1)pass@1: 62.5 -> 66.1Improved
Text2Grad instead of PPO on QALlama3-8B-Instruct · UltraFeedback AlpacaEval 2.0win rate: 32.4 -> 34.7Improved
Text2Grad instead of PPO/DPO on Mistral-7BMistral-7B-Instruct-v0.2 · MT-Benchscore: PPO 6.55 -> 6.78; base 6.60Mixed
add CoT reasoning before span annotation (ablation)Text2Grad · SLF5KROUGE-L: 0.275 -> 0.291Improved
add CoT reasoning before span annotation (ablation)Text2Grad · KodCode (avg pass@1)pass@1: 59.2 -> 66.1Improved
add CoT reasoning before span annotation (ablation)Text2Grad · UltraFeedback AlpacaEval 2.0win rate: 28.6 -> 34.7Improved
use span-level labels instead of dense token-level labels (ablation)Text2Grad · SLF5KROUGE-L: 0.196 -> 0.291Improved
add CoT to reward model annotation (span precision) (ablation)reward model · SLF5K positive-token precision/recallprecision/recall: precision 0.63 -> 0.58; recall 0.46 -> 0.63Mixed
add CoT to reward model (ablation)reward model · UltraFeedback positive-token recallrecall: 0.59 -> 0.43Hurt
Text2Grad on code vs ablated no CoT on ARC-Challenge (ablation)Text2Grad · UltraFeedback ARC-Caccuracy: 83.1 -> 84.4Improved

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