AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Language-based Trial and Error Falls Behind in the Era of Experience

Haoyu Wang, Guozheng Ma, Shugang Cui, Yilun Kong, Haotian Luo, Li Shen, Mengya Gao, Yichao Wu, Xiaogang Wang, Dacheng Tao

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 0 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
add multi-turn PPO Evolving Stage after scout distillation (ablation)SCOUT (Qwen2.5-0.5B-It) · Average over 6 unseen tasksaverage success rate: 0.69 -> 0.81Improved
add multi-turn PPO Evolving Stage after scout distillation (ablation)SCOUT (Qwen2.5-1.5B-It) · Average over 6 unseen tasksaverage success rate: 0.71 -> 0.85Improved
add multi-turn PPO Evolving Stage after scout distillation (ablation)SCOUT (Qwen2.5-3B-It) · Average over 6 unseen tasksaverage success rate: 0.73 -> 0.86Improved
add multi-turn PPO Evolving Stage after scout distillation (ablation)SCOUT (Qwen2.5-7B-It) · Average over 6 unseen tasksaverage success rate: 0.78 -> 0.88Improved
add multi-turn PPO Evolving Stage after scout distillation (ablation)SCOUT (LLaMA3.1-1B-It) · Average over 6 unseen tasksaverage success rate: 0.63 -> 0.83Improved
add scout-trajectory Exploration & Distillation warm-up before multi-turn PPOQwen2.5-3B-It · Average over 6 unseen tasksaverage success rate: 0.38 -> 0.86Improved
add Evolving Stage RL after distillation on Sudoku (ablation)Qwen2.5-3B-It · Sudokusuccess rate: 0.29 -> 0.97Improved
use DQN instead of PPO to train the scoutScouts · Average over 6 unseen tasksaverage success rate: 0.79 -> 0.83Mixed
add scout warm-up with a sub-optimal scout instead of direct RLQwen2.5-3B-It · Rubik's Cube Rotation3success rate: 0.14 -> 0.60Improved
add scout warm-up with a sub-optimal scout instead of direct RLQwen2.5-3B-It · Sudokusuccess rate: 0.06 -> 0.98Improved
use SCOUT instead of direct RL on vision inputsQwen2.5-VL-3B-It · Visual-Sokobansuccess rate: 0.57 -> 0.95Improved
use SCOUT instead of direct PPO for training costQwen2.5-3B-It · Rubiks' Cube Rotation3 training costGPU hours: 24.0 -> 9.6Improved

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