AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

ATPO: ADAPTIVE TREE POLICY OPTIMIZATION FOR MULTI-TURN MEDICAL DIALOGUE

Ruike Cao, Shaojie Bai, Fugen Yao, Liang Dong, Jian Xu, Li Xiao

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 5 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 uncertainty-based adaptive pruning/expansion (ATPO U1+U2) over fixed binary TreePOQwen3-8B · MedQAaccuracy (%): 61.81 -> 64.07Improved
add uncertainty-based adaptive pruning/expansion (ATPO U1+U2) over TreePOQwen3-4B · MedQAaccuracy (%): 53.74 -> 55.47Improved
add uncertainty-based adaptive pruning/expansion (ATPO U1+U2) over TreePOQwen3-1.7B · MedQAaccuracy (%): 42.05 -> 42.87Improved
add second uncertainty metric U2 (variance) to U1 (ablation)ATPO Qwen3-8B · MedMCQAaccuracy (%): 53.22 -> 53.66Mixed
add second uncertainty metric U2 (variance) to U1 (ablation)ATPO Qwen3-1.7B · MedicalExamaccuracy (%): 45.73 -> 43.20Hurt
use tree search credit assignment (TreePO) instead of GRPO trajectory-level advantageQwen3-8B · MedQAaccuracy (%): 57.92 -> 61.81Improved
use hierarchical MDP instead of token-level MDP for PPOPPO Qwen3-8B · MedQAaccuracy (%): 57.38 -> 57.81No clear effect
add interactive MEDIQ prompting over direct single-turn promptQwen3-8B zero-shot · MedQAaccuracy (%): 45.22 -> 46.03 on MedQA; MedMCQA 46.16 -> 41.60Hurt
distill from GPT-4o or Gemini-2.5-Pro dialogues in SFTQwen3-8B · MedQAaccuracy (%): 53.25 and 51.51 vs SFT 53.75No clear effect
add RL (ATPO) on top of SFTQwen3-8B · MedQAaccuracy (%): 53.75 -> 64.07Improved
down-weight policy advantage by node visit count (vs no down-weighting, EXP1) (ablation)ATPO · training stabilitystabilityImproved
also down-weight value loss by visit count (EXP2) (ablation)ATPO · training stabilitystabilityHurt

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