AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

AReaL-DTA: Dynamic Tree Attention for Efficient Reinforcement Learning of Large Language Models

Jiarui Zhang, Yuchen Yang, Ran Yan, Zhiyu Mei, Liyuan Zhang, Daifeng Li, Wei Fu, Jiaxuan Gao, Shusheng Xu, Yi Wu, Binhang Yuan

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 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 tree traversal of prefix tree in backward pass (ablation)AReaL-DTA · tau2-bench single-GPU throughputspeedup vs Dense+CKPT: 7.53xImproved
add DFS traversal order (ablation)AReaL-DTA Tree · tau2-bench single-GPU throughputspeedup: 7.53x -> 7.74xImproved
add larger chunk block size (4096) (ablation)AReaL-DTA Tree+DFS · tau2-bench single-GPU throughputspeedup: 7.74x -> 8.31xImproved
add load-balanced prefix-tree partitioning (ablation)AReaL-DTA · tau2-bench data-parallel trainingwall-clock performance: 11.93%Improved
use AReaL-DTA instead of Sparse tree-mask with checkpointingtraining · tau2-benchthroughput: 1.52x-1.70xImproved
use AReaL-DTA instead of dense trainingQwen3-1.7B · tau2-benchpeak memory: over 50% reductionImproved
use AReaL-DTA instead of AReaL for end-to-end RLQwen3-8B · tau2-benchend-to-end throughput: 2.28xImproved
use AReaL-DTA instead of AReaL for end-to-end RLQwen3-1.7B · tau2-benchend-to-end throughput: 1.28xImproved
use AReaL-DTA instead of AReaL for end-to-end RLQwen3-1.7B/8B · tau2-benchreward vs training stepNo clear effect
shift GPU split to 5/3 rollout/training with AReaL-DTA (ablation)AReaL-DTA · tau2-bench Qwen3-8Bthroughput K tokens/s: 27.5 -> 51.4Improved
apply AReaL-DTA on low-compression GSM8K workloadQwen3-1.7B · GSM8K (C=1.09)relative throughput: 1.00x -> 0.62xHurt
apply AReaL-DTA on AIME workloadQwen3-1.7B · AIME (C=1.008)relative throughput: 1.00x -> 1.19xImproved

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