Papers5

#Credit Assignment

Blockwise Advantage Estimation for Multi-Objective RL with Verifiable Rewards

Kirill Pavlenko, Alexander Golubev et al.Feb 10arXiv

The paper fixes a common mistake in training language models for multi-part tasks: giving the same reward signal to every token, even when different text parts aim at different goals.

#Blockwise Advantage Estimation#Outcome-Conditioned Baseline#Group Relative Policy Optimization

MatchTIR: Fine-Grained Supervision for Tool-Integrated Reasoning via Bipartite Matching

Intermediate

Changle Qu, Sunhao Dai et al.Jan 15arXiv

MatchTIR teaches AI agents to judge each tool call step-by-step instead of giving the same reward to every step.

#Tool-Integrated Reasoning#Credit Assignment#Bipartite Matching

AT$^2$PO: Agentic Turn-based Policy Optimization via Tree Search

Intermediate

Zefang Zong, Dingwei Chen et al.Jan 8arXiv

AT2PO is a new way to train AI agents that work in several turns, like asking the web a question, reading the result, and trying again.

#Agentic Reinforcement Learning#Turn-level Optimization#Tree Search

VAR RL Done Right: Tackling Asynchronous Policy Conflicts in Visual Autoregressive Generation

Intermediate

Shikun Sun, Liao Qu et al.Jan 5arXiv

Visual Autoregressive (VAR) models draw whole grids of image tokens at once across multiple scales, which makes standard reinforcement learning (RL) unstable.

#Visual Autoregressive (VAR)#Reinforcement Learning#GRPO

E-GRPO: High Entropy Steps Drive Effective Reinforcement Learning for Flow Models

Intermediate

Shengjun Zhang, Zhang Zhang et al.Jan 1arXiv

This paper shows that when teaching image generators with reinforcement learning, only a few early, very noisy steps actually help the model learn what people like.

#E-GRPO#Group Relative Policy Optimization#Flow Matching