AIAny
Icon for item

APPO: Agentic Procedural Policy Optimization

Shifts branching and credit assignment in agentic RL from coarse units to fine-grained decision points in generated sequences. Uses a Branching Score combining token uncertainty and policy-induced likelihood gains plus procedure-level advantage scaling; improves performance across 13 benchmarks while keeping efficient tool calls.

Introduction

Most agentic RL methods assign credit over coarse units such as tool-call boundaries or fixed workflows, which hides which intermediate decisions actually drive downstream success. APPO's central insight is that influential decision points are widely distributed across generated sequences and that token entropy alone poorly predicts long-term impact — so branching and credit must move to finer procedural units.

Key Findings
  • Branching Score: combines token uncertainty with the policy-induced likelihood gain of subsequent continuations, so branching focuses on tokens that carry downstream value rather than merely high entropy. This reduces spurious exploration at noisy positions.
  • Procedure-level advantage scaling: rescales advantages across branched rollouts to distribute credit more faithfully across complete procedures, improving learning signal quality for multi-step behaviors.
  • Empirical gains: evaluated on 13 tasks spanning information seeking, knowledge-intensive reasoning, and computational problem solving; APPO yields consistent improvements (~+4 points over strong baselines) while preserving efficient tool usage and interpretable behavior.
Who It's For and Tradeoffs

Great fit if you train or research multi-turn LLM agents where delayed rewards and long decision chains make trajectory-level credit noisy — e.g., tool-using assistants, information retrieval agents, or program-synthesis workflows. Look elsewhere if you need an off-the-shelf drop-in for single-turn or token-level alignment problems: APPO targets procedural credit in interactive, multi-step settings and requires an agentic RL training pipeline and tuning (branching selection, advantage scaling). It emphasizes targeted exploration and interpretability but increases algorithmic complexity compared to simple trajectory-level baselines.

Where It Fits

APPO sits alongside other step- and graph-based credit-assignment approaches but differs by (1) selecting branching points at the token/procedure level using a value-aware Branching Score rather than fixed step boundaries, and (2) explicitly scaling advantages at the procedure level to encourage exploration of high-value continuations. This makes it a middle ground between token-centric RL and coarse trajectory attribution for agentic LLM training.

Information

  • Websitearxiv.org
  • AuthorsXucong Wang, Ziyu Ma, Yong Wang, Yuxiang Ji, Shidong Yang, Guanhua Chen, Pengkun Wang, Xiangxiang Chu
  • Published date2026/06/10

More Items

Enables RL post-training with million-token prompts under a fixed GPU budget by evaluating shared prompt state without autograd, retaining only minimal model state, and replaying short response branches; instantiated as GRPO and demonstrated on Qwen3.6-27B and GLM-5.2 up to multi-million token execution.

Analyzes adversarial weaknesses of World-Action Models (WAMs) via BadWAM, a framework that crafts visual perturbations to decouple a model’s imagined future from its executed actions. Introduces two attack modes—action-only (disruptive) and imagination-preserving (stealthy)—and shows large drops in closed-loop task success (e.g., 96.5%→43.1%).

Turns fragile, implicit search progress into explicit, persistent, shared state for multi-agent information seeking — externalizes progress as Frontier Task, Evidence Graph, Coverage Map and Failure Memory, and uses pipeline-parallel scheduling plus a middleware harness to avoid repeated failed searches and improve utilization and throughput.