English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

StraTA Explained: How 'Plan Before You Act' Lets a 7B Model Beat Closed-Source Giants

Forum topic · 小凯 · 2026-05-17

Summary

This article breaks down the paper StraTA (arXiv:2605.06642), which introduces a hierarchical reinforcement learning framework for LLM agents. Instead of purely reactive step-by-step action selection, StraTA splits long-horizon decision-making into a Strategy Generator that emits a compact natural-language plan and an Action Executor that follows it while conditioning on the fixed strategy at every step. Training uses a two-level GRPO: a strategy-level GRPO compares entire plans and a lower-level GRPO compares execution variants within each plan. Reward is computed from the best rollouts per strategy (not the mean), and strategy candidates are selected via Farthest Point Sampling in embedding space to avoid homogeneous plans, followed by a self-critique filter. A 7B open-source model trained with StraTA reaches 93.1% on ALFWorld, 84.2% on WebShop (vs. 5.3% baseline), and 63.5% on SciWorld, surpassing frontier closed-source models.

Key Points

  • Problem: Reactive LLM agents suffer from exploration collapse and credit-assignment failure on long-horizon tasks, because they never learn to plan globally before acting.
  • Core idea: Decouple the trajectory into two jointly trained modules — a *Strategy Generator* that produces a compact natural-language plan, and an *Action Executor* that conditions on the fixed plan plus the current observation at every step.
  • Training mechanism:
  • Two-level GRPO compares plans at the strategy level and execution variants at the action level, separating "wrong strategy" from "bad execution".
  • Per-strategy reward uses the best rollouts rather than the mean, giving a more reliable quality estimate.
  • Farthest Point Sampling over strategy embeddings enforces semantic diversity instead of near-duplicate plans.
  • A lightweight self-criticism step lets the same LLM reject obviously flawed strategies before rollout.
  • Results:
  • ALFWorld (text household tasks): 93.1% vs. prior best ~70%.
  • WebShop (web shopping): 84.2% vs. 5.3% baseline (~16× lift).
  • SciWorld: 63.5%, surpassing frontier closed-source models; SciWorld Lifespan subtask reaches 100%.
  • Practical takeaway for engineers: Even without RL compute, the same idea applies in prompt engineering — put a global planner on top of a workflow executor that keeps the plan in context at every step.
  • Limitations: Strategy quality depends on the base LLM's common-sense reasoning, self-judgment is imperfect, plan granularity is hard to tune automatically, and hierarchical GRPO is still rollout-expensive.
  • Takeaway: The bottleneck of LLM agents is training paradigm, not model size — a 7B model taught to "plan before it acts" can outperform much larger closed-source models on long-horizon tasks.

Source

Paper: *StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction* — Xue et al., arXiv:2605.06642. https://arxiv.org/abs/2605.06642 Code: https://github.com/xxyQwQ/StraTA

Tags

#strata#llm-agents#reinforcement-learning#grpo#long-horizon-planning#hierarchical-rl#open-source-llm#paper-breakdown

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620179