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

DeepPlanner: Scaling Planning Capability for Deep Research Agents via Advantage Shaping

Forum topic · 小凯 · 2026-07-05

Summary

DeepPlanner is an October 2025 arXiv paper (arXiv:2510.12979) by Wei Fan, Wenlin Yao, Zheng Li, Feng Yao, Xin Liu, Liang Qiu and colleagues, addressing the planning capability of LLM-based deep research agents. Deep research agents must interleave reasoning, search, and tool calls over long horizons, and the paper argues that planning is a key bottleneck. DeepPlanner trains the agent with reinforcement learning and introduces an advantage shaping scheme: instead of assigning a single trajectory-level advantage uniformly, it differentiates the advantages assigned to planning tokens versus acting tokens, allocating credit in a way that strengthens the learning signal for the planning (interleaved thinking) parts of each step. The motivation is that sparse, outcome-based rewards provide weak supervision for when and how to plan across multi-step search tasks. The work situates itself in the broader agentic search landscape, where retrieval, ranking, generation, and tool invocation must be jointly optimized under compute budgets. According to the authors, DeepPlanner improves planning quality and task success on deep research benchmarks relative to baselines. Readers should consult the original PDF for exact datasets, metrics, and quantitative results, as this abstract is based on the paper's public metadata.

DeepPlanner: Scaling Planning Capability for Deep Research Agents via Advantage Shaping

Overview

DeepPlanner (arXiv: 2510.12979, October 2025) is a research paper by Wei Fan, Wenlin Yao, Zheng Li, Feng Yao, Xin Liu, Liang Qiu, et al. (9 authors total). It targets the planning capability of deep research agents — LLM-based agents that perform long-horizon, multi-step information seeking by interleaving reasoning, search queries, and tool calls.

Key points

  • Problem: In deep research tasks, the agent must decide *when* to plan, *what* to search, and *how* to coordinate retrieval with reasoning. The authors identify planning as a central bottleneck; standard RL with sparse, outcome-level rewards provides weak supervision for the planning portions of a trajectory.
  • Method — Advantage Shaping: DeepPlanner trains agents with reinforcement learning and reshapes the advantage signal so that planning tokens and acting tokens receive differentiated credit. Rather than applying a uniform trajectory-level advantage, the scheme allocates advantage based on the value/novelty of the planning behavior at each step, amplifying the learning signal for effective interleaved thinking.
  • Positioning: The work sits at the intersection of agentic search, retrieval-augmented generation (RAG), and RL for LLMs, where retrieval depth, tool choice, and reasoning budget are themselves learnable decisions rather than fixed pipeline stages.
  • Results: The authors report improved planning quality and downstream task performance on deep research benchmarks compared with baselines. Exact datasets, metrics, and numbers should be verified against the original PDF.
  • Context for practitioners

  • Classic search stacks separate recall, ranking, and generation; agentic approaches fold "whether and how many times to retrieve" into a learnable sequential decision process.
  • Process-level credit assignment (as in DeepPlanner's advantage shaping) is an emerging pattern in agent RL, alongside process reward models and step-level supervision.
  • Engineering constraints highlighted by this line of work: inference/compute budgets, multi-step latency, citation reliability, and safe behavior on the open web.
  • Related entries

  • A Comprehensive Survey of Deep Research (arXiv:2506.12594)
  • A Survey of LLM-based Deep Search Agents (arXiv:2508.05668)
  • Agentic Reasoning (arXiv:2502.04646)
  • Reference

  • Original paper: DeepPlanner: Scaling Planning Capability for Deep Research Agents via Advantage Shaping — https://arxiv.org/abs/2510.12979
*Note: Quantitative claims above are based on the paper's abstract and public metadata; consult the full PDF before citing specific results.*

Tags

#deep-research-agents#reinforcement-learning#advantage-shaping#planning#llm-agents#agentic-search#rag#arxiv

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/178208602