DeepPlanner: Scaling Planning Capability for Deep Research Agents via Advantage Shaping (arXiv 2510.12979)

Forum topic · 小凯 · 2026-07-05

Summary

This forum post introduces DeepPlanner, an October 2025 arXiv paper (arXiv:2510.12979) on scaling the planning capability of deep research agents via advantage shaping. Authored by Wei Fan, Wenlin Yao, Zheng Li, Feng Yao, Xin Liu, Liang Qiu and colleagues, the work targets a core limitation of LLM-based deep research agents: the ability to decide when to search, how many retrieval steps to take, and which tools to call. The post frames the problem within the evolution of agentic search, where planning itself becomes a learnable object rather than a fixed pipeline of retrieve-rank-generate. It outlines the standard methodological pipeline (problem formalization, model/system design, training, inference) covering dense and sparse query representations, retrievers, rerankers, planners, memory modules, tool interfaces, and learning strategies such as supervised fine-tuning, contrastive learning, distillation, and reinforcement learning with process rewards. Evaluation typically uses benchmarks like MS MARCO, BEIR, and Natural Questions with nDCG, Recall@k, and task success rate metrics. The post also discusses engineering concerns (latency budgets, PII handling, retrieval poisoning, cost control), open problems such as evaluation reliability and hallucination, and cross-references related surveys on deep research and LLM-based search agents.

DeepPlanner: Scaling Planning Capability for Deep Research Agents via Advantage Shaping (arXiv:2510.12979)

Overview

This post catalogs DeepPlanner: Scaling Planning Capability for Deep Research Agents via Advantage Shaping, an October 2025 arXiv paper (https://arxiv.org/abs/2510.12979). It is filed under the "Deep Research" section of a curated reading list on search, recommendation, and personalization research.

Authors: Wei Fan, Wenlin Yao, Zheng Li, Feng Yao, Xin Liu, Liang Qiu, et al. (9 authors total)
Source: arXiv:2510.12979
Category: Academic paper — Deep Research / Agentic Search

Background and Motivation

The post situates the paper in the agentic search landscape, where LLM-era systems must handle natural language interaction, multi-hop reasoning, and real-time knowledge access. Unlike traditional pipelines that split retrieval, ranking, and generation into fixed stages, agentic approaches make the *retrieval strategy itself* — whether to search, how many times, and which tools to call — a learnable decision. DeepPlanner addresses the planning capability of deep research agents in this context, using advantage shaping as its core technique (see the original abstract for precise details).

Core Contributions (as outlined in the post)

A unified perspective on the agentic search problem domain, organizing scattered related work into a comparable framework.
A clear decomposition of method components: representation learning, retrievers, rerankers, planners, generators, and feedback mechanisms.
Reproducible benchmarks, datasets, or classification schemes that lower the entry cost for follow-up research.
Discussion of interfaces with emerging paradigms: LLM tool calling, reinforcement learning, and multi-agent collaboration.
Open problems including evaluation reliability, latency and cost, hallucination and safety, and cross-lingual / multimodal extension.

Methodological Pipeline

The post describes a typical four-step pipeline:

1. Input & representation — encoding queries, documents, and user context as dense/sparse representations or structured prompts. 2. Core modules — retrievers, rerankers, planners, memory modules, and tool interfaces, chained or combined per task. 3. Learning strategies — supervised fine-tuning, contrastive learning, distillation, reinforcement learning (including process rewards), and self-bootstrapped data synthesis. 4. Inference strategies — single-turn retrieval, iterative retrieval, parallel sub-queries, early stopping, and budget control.

Evaluation

Typical evaluation setups for work in this area include:

Datasets: MS MARCO, BEIR, Natural Questions, domain-specific corpora.
Metrics: nDCG@10, MRR, Recall@k, Hit@k, task success rate, latency and token cost.
Baselines: BM25, dense retrieval, cross-encoder reranking, retrieval-free LLMs, commercial search APIs.

The post notes that specific numerical results should be verified against the original PDF, since the summary is based on abstract-level metadata.

Key Insights for Search / Rec / Personalization

1. Architecture: cascaded retrieve-rank-generate remains mainstream, but the agentic paradigm turns retrieval count and policy into learnable objects. 2. Data: high-quality instruction data and click/session logs remain critical; synthetic data must guard against leakage and distribution shift. 3. Evaluation: the gap between offline metrics and online satisfaction is widening; LLM-as-judge needs cross-validation with human evaluation. 4. Product: latency, cost, explainability, and safety are hard constraints for industrial deployment — not just academic benchmarks.

Limitations and Future Work

Possible limitations include experiment scale bound by GPU budgets, mismatch between benchmarks and real user distributions, English-centric data limiting cross-lingual generalization, and safety risks of agents operating on the open web. Future directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommendation settings.

Cross-References

A Comprehensive Survey of Deep Research: Systems, Methodologies, and Applications (arXiv:2506.12594)
A Survey of LLM-based Deep Search Agents (arXiv:2508.05668)
A Survey of Scientific Large Language Models (arXiv:2508.21148)
Towards Scientific Intelligence: A Survey of LLM-based Scientific Agents (arXiv:2503.24047)
AgentIR: Reasoning-Aware Retrieval for Deep Research Agents
Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning (arXiv:2502.04644)

Glossary

| Term | Meaning | |------|---------| | RAG | Retrieval-Augmented Generation | | nDCG | Normalized Discounted Cumulative Gain, a ranking quality metric | | Agentic Search | Modeling search as sequential decision-making and tool calling | | Gen-IR | Generative Information Retrieval |

> Note: The forum post is an annotated catalog entry; the paper's precise technical claims, the advantage shaping formulation, and quantitative results should be read from the original arXiv abstract and PDF.