More Works from Our Lab

MuSEAgent: A Multimodal Reasoning Agent with Stateful Experiences

A multimodal reasoning agent that abstracts interaction data into atomic decision experiences via hindsight reasoning, and adaptively exploits a quality-filtered experience bank with wide- and deep-search policies.

arXiv 2026
GlyphBanana: Advancing Precise Text Rendering Through Agentic Workflows

An agentic workflow that progressively refines text rendering for high-fidelity multilingual typography in image generation.

arXiv 2026

AgentDisCo: Towards Disentanglement and Collaboration
in Open-ended Deep Research Agents

Jiarui Jin, Zexuan Yan, Shijian Wang, Wenxiang Jiao, Yuan Lu
Xiaohongshu Inc.
Paper arXiv Code GALA Benchmark Demo Renderer Gallery
TL;DR

A disentangled & collaborative agentic architecture for open-ended deep research.

AgentDisCo decouples outline synthesis from search-query planning into a Critic Agent and a Generator Agent that adversarially co-evolve through structured blueprints over a shared Document Bank — instead of fusing information exploration and exploitation into a single, undifferentiated module.

  • Disentangled Architecture

    A critic / generator decomposition that turns deep research into adversarial optimization between information exploration and exploitation, mediated by structured blueprints.

  • Meta-Optimization Harness

    A code-generation-driven harness (Claude-Code / Codex) that automatically explores agent configurations and assembles a reusable policy bank of design strategies.

  • GALA Benchmark

    A new benchmark that mines latent deep research interests from real user browsing history, reflecting organic everyday information needs beyond academic queries.

  • Multimodal Renderer

    A lightweight rendering agent that converts research reports into HTML, posters, and slides, powering the AutoResearch Your Interest product demo.

System Architecture

Comparison of deep research paradigms and the AgentDisCo architecture.
AgentDisCo decouples outline synthesis from search-query planning into a Critic Agent and a Generator Agent that adversarially co-evolve through structured blueprints over a shared Document Bank, in contrast to conventional pipelines (a, b) that fuse the two roles into a single, undifferentiated module.

AgentDisCo is built around an outline optimization loop in which a cold, evaluative Critic Agent and a hot, generative Generator Agent communicate through structured blueprints. Each blueprint binds a planned narrative section to a list of targeted search queries, ensuring that retrieval at every iteration is systematically aligned with the intended scope of the final report.

  • Planner Agent classifies the user query into a response style (decision-making vs. information-seeking) and configures the downstream loop.
  • Critic Agent rates the current outline, identifies information gaps, and emits blueprints with section-level subqueries.
  • Generator Agent retrieves updated evidence and revises the outline draft along with extracted references.
  • Document Bank & Memory Bank sustain citation fidelity across turns and prevent citation drift.
  • Writer Agent turns the converged outline into a long-form report with intent-driven hierarchical generation.
  • Render Agent transforms the report into HTML, posters, slides and other multimodal deliverables.

GALA: General AI Life Assistants

Existing deep research benchmarks predominantly focus on academic or domain-specific consulting queries, which diverge from the breadth and diversity of real-world user needs. We introduce GALA — a benchmark constructed through an agentic workflow that automatically mines latent deep research interests from users' historical browsing behavior on the Xiaohongshu platform, enabling a more faithful reflection of organic, everyday information needs.

On top of GALA, we build “AutoResearch Your Interest”, an end-user demonstration that automatically curates personalized deep research recommendations from a user profile, runs them through AgentDisCo, and delivers the results as visually rich rednote-style posters.

GALA benchmark mining pipeline and AutoResearch Your Interest demo.
From a user's interaction history, the mining pipeline extracts GALA queries that reflect real information needs. AgentDisCo (Planner → Critic & Generator → Writer → Render) transforms each query into a structured outline, a long-form report, and a gallery of rednote-style poster pages.

Demonstrations

Each case below is an end-to-end recording of AgentDisCo: from the user query, through the critic–generator outline loop and long-form writing, to the final multimodal poster rendering produced by the Render Agent.

Case 1 · Policy Interpretation

“2026年北京《非机动车管理条例》新规详解：蓝牌轻便摩托车 F 证考核要求、号牌十年有效期届满后的续期手续，以及非法改装判定的具体技术标准？”

A multi-faceted policy-interpretation query about Beijing's 2026 non-motorized vehicle regulations, covering F-license testing, license-plate renewal after the 10-year validity, and technical criteria for illegal modifications.

Case 2 · Industry Report

“2025 小红书发展报告”

An open-ended industry-report query — 2025 Xiaohongshu (rednote) Development Report — that requires broad evidence aggregation, structured outlining, and visually rich poster rendering.

BibTeX

@misc{agentdisco,
        title={AgentDisCo: Towards Disentanglement and Collaboration in Open-ended Deep Research Agents}, 
        author={Jiarui Jin and Zexuan Yan and Shijian Wang and Wenxiang Jiao and Yuan Lu},
        year={2026},
        eprint={2605.11732},
        archivePrefix={arXiv},
        primaryClass={cs.IR},
        url={https://arxiv.org/abs/2605.11732}, 
  }