Technical Summary: AgentCPM-Report

Problem Statement

Open-ended deep research requires agents to navigate vast information landscapes and synthesize findings into coherent, insightful reports. Current language models face significant challenges in this domain, primarily due to the limitations of existing paradigms:

1. Retrieval-then-Write: Early approaches generate content sequentially based on retrieved evidence. While flexible, this often leads to incoherence over long horizons.
2. Plan-then-Write: More recent frameworks (e.g., WebWeaver) freeze a comprehensive outline before writing to ensure structural consistency. However, this paradigm assumes initial information completeness, which is often violated in open-ended research. It rigidly separates planning from writing, preventing agents from capturing emergent insights that only surface during the drafting process. This creates an "insight ceiling" where reports are structurally sound but intellectually shallow.
3. Dependency on Large Models: Constructing high-quality initial outlines demands strong reasoning and domain knowledge, forcing most systems to rely on closed-source or online large models. This reliance introduces practical barriers to deployment, particularly regarding security and privacy when handling user-authored local data.

There is a critical need for a fully local, on-device solution that does not depend on external large-scale models while overcoming the insight ceiling of static planning.

Methodology: AgentCPM-Report

The authors propose AgentCPM-Report, a lightweight (8B-parameter) deep research agent built on a novel framework called WARP (Writing As Reasoning Policy).

1. The WARP Framework

WARP reformulates deep research as an iterative, policy-level refinement process where planning decisions dynamically emerge from the writing itself. Instead of a static blueprint, the agent alternates between two macro-states:

• Evidence-Based Drafting: The agent converts structural plans into substantiated content. It uses a retrieve-then-write strategy where retrieval queries are conditioned on the accumulating narrative to ensure contextual consistency. Content is grounded in retrieved evidence to guarantee faithfulness.
• Reasoning-Driven Deepening: The agent treats the newly generated draft as a fresh observation for reasoning. It evaluates the semantic density and logical coherence of the current draft to detect gaps or superficial arguments. If depth is lacking, the agent autonomously decomposes high-level sections into granular inquiries, updates the dynamic outline, and triggers a targeted drafting cycle.

This loop allows the agent to discover and bridge logical gaps that emerge only during the articulation of concrete arguments, mirroring the human "knowledge-transforming" process.

2. Multi-Stage Agentic Training

Training small models (8B) for such long-horizon decision-making is non-trivial due to ambiguous termination points and sparse rewards. The authors introduce a three-stage curriculum:

• Cold-Start (SFT): Establishes basic instruction following and format adherence using trajectory data.
• Atomic Skill RL: Decomposes the global objective into atomic abilities (Planning, Retrieval, Writing, Decision-Making). The agent is trained with specific reward functions for each skill (e.g., relevance recall for retrieval, faithfulness for writing) to master local execution before optimizing global strategy.
• Holistic Pipeline RL: Shifts to end-to-end optimization of the final report quality (Comprehensiveness, Insight). This stage uses a trajectory pruning strategy where teacher models are forced to "over-expand" recursively. The system then retroactively identifies the optimal stopping point based on report scores, pruning the trajectory to provide a supervision signal for information saturation. This teaches the agent to terminate based on quality rather than arbitrary imitation.

Key Contributions

1. WARP Framework: A policy-level reformulation that interleaves drafting and deepening, enabling dynamic outline evolution and breaking the insight ceiling of static planning.
2. Multi-Stage Agentic Training: A curriculum strategy combining cold-start, atomic skill RL, and holistic pipeline RL, specifically designed to handle long-horizon credit assignment and sparse rewards in small models.
3. Trajectory Pruning: A data preparation technique that solves the "optimal stopping problem" by filtering teacher trajectories to identify the highest-quality intermediate drafts, providing cleaner supervision signals for termination decisions.
4. Local, Privacy-Preserving Solution: A fully local system using an 8B parameter model (MiniCPM4.1-8B) that eliminates the need for closed-source APIs, addressing security and privacy concerns.

Experimental Results

The system was evaluated on three benchmarks: DeepResearch Bench (scientific tasks), DeepConsult (business/financial queries), and DeepResearch Gym (general information seeking).

• Performance vs. Closed-Source Systems: AgentCPM-Report (Pipeline RL) outperformed leading closed-source systems, including Gemini-2.5-Pro-deepresearch, on DeepResearch Bench. Specifically, it achieved an Insight score of 52.64 (vs. 49.45 for Gemini) and a Comprehensiveness score of 50.54 (vs. 49.51). On DeepResearch Gym, it achieved state-of-the-art results with an average score of 98.48.
• Impact of Training Stages: Performance steadily improved from SFT to Atomic RL and finally to Pipeline RL. For instance, on DeepResearch Bench, the Comprehensiveness metric rose from 46.24 (SFT) to 50.54 (Pipeline RL), and Insight from 48.10 to 52.64.
• Ablation Studies:
  • WARP Effectiveness: Even without training, applying the WARP paradigm to a larger model (Qwen3-235B) outperformed the Plan-then-Write paradigm, confirming that draft-aware deepening is a key driver of insight.
  • Deepening Depth: Experiments showed that report quality (Insight and Comprehensiveness) increases with deepening steps, plateauing around 9 steps. The trained Pipeline RL agent learned to stop near this optimal depth, whereas SFT agents stopped prematurely.
  • Trajectory Pruning: Models trained on pruned trajectories significantly outperformed those trained on raw teacher trajectories, validating the efficacy of the pruning strategy in teaching optimal termination.

Significance and Claims

The paper claims that AgentCPM-Report establishes a strong foundation for safe, privacy-preserving, and fully local deep research report generation.

The authors argue that for deep research tasks, the primary bottleneck is not model size, but the design of effective cognitive and planning processes. By integrating the WARP inference paradigm with multi-stage agentic training, small-size, open-source models can achieve deep research capabilities previously associated with proprietary large-scale systems. The work demonstrates that a specialized policy design can serve as a viable alternative to model scaling for advancing deep research capabilities.

The paper acknowledges limitations, noting that current systems rely on textual knowledge bases (lacking multi-modal or personalized data) and that table/figure generation remains a challenge for smaller models due to the structural reasoning required. Future work is suggested to decouple presentation from content generation and expand knowledge sources to multi-modal and local personalized data.