Source-linked AI summary
AgentCPM-Report: Interleaving Drafting and Deepening for Open-Ended Deep Research
Yishan Li, Wentong Chen, Yukun Yan, Mingwei Li, Sen Mei, Xiaorong Wang, Kunpeng Liu, Xin Cong, Shuo Wang, Zhong Zhang, Yaxi Lu, Zhenghao Liu, Yankai Lin, Zhiyuan Liu, Maosong Sun
TL;DR
Existing deep-research systems depend on comprehensive upfront outlines and often online large models, creating reasoning, deployment, safety, and privacy challenges. AgentCPM-Report addresses this with WARP’s interleaved drafting and deepening plus multi-stage training for an 8B local agent. Across multiple benchmarks, it reports substantial quality gains, particularly in Insight, and matches or outperforms leading closed-source systems.
Problem
Plan-then-write research systems depend on comprehensive initial outlines and online large models, limiting support for private local data.
Method
AgentCPM-Report combines WARP’s dynamic outline revision with alternating Evidence-Based Drafting and Reasoning-Driven Deepening in an 8B-parameter local agent.
Results
AgentCPM-Report substantially improves overall report quality, particularly Insight, across DeepResearch Bench, DeepConsult, and DeepResearch Gym, matching or outperforming leading closed-source systems.
Takeaways & Limitations
WARP and multi-stage training can bring a small agent to the performance level of leading proprietary deep-research systems.
Takeaways & Limitations
The textual knowledge base limits coverage and timeliness and excludes images, videos, domain-specific corpora, and personalized data.
Abstract
from arXiv · showhide
Generating deep research reports requires large-scale information acquisition and the synthesis of insight-driven analysis, posing a significant challenge for current language models. Most existing approaches follow a plan-then-write paradigm, whose performance heavily depends on the quality of the initial outline. However, constructing a comprehensive outline itself demands strong reasoning ability, causing current deep research systems to rely almost exclusively on closed-source or online large models. This reliance raises practical barriers to deployment and introduces safety and privacy concerns for user-authored data. In this work, we present AgentCPM-Report, a lightweight yet high-performing local solution composed of a framework that mirrors the human writing process and an 8B-parameter deep research agent. Our framework uses a Writing As Reasoning Policy (WARP), which enables models to dynamically revise outlines during report generation. Under this policy, the agent alternates between Evidence-Based Drafting and Reasoning-Driven Deepening, jointly supporting information acquisition, knowledge refinement, and iterative outline evolution. To effectively equip small models with this capability, we introduce a Multi-Stage Agentic Training strategy, consisting of cold-start, atomic skill RL, and holistic pipeline RL. Experiments on DeepResearch Bench, DeepConsult, and DeepResearch Gym demonstrate that AgentCPM-Report outperforms leading closed-source systems, with substantial gains in Insight.
1 Introduction
AgentCPM-Report addresses the insight ceiling and deployment limitations of rigid plan-then-write systems with a local framework that interleaves drafting, deepening, and outline revision. Its 8B-parameter agent uses WARP and multi-stage training to improve deep-research report quality while supporting local operation.
- Motivation: Plan-then-write systems can produce structurally sound but intellectually shallow reports when their frozen outlines miss information discovered during research.Their reliance on a comprehensive initial outline assumes information completeness, which open-ended research often violates.
- Motivation: Generating comprehensive initial outlines demands reasoning and domain knowledge, encouraging reliance on closed-source or online large models and complicating private-data use.The paper identifies a need for fully local, on-device research that avoids external large-scale models.
- Contribution: AgentCPM-Report is a lightweight local system combining the WARP framework with an 8B-parameter deep research agent.WARP is presented as a response to both the insight ceiling and the challenges of on-device deep research.
- Contribution: WARP alternates Evidence-Based Drafting with Reasoning-Driven Deepening, allowing the agent to revise its outline from feedback arising during writing.The deepening state evaluates draft density and coherence, then decomposes sections and updates the outline when further refinement is warranted.
- Contribution: Multi-Stage Agentic Training combines trajectory pruning with curriculum-based reinforcement learning to train the dynamic policy under long-horizon credit-assignment and expanded-action-space challenges.The strategy progressively optimizes local atomic actions before end-to-end behavior and balances research depth against computational cost.
- Results: AgentCPM-Report substantially improves report quality, particularly Insight, across DeepResearch Bench, DeepConsult, and DeepResearch Gym while matching or outperforming leading closed-source systems.The reported results position the system as a foundation for safe, privacy-preserving, fully local report generation.
2 Method
The method formulates deep research as iterative hierarchical decision-making, unifying outline planning and content generation through WARP’s alternating drafting-and-deepening policy. Multi-stage training prepares an 8B agent by pruning supervision and progressing from atomic skills to holistic report optimization.
- Formulation: AgentCPM-Report represents deep research as an iterative hierarchical decision process with states containing the query, dynamic outline, draft, and retrieved context.Actions are selected from INITIALIZE, SEARCH, WRITE, EXPAND, and TERMINATE.
- Formulation: Outline adjustments and content generation are treated as equivalent policy-driven state transitions, thereby unifying planning and writing.The workflow is explicitly iterative rather than linear.
- WARP Inference: WARP initializes a sparse high-level outline, then alternates contextual retrieve-then-write drafting with global reasoning that diagnoses gaps and expands shallow sections.Draft context guides retrieval and supports coherent evidence integration, while deepening can create local subsections and trigger targeted drafting.
- WARP Inference: The process terminates after the agent verifies that the logical chain is complete and content depth matches the query’s complexity.Deepening is therefore governed by a quality-oriented stopping decision rather than a fixed linear schedule.
- Data Preparation: Trajectory pruning creates termination supervision by over-expanding teacher trajectories, selecting the highest-scoring draft, and relabeling that point as TERMINATE.This teaches stopping based on report quality rather than arbitrary teacher imitation.
- Multi-Stage Agentic Training: Training proceeds through SFT cold start, Atomic Skill RL for local abilities, and Holistic Pipeline RL for end-to-end report quality.Atomic RL decomposes planning, retrieval, writing, and termination, while Pipeline RL optimizes global metrics such as Comprehensiveness and Insight.
3 Experiments
Experiments evaluate AgentCPM-Report across three benchmarks and compare its WARP framework, training curriculum, action behavior, deepening depth, and trajectory pruning. The results show strong report quality, effective draft-aware deepening, improved behavior after reinforcement learning, and benefits from pruning.
- 3 Experiments: AgentCPM-Report uses MiniCPM4.1-8B and is evaluated on DeepResearch Bench, DeepConsult, and DeepResearch Gym against proprietary, prompt-based, and trained open-model systems.The benchmarks cover scientific, business and financial, and general information-seeking tasks.
- 3.2 Main Results: 52.64 Insight and 50.54 Comprehensiveness on DeepResearch Bench surpass Gemini-2.5-Pro-deepresearch’s 49.45 and 49.51, respectively.On DeepResearch Gym, AgentCPM-Report also achieves 100.0 in Depth, Breadth, and Insightfulness.
- 3.2 Main Results: 50.11 Overall on DeepResearch Bench surpasses Gemini-2.5-Pro-deepresearch’s 49.71, while the system reaches 98.48 average on DeepResearch Gym.The reported results characterize the small-scale system as competitive with leading proprietary research systems.
- 3.3.1 Does WARP remain effective without training?: WARP outperforms Plan-then-write across all metrics, including +1.19 Insight and +0.98 Comprehensiveness.WARP starts from a simple outline and interleaves writing with iterative deepening, whereas Plan-then-write uses a fixed detailed outline.
- 3.3.2 How multi-stage training shapes agent actions and report structure?: Expand actions nearly double from 4.44 to around 8.8 as training moves from SFT to reinforcement learning, while Level-3 sections increase from 4.86 to 17.32.RL training drives at least 4 Expand steps and produces richer hierarchical report structure.
- 3.3.3 How does the number of deepening influence report quality?: Performance plateaus around nine expansion steps, with Comprehensiveness and Insight improving by nearly 6 points from shallow to sufficiently deep regimes.SFT usually stops within 6 steps, whereas Atomic RL and Pipeline RL shift stopping toward 6–15 steps, closer to the empirically optimal depth.
- 3.3.4 How does the trajectory pruning affect the agent training?: Pruned teacher trajectories consistently outperform raw trajectories across all evaluation dimensions, improving supervision and termination-policy learning.Pruning selects the best intermediate draft using reward scores instead of cloning potentially suboptimal teacher stopping decisions.
4 Conclusion and Future Works
AgentCPM-Report concludes that WARP and multi-stage training enable a fully local deep research system with improved report insight, while identifying presentation and information coverage as remaining limitations and future directions.
- AgentCPM-Report reformulates deep research as iterative refinement through WARP and uses multi-stage agentic training to support long-horizon decisions.
- Experiments across multiple deep research benchmarks show substantial report-quality improvements, particularly in insight, despite using an 8B-parameter model.
- Better report presentation: Inline generation couples report presentation with prose writing, placing heavy structural and formatting demands on smaller models.
- Better report presentation: Current readability evaluation is largely text-based and weakly reflects the visual structure of rendered reports, motivating visual-modality evaluation.
- More information sources: The locally deployed textual knowledge base ensures stability and reproducibility but limits coverage, timeliness, modalities, and access to personalized data.
5 Contributions and Acknowledgments
The acknowledgments credit the collective team effort behind AgentCPM-Report and identify project leads, contributors, and advisors.
- AgentCPM-Report is credited to the collective efforts of the full research team.
- Yishan Li and Wentong Chen are listed as project leads.
- The contributor list includes Yishan Li, Wentong Chen, Yukun Yan, Mingwei Li, Sen Mei, Xiaorong Wang, Kunpeng Liu, Xin Cong, Shuo Wang, Zhong Zhang, Yaxi Lu, Zhenghao Liu, Yankai Lin, Zhiyuan Liu, and Maosong Sun.
- Yukun Yan, Yankai Lin, Zhiyuan Liu, and Maosong Sun are listed as advisors.
A Method Details
The method details describe WARP’s action space, prompting scheme, local retrieval environment, and training data construction for iterative report generation.
- WARP uses five actions across its stages: Initialize, Search, Write, Expand(Deepen), and Terminate.Initialization creates the initial Level-1 outline; drafting uses Search and Write, while expansion and termination prompts govern later decisions.
- The searcher selects up to five relevant keywords using the user query and current article outline, while matching the user’s language.
- The writer composes coherent, accurately cited paragraphs from analyst instructions, current writing status, and recently retrieved information.The prompt encourages breadth, depth, analytical comparison, and faithful support for claims through citations.
- The expansion policy selects one section for a one-level extension or terminates when expansion is unnecessary, redundant, or overly detailed.The prompt restricts expansion to levels 1 and 2 and prevents extending an already expanded section.
- The local retrieval environment contains approximately 2.86 million documents, including about 2.71 million arXiv abstracts and 150k summarized web pages.
A.4 Trajectory Data Construction
The trajectory-data construction process collected WARP executions with a large teacher model and pruned trajectories to improve training signals for decision-making.
- 1,500 actual WARP execution trajectories were collected using Qwen3-235B-A22B-Instruct-2507 as the teacher model.
- Trajectory pruning was introduced because large language models still struggle to determine reliably when to stop.The strategy addresses suboptimal termination decisions during trajectory collection.
A.5 Action Data Distribution
The action-data design maps agent actions to planning, retrieval, writing, and decision-making capabilities, using specialized rewards and evaluation criteria. Process-level RL evaluates complete reports rather than intermediate actions, enabling freer exploration.
- Action distribution: Agent trajectories contain initialize, terminate, expand, search, and write actions representing planning, retrieval, writing, and decision-making.The action distribution is imbalanced: critical planning and decision-making abilities are underrepresented relative to easier abilities.
- Reward design: The reward system defines ability-specific functions for four capabilities and optimizes actions including initialize, expand-plan, search, and writing.These rewards are primarily used for single-step RL training.
- Planning rewards: Planning rewards evaluate one-level outlines using basic properties, holistic quality, and faithfulness.The outline criteria include reasonable section counts, hierarchical clarity, logical coherence, content-generation guidance, and factual reliability.
- Retrieval and writing rewards: Retrieval rewards use recall by comparing retrieved documents with golden documents, while writing rewards assess paragraph properties, holistic quality, faithfulness, and citation precision.Writing constraints include paragraph lengths of 100–2000 tokens and 0–12 citations; citation hallucinations reduce citation precision to zero.
- Decision-making and content evaluation: Decision-making rewards evaluate terminate actions with accuracy, assigning 1.0 for matching the reference decision and 0.0 otherwise.Content evaluation also considers relevance, coverage, depth, and novelty.
- Process-level RL: Process-level RL evaluates the final report across several aspects without reference answers or intermediate-action scores, allowing freer and more diverse exploration.The stated goal is to improve the complete report beyond the teacher model’s capabilities.
B.1 Training Details
Training uses a three-stage curriculum that progresses from supervised cold-start training to atomic skill RL and end-to-end holistic pipeline RL.
- Training curriculum: The training pipeline consists of cold-start training, atomic skill RL, and holistic pipeline RL.Experiments use 8 A100 GPUs.
- Cold-start training: Cold-start training uses approximately 33k action-level samples from 1,200 trajectories for supervised fine-tuning over 4 epochs.Training uses a 1.5e-5 learning rate and batch size 32, taking about 2 days.
- Atomic skill RL: Atomic skill RL uses approximately 5150 action-level samples from 300 trajectories for single-step reinforcement learning.It uses a 2.5e-6 learning rate, batch size 8, eight rollouts, and 200 optimization steps over about 2 days.
- Holistic pipeline RL: Holistic pipeline RL performs end-to-end optimization on 500 user queries.It uses a 1e-6 learning rate, batch size 8, four rollouts, and 50 training steps over about 4 days.
B.2 Metrics Details
Metrics are reported on three deep-research benchmarks using complementary evaluation frameworks for report quality, citations, and pairwise performance.
- Evaluation benchmarks: Evaluations cover DeepResearch Bench, DeepConsult, and DeepResearch Gym.The benchmarks span academic research tasks, business and consulting queries, and research questions.
- DeepResearch Bench: DeepResearch Bench contains 100 PhD-level tasks across 22 academic domains and evaluates comprehensiveness, insight/depth, instruction following, readability, citation effectiveness, and citation reliability.It uses the RACE and FACT frameworks and is evaluated by Gemini-2.5-Pro.
- DeepConsult: DeepConsult contains 102 business and consulting queries and compares systems with win, tie, and loss rates plus average quality scores.The quality dimensions are instruction following, comprehensiveness, completeness, and writing quality; evaluation uses o3-mini-2025-01-31.
- DeepResearch Gym: DeepResearch Gym samples 100 queries from the top 1,000 test questions and evaluates clarity, depth, balance, breadth, support, and insightfulness.Evaluation uses GPT-4.1-mini-20250414.
- Rubric-based scoring: The metrics materials use rubric-based evaluation prompts containing an instruction, response, reference answer scored 5, and evaluation criteria.Evaluators identify major and minor errors before assigning an integer score from 1 to 5.