Cognitive Kernel-Pro: A Framework for Deep Research Agents and Agent Foundation Models Training

• Open-source agents often rely on paid proprietary tools or lack multimodal and general agentic abilities, limiting accessibility and reproducibility.The paper identifies these limitations as motivating a framework that maximizes LLM and VLM capabilities without external dependencies.Source: §1 · p. 2
• Cognitive Kernel-Pro uses a hierarchical main agent and specialized web and file agents, with Python code as the action space and a shared foundation model.The framework is designed to minimize proprietary-tool reliance and simplify task-specific training-data collection.Source: §1 · p. 2, Figure 3 · p. 3
• Its training recipe constructs verifiable query-answer pairs across web, file, code-generation, and reasoning tasks, using process hints and hint-based rejection sampling to improve collected data.The paper describes this structured recipe as supporting robust performance across tasks while maintaining open-source compatibility.Source: §1 · p. 2
• Retries and ensemble-based multi-run strategies address randomness in tasks such as web browsing and are presented as improving reliability and consistency.The framework also includes a standalone reflection module that assesses task completion and retries incomplete tasks.Source: §1 · p. 3, Figure 3 · p. 3

• The main agent decomposes tasks, delegates subtasks, and aggregates responses, while both it and the sub-agents inherit from a shared base class and execute intermediate actions as Python code.The main agent does not itself have specialized web-browsing or file-processing skills.Source: §2 · p. 3
• WebDancer and WebSailor support PDF fetching but lack general file-agent capabilities; the feature comparison excludes Google Search API and notes it can be switched to DuckDuckGo.Source: Table 1 · p. 4
• The web agent navigates live pages using browser observations and defaults to the text accessibility tree, with screenshot mode enabling image processing by a multimodal model.It can save a web file locally for the file agent to process.Source: §2 · p. 4
• The file agent handles PDFs, spreadsheets, and images by splitting files into pages and reading selected subsets as text or screenshots.Screenshot reading is described as essential for image-based tasks.Source: §2 · p. 4
• The framework's two-tier design decouples the main agent's task-solving procedure from sub-agent execution and can be extended with additional specialized agents.Source: §2 · p. 4

3.1 Overall Data Recipe

• The overall recipe divides deep-research-agent abilities into Web, File, and Reasoning, converting existing benchmarks or constructing new queries for each category.Source: §3.1 · p. 5

3.2 Multi-hop Web Search Data Construction

• Multi-hop web QA construction starts from topic-diverse seed URLs, then has an agent browse pages and compose questions requiring retrieval and information aggregation.Composition rules use operations suited to the sources, including calculation for numbers, sorting candidate sets, and analyzing tables.Source: §3.2 · p. 5

3.3 Agent Exploration-based Data Construction

• For agent-exploration data, the main agent constructs queries by combining information from multiple verifiable sources, while an LLM generates broad topics that are diversity-sampled.The synthesis procedure reuses the ordinary framework's sub-agents while adapting the main agent for data synthesis.Source: §3.3 · p. 6
• Intermediate results are added as hints only during training-data trajectory collection, then removed from model inputs and outputs before training.The authors report that these hints significantly improve the task-solving agent's success rate during data collection.Source: §3.3 · p. 6

3.4 Persona Hub-based Data Augmentation

• Persona Hub-based synthesis uses an LLM to generate deep-research queries from synthetic personas, seeded by a manually crafted question and persona example.The authors state that using more personas can scale query diversity despite the limited number of manually crafted questions.Source: §3.4 · p. 6
• Cross-validating trajectories from different agent systems supplies 1k synthetic Persona Hub queries and trajectories for training, while manual validation and response annotation remain excluded.The ablation is described as suggesting that a small number of these trajectories can improve performance.Source: §3.4 · p. 6

3.5 Reasoning Data Construction

• The reasoning data converts NumiaMath, LogiCoT, TACO, and RiddleSense into agent inputs and expected outputs, extracting code-task input/output pairs from TACO descriptions and cases.Source: §3.5 · p. 7

3.6 Trajectory Sampling

• GPT-4.1 generates trajectories for constructed query-answer pairs, followed by similarity-based rejection sampling and up to three attempts per query.Hint-based sampling excludes text bracketed by <secret> and </secret> to prevent information leakage.Source: §3.6 · p. 7

3.7 Experiment Setup

• The evaluation uses GAIA, while the OWL reproduction follows default settings but omits Whisper API and uses a different network environment.Those differences may have contributed to lower reproduced performance than OWL’s originally reported results.Source: §3.7 · p. 8
• Cognitive Kernel-Pro uses Claude-3.7 as its framework backbone and CK-Pro-8B, based on Qwen-3-8B, as its agent foundation model, with Google Search API as its only paid tool.Source: §3.7 · p. 8

3.8 Results

• 5% in Pass@1 and 7% in Pass@3: Cognitive Kernel-Pro with Claude-3.7 surpasses Smolagents on the full GAIA dev set under identical conditions and rivals OWL.Source: §3.8 · pp. 8–9
• 38.18% Pass@3: the fine-tuned Qwen-3-8B model, supported by GPT-4.1 for multimodal functions, remains 30% below the state-of-the-art Claude-3.7 model.Source: §3.8 · p. 9
• CK-Pro-8B achieves the best Pass@1 and Pass@3 performance across all levels of GAIA’s text-only subset.Source: §3.8 · p. 9
• The reflection ablation finds Qwen-3-32B performs comparably to GPT-4.1, while CK-Pro-8B without reflection fine-tuning shows only marginal improvement.The authors identify training agent foundation models for reflection as a future direction.Source: §3.8 · p. 9
• Replacing Qwen-2.5-VL-72B with GPT-4.1 as the multimodal-language-model backbone yields only marginal performance improvements.The authors propose developing a fully multimodal backbone for future work.Source: §3.8 · p. 9

4.1 Data Construction and Filtering

• Candidate samples are checked by deriving answers from reference reasoning and verifying sources and reasoning; inconsistent samples are discarded.The data format is <Question, Answer, Reference Solution, Grounded URLs>.Source: §4.1 · p. 10
• Four CK-Pro SFT-checkpoint rollouts identify 400 high-quality partially correct samples and yield a curriculum signal based on verifiability and logical depth.Training progresses from high-consistency samples toward more complex, lower-consistency tasks.Source: §4.1 · p. 10
• Structured Subgoal Recovery reconstructs long-form reasoning from synthesis metadata into a dictionary of discrete, verifiable milestones.The milestones are intended to support more granular credit assignment for sparse, delayed rewards.Source: §4.1 · p. 10

4.2 Subgoal-Conditioned GRPO

• Each main-agent step is represented as a tuple containing the state, action, sub-agent trajectory, and active subgoal index.The action includes reasoning, planning, and executable Python code; the state includes the observation and accumulated context.Source: §4.2 · pp. 10–11
• Subgoal-conditioned GRPO separately normalizes local subgoal-level and global task-level feedback, then combines them into a unified step-wise advantage.Source: §4.2 · p. 11

1. Trajectory sampling.

• Subgoal-conditioned GRPO separately normalizes subgoal-level and task-level rewards, then combines their advantages with λ controlling the trade-off.The subgoal stream uses trajectories reaching the same phase; the final task stream uses statistics across all sampled trajectories.Source: §1 · p. 11

4.3 Inference-time Scaling

• Reflection retries rubric-flagged trajectories with targeted feedback, while voting compares multiple attempts and retains the highest-scoring verifier candidate.The verifier assesses sources, reasoning, task understanding and decomposition, action execution, and trajectory efficiency.Source: §4.3 · p. 12

4.4 Experiments

• 54.61% average Pass@3 on GAIA-text with Subgoal-GRPO plus verifier, and 58.00% on xbench-DeepSearch with RL + Pass@3 + Verifier.On GAIA-text Level 3 tasks, Subgoal-GRPO Pass@3 + Verifier reaches 41.67%.Source: Table 7 · p. 13, §4.4 · p. 13

5.1 Deep Research Agents

• Open-source deep research frameworks have advanced, but the section says they lag proprietary systems and often rely on proprietary tools, while open-source agent foundation models remain underexplored.The cited frameworks include WebDancer, WebSailor, OWL, SmolAgents, TapeAgent, AutoAgent, and OAgent.Source: §5.1 · p. 14

5.2 Process Credit Assignment and Multi-Reward Optimization in Agentic RL

• Prior RL work either induces step-level supervision from sparse outcomes or combines explicit intermediate rewards; Cognitive Kernel-Pro assumes hierarchical subgoal rewards.The section contrasts sparse-outcome approaches such as TreeRPO and ARPO with multi-reward optimization methods including GDPO.Source: §5.2 · pp. 14–15

6 Conclusion

• The 8B-based model surpasses WebDancer and WebSailor, and the paper reports further performance improvement through reinforcement learning with subgoal-conditioned GRPO.Future work will focus on more capable multimodal agent foundation models for increasingly complex tasks.Source: §6 · p. 15

Code-based Action and Tool-using

• Agents express actions, including tool and sub-agent calls, as Python functions whose executed outputs are fed into later steps.The described format asks for a one-line thought and a Python code block for the next action.Source: A Technical Details of Cognitive… · p. 19

State-enhanced Problem-solving Workflow

• Before each action, the workflow plans from prior steps and observations, while a progress state records completed work, future tasks, experience, and information.The information record can support counting tasks by avoiding redundancy.Source: A Technical Details of Cognitive… · p. 19

Unified Multi-module Communication

• Sub-agents accept a task string and return an output string plus a log, with optional arguments such as file paths.Their docstring-style definitions describe utilities and use cases to the main agent.Source: A Technical Details of Cognitive… · pp. 19–20

Technical Details

• The web agent can download files but cannot process or analyze them, so file analysis returns to an external planner or file agent.Its output dictionary contains a formatted answer and a log.Source: A Technical Details of Cognitive… · p. 20

DATA SYNTHESIZING REQUIREMENTS

• Queries should use verifiable sources, require cross-source synthesis, and produce novel, stable, unambiguous answers in a self-contained, clear, feasible format.They should focus on the information need rather than prescribing steps, and may require multiple abilities such as browsing, file handling, and multimodal processing.Source: B Details of Agent-Based Data… · pp. 20–21
• Example query topics range from natural-language-processing projects and jazz history to literature, space exploration, the Olympics, and world languages.Source: B Details of Agent-Based Data… · p. 21

HINT-BASED QUERY AUGMENTATION

• The hint-based augmentation instructions tell agents not to disclose or imply access to confidential hints, to continue using tools, and to prioritize tool results over conflicting hints.Source: B Details of Agent-Based Data… · p. 21