Deep Cognition: A Multi-Agent Framework for Collaborative Research with Real-Time Cognitive Oversight

ICLR 2026 Conference Submission25395 Authors

Keywords: Interactive AI Systems, Human-in-the-Loop, Multi Agent Framework

Abstract:

Despite advances in large language models, current systems for deep research are limited by an asynchronous, "input-wait-output" interaction paradigm. This model creates a critical disconnect between human intent and AI execution, leading to error propagation and an inability to dynamically course-correct during complex problem-solving. We propose that a more effective form of human-AI partnership requires a shift from passive command-giving to cognitive oversight, where humans actively guide and intervene in the AI's thinking process. This perspective treats interaction as a core component of intelligence, rather than a peripheral interface.

We introduce Deep Cognition, a system designed to enable this paradigm through three technical pillars: transparent and interruptible AI reasoning, fine-grained bidirectional dialogue, and a shared cognitive context. At the core of our system is a layered StateManager architecture and a novel multi-stage budget allocation algorithm. This architecture ingests and normalizes all interaction data (e.g., dialogue trajectories and user artifacts) into a perpetually optimized, high-information-density working memory. By dynamically prioritizing context based on a combination of static heuristics and a time-sensitive scoring function, our system mitigates error cascades and allows the AI to adapt its reasoning pathways based on the user's implicit focus.

We conduct a comprehensive user study on challenging deep research tasks to evaluate the efficacy of our system. Results show that our approach significantly enhances the user experience, yielding improvements of up to 29.2% in Fine-Grained Interaction and 27.7% in Ease of Collaboration compared to a competitive baseline. Most notably, our system demonstrates a 31.8% to 50.0% points improvement in overall task performance. These results highlight the critical importance of designing interactive AI systems that facilitate continuous human guidance and transparent reasoning, rather than merely responding to isolated commands.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 25395