Metacognitive Control

Deep Search with Hierarchical Meta-Cognitive Monitoring Inspired by Cognitive Neuroscience

DS-MCM brings meta-cognitive monitoring into long-horizon search agents. It adds fast monitoring, slower reflective diagnosis, and experience updating so that agents can detect abnormal states, attribute failures, correct themselves, and reuse lessons across tasks.

SIGIR 2026 First author Zhongxiang Sun, Qipeng Wang, Weijie Yu, Jingxuan Yang, Haolang Lu, Jun Xu

Paper Code Research Program

Research Question

A deep-search framework that calibrates reasoning uncertainty against evidence uncertainty and triggers experience-driven correction.

How can a deep-search agent monitor its own execution, diagnose errors, and revise its behavior during long-horizon tasks?

Paper Walkthrough

From the research problem to mechanism and evidence

The figures below follow their order in the paper. Each section connects the figure to the question it resolves and the conclusion it supports.

DS-MCM Why deep-search agents need metacognitive monitoring figure

The top row maps human monitoring onto execution: the fast monitor checker, associated with ACC, triggers low-cost anomaly detection, while the slow monitor checker, associated with PFC and hippocampus, performs explicit memory-augmented reflection. Execution proceeds from task start through steps to completion.

The middle row shows the failure mode in a deep-search agent: query, tool use, reasoning, repeated execution, and then an error. The bottom row adds MCM. In the visible example, the query about a missing 16-year-old is processed through tool and reasoning steps; the system identifies a failure to understand the query, triggers rethinking, and produces a corrected final state.

How to read this figure

Top: use ACC versus PFC+hippocampus labels to distinguish fast anomaly detection from slow reflection.
Middle: the ordinary deep-search loop has tool and reasoning modules but no explicit correction point.
Bottom: MCM inserts query-understanding diagnosis and rethinking before the final answer.

The top execution pipeline starts from user query q. Step s_t interleaves retrieval action, document set D_t, semantic clusters C_t, reasoning trace, and token distribution p_r. These produce searching entropy SE_t and reasoning entropy RE_t, which enter the Fast Consistency Monitor.

The yellow monitor box calibrates expected reasoning entropy from searching entropy, computes epsilon_t, and triggers monitoring when |epsilon_t| exceeds threshold tau. The lower blue region builds memory from historical trajectories, retrieves success and failure experiences, uses a critical model C to produce corrective suggestion delta_t or error flag err_t, and then updates memory through deduplication and consolidation.

How to read this figure

SE_t comes from the retrieval/search side; RE_t comes from the reasoning/token-distribution side.
The fast monitor decides whether the current step is divergent or consistent by checking calibrated entropy mismatch.
The slow monitor turns retrieved success/failure experiences into corrective suggestions and memory updates.

DS-MCM Checking robustness of the monitoring mechanism (a) Entropy threshold k panel

DS-MCM Checking robustness of the monitoring mechanism (b) Retrieved monitoring memories |R| panel

Panel (a) varies the entropy threshold k. Panel (b) varies the number of retrieved memories |R|. Panel (c) varies the number of retrieved documents top-K. In all three panels, the blue Tongyi-Deepresearch curve stays clearly above the purple dashed Mirothinker-Deepresearch curve.

The curves also show that the best settings are not always at the largest value: the blue curve peaks around moderate k and |R| values, and panel (c) peaks around a mid-range top-K setting. This is why the result belongs in a monitoring paper: thresholds and retrieval budgets are part of the control mechanism.

How to read this figure

Panel (a) checks the trigger sensitivity of the fast monitor through entropy threshold k.
Panel (b) checks how much retrieved monitoring memory is useful.
Panel (c) checks the evidence retrieval budget through top-K documents.

Experimental Evidence

Main experimental results

DS-MCM is evaluated on four deep-search benchmarks using three open-source backbones. The table compares each unmodified system, an LLM-as-Critic baseline, and DS-MCM; GPT-5 is included only as a proprietary reference point reported in the paper.

Deep-search benchmark accuracy (%)

System	BrowseComp+	BrowseComp-ZH	X-Bench	GAIA	Average
GPT-5 (reference)	70.0	34.3	75.0	62.8	60.5
Tongyi-DR	51.0	38.0	69.0	70.1	57.0
Tongyi-DR + Critic	55.0	42.0	68.0	69.3	58.6
Tongyi-DR + DS-MCM	62.0*	45.0*	74.0*	71.7	63.2*
MiroThinker-DR	21.0	31.0	62.0	63.8	44.5
MiroThinker-DR + Critic	24.0	30.0	64.0	63.0	45.2
MiroThinker-DR + DS-MCM	26.0	34.0*	68.0*	69.3*	49.3*
Qwen3-30B-MoE	5.0	21.0	42.0	40.9	27.2
Qwen3-30B-MoE + Critic	22.0	27.0	47.0	43.3	34.8
Qwen3-30B-MoE + DS-MCM	29.0*	35.0*	53.0*	47.2*	41.1*

An asterisk marks DS-MCM improvements reported as statistically significant in the paper's main table.

What the results show

DS-MCM improves the average over the unmodified backbone by 6.2 points for Tongyi-DR, 4.8 for MiroThinker-DR, and 13.9 for Qwen3-30B-MoE.
It also exceeds the generic LLM-as-Critic baseline on every backbone, showing that structured fast and slow monitoring contributes beyond additional critique calls.
Tongyi-DR + DS-MCM reaches a 63.2 average, 2.7 points above the GPT-5 reference reported in the table.

Main Takeaways

What this project establishes

Fast monitoring calibrates reasoning against evidence

Searching Entropy measures semantic fragmentation in retrieved documents, while Reasoning Entropy measures uncertainty in the model's reasoning tokens. Successful trajectories define their expected relationship.

Slow monitoring diagnoses with success and failure memory

The slow monitor retrieves similar sessions from separate success and failure memory pools. These examples give the critical model concrete behavioral context for diagnosing the flagged step.

Selective triggering connects detection to control

The fast monitor runs at every step, while the slower critical process is activated only when uncertainty calibration becomes abnormal. This separates lightweight detection from more expensive reflection.

The framework improves deep-search robustness

Across multiple deep-search benchmarks and backbone models, DS-MCM consistently improves performance and robustness. The sensitivity analyses test the entropy threshold, retrieved-memory count, and document budget.

Research Impact

Position in the broader research program

DS-MCM is the metacognitive-control endpoint of the current project sequence. ReDeEP studies evidence use, RHD studies reasoning dynamics, ReARTeR studies process rewards, and DS-MCM integrates these ideas into a monitoring and correction loop.

The long-term impact is a framework for agents that can inspect their own execution state. As AI agents become more autonomous in research, legal, scientific, and decision-support tasks, this kind of monitoring layer becomes central to reliability.

Resources and Citation

Paper Code Full Publication List

@inproceedings{sun2026deep,
  title={Deep Search with Hierarchical Meta-Cognitive Monitoring Inspired by Cognitive Neuroscience},
  author={Sun, Zhongxiang and Wang, Qipeng and Yu, Weijie and Yang, Jingxuan and Lu, Haolang and Xu, Jun},
  booktitle={Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  doi={10.1145/3805712.3809549},
  year={2026}
}