The cognitive companion: a lightweight parallel monitoring architecture for detecting and recovering from reasoning degradation in LLM agents

arXiv cs.AI / 4/16/2026

📰 NewsSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper reports that LLM agents performing multi-step tasks can experience reasoning degradation (e.g., looping, drift, stuck states) at rates up to 30% on hard tasks, motivating improved monitoring and recovery approaches.
It introduces “Cognitive Companion,” a lightweight parallel monitoring architecture with two variants: an LLM-based companion (with ~11% overhead) and a zero-overhead probe-based companion trained on hidden states.
In feasibility experiments centered on Gemma 4 E4B, the LLM-based companion reduced repetition on loop-prone tasks by 52–62% while adding about 11% per-step overhead.
The probe-based companion achieved positive effects with zero measured inference overhead, reaching up to cross-validated AUROC 0.840 on a small proxy-labeled dataset.
The authors find strong task-type sensitivity (largest gains on loop-prone/open-ended tasks, neutral or negative on structured tasks) and suggest a potential scale boundary where small-model companions did not improve the measured quality proxy for 1B–1.5B models; the study is explicitly framed as feasibility rather than definitive validation.

Abstract

Large language model (LLM) agents on multi-step tasks suffer reasoning degradation, looping, drift, stuck states, at rates up to 30% on hard tasks. Current solutions include hard step limits (abrupt) or LLM-as-judge monitoring (10-15% overhead per step). This paper introduces the Cognitive Companion, a parallel monitoring architecture with two implementations: an LLM-based Companion and a novel zero-overhead Probe-based Companion. We report a three-batch feasibility study centered on Gemma 4 E4B, with an additional exploratory small-model analysis on Qwen 2.5 1.5B and Llama 3.2 1B. In our experiments, the LLM-based Companion reduced repetition on loop-prone tasks by 52-62% with approximately 11% overhead. The Probe-based Companion, trained on hidden states from layer 28, showed a mean effect size of +0.471 at zero measured inference overhead; its strongest probe result achieved cross-validated AUROC 0.840 on a small proxy-labeled dataset. A key empirical finding is that companion benefit appears task-type dependent: companions are most helpful on loop-prone and open-ended tasks, while effects are neutral or negative on more structured tasks. Our small-model experiments also suggest a possible scale boundary: companions did not improve the measured quality proxy on 1B-1.5B models, even when interventions fired. Overall, the paper should be read as a feasibility study rather than a definitive validation. The results provide encouraging evidence that sub-token monitoring may be useful, identify task-type sensitivity as a practical design constraint, and motivate selective companion activation as a promising direction for future work.