AI summaryⓘ
The authors study common failure behaviors of large language models (LLMs) when they react poorly under stress, like getting stuck or overreacting. They introduce the Gubernaut Cognitive Controller (GCC), a separate control system that monitors simple numeric signals (like intensity and repetition) from the model rather than the text itself, to help regulate the model's responses in real time. Testing GCC across several advanced LLMs showed it generally helped keep responses calmer and more stable. Their method relies on an architectural design that avoids direct manipulation of the text output, and results were validated by multiple independent judges. The authors clearly state they do not claim any form of consciousness in their system.
Large Language ModelsLLM AlignmentRuntime ControlNelson-Narens Monitoring-Control LoopArousal and ValenceMeta-level ControlGubernaut Cognitive ControllerModel-AgnosticBehavioral RegulationEvaluation Protocol
Abstract
Large language model (LLM) agents inherit reactive failure modes: escalation under provocation, sycophantic drift under flattery, perseveration when stuck. These are failures of propensity, not capability; they concern what a model does under sustained pressure, which training-time alignment reduces but does not eliminate at runtime. This research led to the Gubernaut Cognitive Controller (GCC), a model-agnostic runtime control layer in a Nelson--Narens monitoring--control loop: an object level reads and writes text, while a deterministic meta level reads only the numeric telemetry {intensity, valence, repetition} and returns a regulating posture. Because the meta level ingests zero tokens, no injection channel to the controller exists by construction (an architectural property, not yet adversarially tested); the text-exposed arbiter's compliance is measured, not assumed. We evaluate the GCC with a pre-registered, generate-once/judge-many protocol across a 4x4 matrix of four frontier models (GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash, Grok 4.3), each serving as both a generator and a judge. The regulated arm is calmer in 13 of 16 cells at p<.05 and 15 of 16 by sign; the three sub-threshold cells, including a -0.04 null, all fall on the single near-saturated host. The effect survives a lineage-independent fourth judge family (xAI), strong evidence that it is no artifact of shared judge style. The clearest mechanism is the recovery signature: arousal that integrates under attack and then decays, valence-gated, on de-escalation, replicating across all four families. Transcripts and panels ship with SHA-256 provenance and are re-judgeable; five failure modes are pre-registered. No consciousness claims are made.