CognitionLight: Continue, Rethink, or Rollback? Signaling for Persona-Aware Reasoning in Intelligent Agents

ICLR 2026 Conference Submission13785 Authors

Keywords: Symbolic Control Signals; Cognitive Control; Persona Switching; Intelligent Agent Control; Multi-Turn Reasoning

TL;DR: CognitionLight is a lightweight confidence-based controller that adaptively rolls back, switches strategies, or continues reasoning—offloading high-level planning from LLMs.

Abstract: In complex, dynamic scenarios, intelligent agents often proceed with overconfidence, repeating errors or switching strategies inconsistently — this leads to hallucinations, particularly in multi-turn or tool-augmented interactions. Can we equip intelligent agents with human-like cognitive control: to reason adaptively, choose suitable thinking styles, and self-correct in complex, multi-turn tasks? Inspired by human meta-reasoning, we introduce CognitionLight, a cognitively inspired control plugin that regulates agent behavior via a symbolic “traffic-light” mechanism. At each reasoning step, CognitionLight computes a multi-dimensional confidence vector and issues one of three symbolic control signals: Continue (green), Switch Persona (yellow), or Rollback (red), dynamically guiding how the agent proceeds. To operationalize the symbolic signals, CognitionLight incorporates a structured Persona Switching Module. Upon receiving a control signal, the system selects from five predefined cognitive styles: Direct, Reflective, Conservative, Tool-Seeking, and Contextual, each implemented via prompt-level behavioral modulation. The choice is guided by a fused representation of task-level uncertainty, feedback consistency, and historical persona performance, enabling adaptive reasoning modulation. Through extensive experiments on multi-turn reasoning benchmarks, we demonstrate that CognitionLight enhances response consistency, reduces hallucinations, and enables dynamic persona adaptation. Our results validate it as a promising framework for integrating human-like meta-reasoning into large-scale agent systems, offering both stability and flexibility in diverse reasoning environments.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 13785