When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoning

2026-09-29

AI Summary

This study addresses the issues of uncontrolled reasoning and resource exhaustion in large reasoning models caused by redundant verification and repetitive loops. We propose RADAR, a framework that formulates the generation process as a four-state machine, dynamically detects anomalous distributions through real-time attention response analysis, and enforces runtime intervention via attention realignment techniques. This work is the first to mechanistically reveal the degradation pathway from benign reasoning to harmful behavior, enabling early detection of precursors to reasoning anomalies. Experimental results demonstrate that RADAR significantly suppresses infinite loops while preserving performance on standard tasks, offering an actionable intervention strategy for ensuring the safety of large model reasoning.

Abstract

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.

Problem

• Large Reasoning Models
• Uncontrolled Reasoning
• Attention Dynamics
• Generation Loops
• Inference Cost

Innovation

• Reasoning-state Analysis
• Dynamic Attention Responses
• Attention Realignment
• Uncontrolled Reasoning
• Large Reasoning Models