Mitigating Cognitive Inertia in Large Reasoning Models via Latent Spike Steering

TL;DR

This work tackles cognitive inertia in Large Reasoning Models by introducing STARS, a training-free framework that monitors latent hidden-state dynamics to steer reasoning in real time. It identifies Cognitive Pivots via latent $L_2$ spikes, differentiates Functional from Critical pivots, and applies a three-stage loop—Detect, Select, Steer—using robust spike extraction, trajectory-dynamics diagnosis, and lightweight, state-aware suffixes. Empirical results across math, logic, and domain-general benchmarks show that STARS improves accuracy while reducing unnecessary token generation and maintains efficiency, without requiring model fine-tuning. The approach highlights the value of internal latent signals for proactive reasoning control, offering a scalable, unsupervised mechanism to optimize long-horizon reasoning in LRMs.

Abstract

While Large Reasoning Models (LRMs) have achieved remarkable performance by scaling test-time compute, they frequently suffer from Cognitive Inertia, a failure pattern manifesting as either overthinking (inertia of motion) or reasoning rigidity (inertia of direction). Existing detection methods, typically relying on superficial textual heuristics like self-correction tokens, often fail to capture the model's unvoiced internal conflicts. To address this, we propose STARS (Spike-Triggered Adaptive Reasoning Steering), a training-free framework designed to rectify cognitive inertia by monitoring latent dynamics. STARS identifies Cognitive Pivots-critical moments of reasoning transition-by detecting distinct L2 distance spikes in the hidden states. Upon detection, the framework employs geometric trajectory analysis to diagnose the structural nature of the transition and injects state-aware language cues to steer the model in real-time. Our experiments across diverse benchmarks confirm that STARS efficiently curtails redundant loops while improving accuracy through the adaptive correction of erroneous trajectories. STARS offers a robust, unsupervised mechanism to optimize the reasoning process of LRMs without requiring additional fine-tuning.

Figures (15)

• Figure 1: Cognitive Inertia in LLMs. An illustration of two failure modes—overthinking (inertia of motion) and reasoning rigidity (inertia of direction).
• Figure 2: Visualization of latent $L_2$ spike patterns across different models. We plot high-magnitude latent displacement events (spikes), measured by the layer-wise $L_2$ distance between consecutive hidden states, for two distinct models processing the same problem instance. Despite exhibiting similarly prominent spike signals, the two models differ substantially in the distribution and relative prevalence of Functional and Critical pivots, reflecting heterogeneity in their latent reasoning dynamics across architectures.
• Figure 3: Spike-aligned context windows. Red-highlighted tokens indicate detected latent spikes. Each window shows a short preceding context for interpretability.
• Figure 4: Overview of Spike-Triggered Adaptive Reasoning Steering (STARS).
• Figure 5: Average spike and flip counts across model scales. Semi-transparent bars indicate the average number of latent spikes detected per question, whereas solid bars denote directional flips, a structurally meaningful subset of spike events.