Reflection Steering: Disentangling Reflection from Reasoning in Activation Space for Token-Efficient Inference

• URL: http://arxiv.org/abs/2608.25542v1
• Date: Wed, 26 Aug 2026 08:55:28 GMT
• Title: Reflection Steering: Disentangling Reflection from Reasoning in Activation Space for Token-Efficient Inference
• Abstract summary: Reflection merely re-checks established results, it wastes reasoning tokens and increases latency.<n>Most existing reflection steering methods add a label-derived mean-difference direction across preset layers.<n>We propose Reflection Steering, a training-free framework for controlling reflection-associated within LLMs.
• Score: 14.39927122740842
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large reasoning models often produce reasoning traces with verification, revision, and backtracking. When reflection merely re-checks established results, it wastes reasoning tokens and increases latency. Most existing reflection steering methods add a label-derived mean-difference direction across preset layers, but its entanglement with reasoning and length signals destabilizes the accuracy-efficiency trade-off. In this paper, we propose Reflection Steering, a training-free framework for controlling reflection-associated computation within LLMs by disentangling reflection-related activations from general reasoning. Specifically, we contrast reflective and non-reflective hidden states at each LLM layer, denoise the resulting reflection directions with PCA, and orthogonalize them against general-reasoning directions. To limit downstream amplification from early-layer interventions, we calibrate each layer across multiple intervention strengths on a small set, retain only stable layers, and apply bounded projection removal to their residual-stream activations. We conduct extensive experiments across two public benchmarks and three open-weight LLMs against state-of-the-art activation-steering baselines. Results show that Reflection Steering reduces reasoning tokens by 16.9% on average across six matched settings. Besides, our method further introduces a bounded reflection intervention-strength parameter $α$, enabling deployment-time adjustment to balance token savings, accuracy, and generation stability.

Related papers

• Graph-Based Chain-of-Thought Pruning for Reducing Redundant Reflections in Reasoning LLMs [4.494151943799445]
  In this work, we argue that a major source of such redundancy is inefficient reflection, which often manifests in two problematic patterns.<n>Specifically, we convert each linear CoT into a directed acyclic graph (DAG) with explicit dependency edges.<n>Experiments show that our approach reduces the average reasoning tokens by 42% while maintaining or improving accuracy.
  arXiv  Detail & Related papers  (2026-04-07T09:46:03Z)
• Learning from Trials and Errors: Reflective Test-Time Planning for Embodied LLMs [63.88783817420284]
  Embodied robots cannot reflect on what went wrong or why, turning deployment into a sequence of independent trials.<n>We introduce Reflective Test-Time Planning, which integrates two modes of reflection: textitreflection-in-action and textitreflection-on-action<n>We also include retrospective reflection, allowing the agent to re-evaluate earlier decisions and perform model updates with hindsight.
  arXiv  Detail & Related papers  (2026-02-24T18:55:18Z)
• MirrorLA: Reflecting Feature Map for Vision Linear Attention [49.41670925034762]
  Linear attention significantly reduces the computational complexity of Transformers from quadratic to linear, yet it consistently lags behind softmax-based attention in performance.<n>We propose MirrorLA, a geometric framework that substitutes passive truncation with active reorientation.<n>MirrorLA achieves state-of-the-art performance across standard benchmarks, demonstrating that strictly linear efficiency can be achieved without compromising representational fidelity.
  arXiv  Detail & Related papers  (2026-02-04T09:14:09Z)
• From Latent Signals to Reflection Behavior: Tracing Meta-Cognitive Activation Trajectory in R1-Style LLMs [48.33546389897804]
  R1-style LLMs have attracted growing attention for their capacity for self-reflection, yet the internal mechanisms underlying such behavior remain unclear.<n>Using the logit lens to read out token-level semantics, we uncover a structured progression.<n>Our findings suggest a human-like metacognitive process-progressing from latent monitoring, to discourse-level regulation, and to finally overt self-reflection.
  arXiv  Detail & Related papers  (2026-02-02T11:58:24Z)
• ReflCtrl: Controlling LLM Reflection via Representation Engineering [6.828302913581854]
  We study self-reflection through the lens of representation engineering.<n>We propose a stepwise steering method that can control reflection frequency.<n>In experiments, we can save up to 33.6 percent of reasoning tokens while preserving performance.
  arXiv  Detail & Related papers  (2025-12-16T00:38:34Z)
• Meaningless Tokens, Meaningful Gains: How Activation Shifts Enhance LLM Reasoning [53.35553353785948]
  Motivated by the puzzling observation that inserting long sequences of meaningless tokens before the query prompt can consistently enhance reasoning LLM performance, this work analyzes the underlying mechanism driving this phenomenon.<n>We find that the improvements arise from a redistribution of activations in the LLM's layers, where near zero activations become less frequent while large magnitude activations increase.<n>We propose a lightweight inference-time technique that modifies activations directly without altering the input sequence.
  arXiv  Detail & Related papers  (2025-10-01T15:39:38Z)
• Unveiling the Latent Directions of Reflection in Large Language Models [3.396557052704669]
  We investigate reflection through the lens of latent directions in model activations.<n>New reflection-inducing instructions can be systematically identified, and reflective behavior can be directly enhanced or suppressed.<n>This work opens a path toward mechanistic understanding of reflective reasoning in large language models.
  arXiv  Detail & Related papers  (2025-08-23T11:05:15Z)
• Dereflection Any Image with Diffusion Priors and Diversified Data [86.15504914121226]
  We propose a comprehensive solution with an efficient data preparation pipeline and a generalizable model for robust reflection removal.<n>First, we introduce a dataset named Diverse Reflection Removal (DRR) created by randomly rotating reflective mediums in target scenes.<n>Second, we propose a diffusion-based framework with one-step diffusion for deterministic outputs and fast inference.
  arXiv  Detail & Related papers  (2025-03-21T17:48:14Z)
• Two-Stage Single Image Reflection Removal with Reflection-Aware Guidance [78.34235841168031]
  We present a novel two-stage network with reflection-aware guidance (RAGNet) for single image reflection removal (SIRR) RAG can be used (i) to mitigate the effect of reflection from the observation, and (ii) to generate mask in partial convolution for mitigating the effect of deviating from linear combination hypothesis. Experiments on five commonly used datasets demonstrate the quantitative and qualitative superiority of our RAGNet in comparison to the state-of-the-art SIRR methods.
  arXiv  Detail & Related papers  (2020-12-02T03:14:57Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.