Reliable Control-Point Selection for Steering Reasoning in Large Language Models

Dimensions overall score 7.0

• Current methods for detecting reasoning behaviors through keyword matching in chain-of-thought traces are overwhelmingly wrong, with 93.3% of detected boundaries being behaviorally unstable.limitation 95%
• Stability filtering combined with content-subspace projection achieves 0.784 accuracy on MATH-500, representing a +5.0 improvement over the strongest baseline.result 93%
• The proposed steering vectors transfer across models in the same architecture family without re-extraction, improving Nemotron-Research-Reasoning-1.5B by +5.0 and DeepScaleR-1.5B-Preview by +6.0.result 92%
• Stability filtering retains only boundaries where the model consistently reproduces the target behavior, addressing the problem of unstable boundaries diluting steering signals.method 88%
• Many reasoning behaviors such as self-reflection emerge spontaneously and resist prompt-level control, making them difficult to detect via conventional methods.technical 85%
• Unstable boundaries dilute the steering signal when constructing steering vectors for controlling reasoning behaviors.technical 82%
• The paper develops a probabilistic model that formalizes intrinsic reasoning behaviors as stochastic events with context-dependent trigger probabilities.method 90%
• Steering vectors offer a training-free mechanism for controlling reasoning behaviors in large language models.technical 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card