Back to feed
arXiv cs.CL
arXiv cs.CL
6/25/2026
Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models

Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models

Short summary

Research on Gemma and three other language models shows that detecting hallucinations (perfect AUC=1.0) doesn't enable controlling them—the detection direction sits 83° away from the refusal direction. This detection-intervention gap persists across model families and scales, originating in pretraining. Control in LLMs requires high-dimensional intervention, not single-axis manipulation.

  • •Detection and control directions diverge by ~83 degrees in hallucination detection across tested models
  • •Detection-intervention gap originates in pretraining, not instruction tuning
  • •LLM behavior control is fundamentally high-dimensional, not reducible to single activation directions

Generated with AI, which can make mistakes.

#research-breakthrough#ai-agents#certification-education
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Explore more

Adaptive scheduling steers diffusion L
arXiv cs.LG
Investigating Linguistic Steering: An Analysis of Adjectival Effects Across Large Language Model Architectures
arXiv cs.CL
How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits
arXiv cs.CL
Belief or Circuitry? Causal Evidence for In-Context Graph Learning
arXiv CS.AI