A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinations
Contextual hallucinations -- statements unsupported by given context -- remain a significant challenge in AI. We demonstrate a practical interpretability insight: a generator-agnostic observer model detects hallucinations via a single forward pass and a linear probe on its residual stream. This probe isolates a single, transferable linear direction separating hallucinated from faithful text, outperforming baselines by 5-27 points and showing robust mid-layer performance across Gemma-2 models (2B to 27B). Gradient-times-activation localises this signal to sparse, late-layer MLP activity. Critically, manipulating this direction causally steers generator hallucination rates, proving its actionability. Our results offer novel evidence of internal, low-dimensional hallucination tracking linked to specific MLP sub-circuits, exploitable for detection and mitigation. We release the 2000-example ContraTales benchmark for realistic assessment of such solutions.
Code (0)
등록된 구현이 없습니다.
Similar Papers 제목 키워드 기반
Fault Diagnosis and Bad Data Detection of Power Transmission Network - A Time Domain Approach
Fault analysis and bad data are often processed in separate manners. In this paper it is proved that fault as well as bad current measurement data can be modeled as control failure for the power transmission network and …
Fault DiagnosisReservoir observer enhanced with residual calibration and attention mechanism
Reservoir observers provide a data-driven approach to the inference of unmeasured variables from observed ones for nonlinear dynamical systems. While previous studies have demonstrated wide applicability, their performan…
Cascaded Tightly-Coupled Observer Design for Single-Range-Aided Inertial Navigation
This work introduces a single-range-aided navigation observer that reconstructs the full state of a rigid body using only an Inertial Measurement Unit (IMU), a body-frame vector measurement (e.g., magnetometer), and a di…
Sycophancy Hides Linearly in the Attention Heads
We find that correct-to-incorrect sycophancy signals are most linearly separable within multi-head attention activations. Motivated by the linear representation hypothesis, we train linear probes across the residual stre…
Observer-Assisted Relative-Velocity Compensation with LPV-$H_\infty$ Robust Correction for 3D Trajectory Tracking of Underactuated Non-Minimum-Phase AUVs under Ocean Currents
This paper develops an observer-assisted control architecture for 3D trajectory tracking of torpedo-type underactuated AUVs with non-minimum-phase sway/heave dynamics under unknown ocean currents. A three-stage state-cur…