Tag
This paper demonstrates that linear probes decoding concepts from language model activations do not necessarily identify causally relevant features, and introduces a feature-level diagnostic using sparse autoencoders to separate probe alignment from behavioral drivers.
The paper proposes CMR-Mamba, a causal mechanism monitoring method using Mamba state-space models for unsupervised industrial fault detection, focusing on coupling faults and evaluated on electromechanical, hydraulic, and cyber-physical systems.
This paper investigates whether vision-language models exhibit reward valuation circuits analogous to those in humans. By perturbing functionally identified reward-anticipatory units, the authors induce anhedonia-like behavioral deficits in the model, mirroring clinical motivational impairments.
This paper investigates how vision-language models resolve conflicts between visual evidence and world knowledge, revealing that visual grounding is the default while prior knowledge depends on a small set of late-layer attention heads. The authors perform causal analysis across three VLM families, demonstrating an asymmetric structure where ablating these heads shifts predictions from knowledge-grounded to visually grounded answers.