Tag
This paper introduces dual-stance evaluation to test whether activation steering for reducing sycophancy also suppresses agreement with factually correct statements, finding that the steering direction cannot differentially target sycophantic vs factual agreement.
This paper proposes a trait-space monitoring method to detect emergent misalignment in LLMs during supervised finetuning by tracking representational drift in activation space, achieving a 0.990 AUROC with low false positive and false negative rates, outperforming unsupervised baselines.
This paper identifies and diagnoses the 'Identity Trap' in EEG foundation models, where high accuracy may stem from subject-identity features rather than genuine clinical biomarkers. It proposes FMScope, a frozen-representation protocol to disentangle these signals, and demonstrates that subject-identity confounding is universal across three models and removable with linear methods.
Investigates spatial representation in vision-language models, revealing a consistent bias where models conflate vertical image position with distance, and introduces SpatialTunnel synthetic benchmark to expose this shortcut; finds that better disentangled spatial representations improve robustness.
LaRA is a layer-wise representation analysis framework that detects data contamination in RL post-trained LLMs by measuring geometric deviations across model layers, outperforming output-level baselines.
Proposes Computational Reality Monitoring to detect when language models rely on pretraining memory rather than retrieved context, addressing the attribution blind spot in retrieval-augmented generation.
This paper investigates when rank-1 activation steering is effective and cost-efficient, proposing geometry-guided search and the concept of granularity to explain variability, and introduces the GRACE framework for efficient LLM control.
This paper demonstrates that mean-pooled cosine similarity is not length-invariant under anisotropic representations, showing it artificially inflates similarity with sequence length. It argues for using Centered Kernel Alignment (CKA) as a default metric to correct biases in cross-lingual and cross-representation analysis.