Claims Browser

CAST (Concept-guided Artifact Suppression Tuning) uses Sparse Autoencoders to expose sparse, human-auditable features from intermediate Transformer activations for clinical text classification

"CAST uses Sparse Autoencoders to expose sparse, human-auditable features from intermediate Transformer activations"

CAST improves over fine-tuned encoder baselines and remains competitive with strong LLM baselines on MIMIC-IV discharge-note mortality prediction

"On MIMIC-IV discharge-note mortality prediction, CAST improves over its corresponding fine-tuned encoder baselines and remains competitive with strong LLM baselines"

SAE-based approaches can provide auditable, feature-level audit trails showing clinical concepts supporting predictions and artifact concepts suppressed during training

"while producing a feature-level audit trail of the clinical concepts that support each prediction and the artifact concepts suppressed during training"

Large language models organize moral knowledge geometrically, with moral foundation representations spanning near-maximal independent dimensions while sharing a positive common component

"We find the directions neither collapse into a single moral detector nor isolate from one another. Rather, they span a near-maximal number of independent dimensions while sharing a positive common component."

The shared component in moral representation is moral-specific with much higher integration compared to matched non-moral concept batteries

"The shared component is the signature of integration, and it is moral-specific relative to a matched non-moral concept battery built identically (mean pairwise cosine 0.26 vs. 0.013)."

Moral knowledge geometry in LLMs is consistent across architectures and scale, and emerges early in pre-training before probe accuracy saturates

"The geometry is consistent across architectures and scale and reaches its integration regime early in pre-training, well before probe accuracy saturates."

LLM moral representations reflect corpus statistics rather than the individualizing/binding distinction predicted by Moral Foundations Theory

"The structure the model discovers shows no evidence of the individualizing/binding distinction predicted by Moral Foundations Theory (an underpowered test: only 20 candidate partitions exist) but rather reflects corpus statistics."

LLMs represent moral tension itself in dilemmas rather than pre-resolved judgments, with dilemma directions partially composing from component foundations but majority variance encoding conflict-specific structure

"Extending to moral dilemmas, each dilemma direction partially composes from its component foundations, at 2.7x a mismatched-pair baseline, while the majority of its variance encodes conflict-specific structure. The model represents moral tension itself, not a pre-resolved judgment."

Vision-language models contain Visual Retrieval Heads (VRHs), a small subset of about 1.7-2.6% of attention heads that are causally responsible for grounding text descriptions to image regions

"Visual Retrieval Heads (VRHs), a small subset of attention heads (about 1.7-2.6%) that are causally responsible for grounding text descriptions to image regions"

Masking only the top 20 VRHs reduces grounding accuracy by up to 80 percentage points across eleven VLMs and five benchmarks, demonstrating their causal importance

"Across eleven VLMs and five referring-expression benchmarks, masking only the top 20 VRHs reduces grounding accuracy by up to 80 percentage points, while masking the same number of random heads has little effect"

Visual Retrieval Heads generalize across visual reference tasks, remaining causal on attribute, spatial, counting, and visual-math benchmarks despite being discovered through bounding-box prediction

"they generalize across visual reference tasks, remaining causal on attribute, spatial, counting, and visual-math benchmarks despite being discovered through bounding-box prediction"

VRHs are functionally specific, preserving output format while corrupting localization

"they are functionally specific, preserving output format while corrupting localization"

VRHs are architecturally shared, transferring causally across VLMs that share an LLM backbone but differ in vision encoder, projector, and instruction tuning

"they are architecturally shared, transferring causally across VLMs that share an LLM backbone but differ in vision encoder, projector, and instruction tuning"

Misleadingness is frequently associated with heightened emotional arousal and distortive communicative intent

"is frequently associated with heightened emotional arousal and distortive communicative intent"

Reader-level interpretations can often be recovered from lightweight claim-and-context representations without access to richer contextual, evidential, and multimodal information

"reader-level interpretations can often be recovered from such limited representations"

Identifying how misleadingness is produced from lightweight representations remains considerably more challenging than recovering reader interpretations

"whereas identifying how misleadingness is produced remains considerably more challenging"

Reliable understanding of misleading mechanisms requires richer contextual and evidential grounding beyond lightweight representations

"reliable understanding of misleading mechanisms continues to require richer contextual and evidential grounding"

Influential public discourse can mislead not only through what is stated, but also through how information is framed, omitted, contextualized, and communicated

"Influential public discourse shapes public beliefs and can also mislead, not only through what is stated, but also through how information is framed, omitted, contextualised, and communicated."

Less research has focused on how misleadingness arises and shapes reader interpretations

"Yet less research has focused on how such misleadingness arises and shapes the interpretations formed by readers."