Introduces Causal Functional Signatures grounded in causal evidence and ILP-learned architectural signatures to enable explicit, comparable, and portable mechanistic claims across model scales.
hub
author Montani, I
28 Pith papers cite this work, alongside 425 external citations. Polarity classification is still indexing.
28
Pith papers citing it
425
external citations · Crossref
hub tools
citation-role summary
method 2
citation-polarity summary
roles
method 2polarities
use method 2representative citing papers
A new linked multimodal dataset of Russian domestic and foreign policy speeches with texts, images, captions, harmonized metadata, and expert-refined topic annotations is introduced to support analyses in political communication and LLM applications.
Semantic search retrieves substantially more implicit receptions of Locke's work than lexical baselines in 18th-century corpora, yet remains constrained by lexical gatekeeping.
Brazilian YouTube climate videos show a transition from traditional denial of climate science to 'new denial' that undermines solutions, with the latter attracting more engagement from diverse actors.
LLMs exhibit positional bias and context-dependent scoring patterns when judging document similarity, with each model showing a stable scoring fingerprint but a shared hierarchy of sensitivity to different semantic perturbations.
LLMs display a consistent pattern of elevated form-meaning divergence and uniform rhetorical device use in argumentative texts compared to humans, quantified by new metrics FMD, GPR, and RDDE.
SAM 3 introduces promptable concept segmentation that doubles accuracy of prior systems on images and videos while improving standard SAM segmentation performance.
TSVer is a new benchmark dataset for fact verification against time-series evidence, with 304 annotated real-world claims, 400 time series, verdicts, and justifications, plus baseline results showing current models struggle.
IPA-based subword tokenizers trained across 24 languages improve tokenization quality and generalization to unseen languages compared to standard text tokenizers, especially for non-Latin scripts.
Systematic experiments reveal that activation steering trades fluency for concept control, is less effective on instruction-tuned models, and that prompting/SFT excel at injection but not removal, with textual metrics correlating to LLM judges.
AraSEG is a genre-diverse Arabic sentence segmentation corpus showing lightweight encoders and dependency parsers outperform LLMs under challenging punctuation while improving downstream parsing.
Introduces a triangulation-based metric to quantify lexical shifts attributable to preference tuning without requiring manual curation of examples.
Structural probes on UD-invariant wh-movement stimuli reveal phase-count gradients and phase-internal cohesion effects in 12-13 of 13 LLMs, indicating syntactic abstractions beyond UD annotations.
Constructs continuous sign conversation data from isolated signs using retrieval and diffusion models to train a direct sign-to-sign conversational AI.
ATD-Trans is a new geographically annotated Japanese-English travelogue dataset that reveals Japanese-enhanced models perform better on geo-entity translation while domestic Japanese locations remain harder to translate accurately.
An encoding probe reconstructs transformer representations from acoustic, phonetic, syntactic, lexical and speaker features, showing independent syntactic/lexical contributions and training-dependent speaker effects.
MediaGraph uses co-occurrence networks from Indian news on farmer protests and a new link predictability metric to reveal source-specific reporting preferences and under-representation of farmer leaders.
A framework using language models to simulate non-existent experiments and derive novel testable hypotheses on dative verb acquisition and cross-structural generalization in children.
LLMs exhibit different trade-offs between rule compliance and communicative success across prompting, generation constraints, and representation interventions, but remain substantially weaker than humans at guessing under lexical constraints.
Dual-encoder VLMs gain robust compositional generalization by learning localized alignments from frozen patch and token embeddings instead of using global similarity.
Contradictions between highly similar medical abstracts degrade the factual accuracy and consistency of LLM responses in retrieval-augmented generation.
Entity-based chunk filtering reduces RAG vector index size by 25-36% with retrieval quality near baseline levels.
No single privacy technique wins; combining local inference, redaction, and semantic rephrasing limits PII leaks to 0.6% and proprietary code leaks to 31.3% on a 1,300-sample benchmark, with code released.
Empirical comparison finds tokenization most important and recommends specific preprocessing order for Twitter sentiment analysis models.