A hybrid randomized smoothing method yields a closed-form certificate for joint discrete-continuous perturbations that generalizes prior Gaussian and discrete smoothing approaches.
Mixed citations
Probing classifiers: Promises, shortcomings, and advances
Mixed citation behavior. Most common role is background (60%).
27
Pith papers citing it
134
external citations · OpenAlex
Background
60% of classified citations
citation-role summary
background 4
dataset 1
citation-polarity summary
representative citing papers
Memorized facts in fine-tuned LLMs often sit off the mid-layer reasoning path; relocating those representations recovers most multi-hop generalization failures.
Fragility, the activation noise level causing probe accuracy collapse, reveals evolving lexical-to-compositional moral encoding, layer robustness gradients, and fine-tuning differences invisible to saturated probing accuracy.
Steering vectors from frozen LM layers enable a lightweight classifier to detect machine-generated text robustly across domains, source models, and editing attacks.
Training-language dominance, not English inherent properties, determines brain-LLM alignment across English, Chinese, and French, with additional independent effects from typological distance concentrated in syntactic brain regions.
Bounded fitting can be extended to expressive description logics while retaining generalization guarantees and implemented practically via SAT solvers.
Training-free graph method with LM edge scoring and max-regret path cover recovers 95% successor edges on Glossa wrap-around layouts vs 50% for XY-cut and 88% on OmniDocBench multi-column vs 75% XY-cut.
CLAE steers frozen multi-quadrotor policies at inference by training a sparse autoencoder on activations, probing for relevant latents, and learning an RL policy to apply state-dependent affine edits for velocity control, formation preservation, and camera avoidance.
Introduces the directional linear separability measure (LSM) as an asymmetric diagnostic for one-sided affine separability of neural representations.
Epi-LLM integrates LLMs as agents in ABM epidemic simulations, finding reduced peak infections, 58-65% quarantine compliance, and perceived severity as top predictor with pseudo-R² 0.055 comparable to human data.
ACT*ONOMY is a Grounded-Theory-derived hierarchical taxonomy and open repository that enables systematic comparison and characterization of autonomous agent behavior across trajectories.
Instructions trigger a production-centered mechanism in language models, with task-specific information stable in input tokens but varying strongly in output tokens and correlating with behavior.
Visual fingerprints represent distributions of linguistic choices extracted from repeated LLM samples to enable direct comparison of behaviors under different generation conditions.
Later LLM layers align better with human cognitive effort in syntactic ambiguity than early layers do, indicating dual processing modes and complementary benefits from multi-layer probability updates.
Prompt injection works because models internally treat text that sounds like a trusted role as if it were tagged as that role, and this confusion can be measured before generation.
LLMs compress U.S. climate opinion diversity and apply uniform gender assumptions that match reality for White and Hispanic Americans but misrepresent Black Americans.
Emotion AI encounters an epistemic gap preventing recovery of individual emotion meanings from annotator distributions, supporting the norm of affective sovereignty.
Longitudinal study of 56,800 AI papers finds sixfold increase in code+data sharing from 2014-2024 with inferred reproducibility rising from 28% to 64%.
Video foundation models encode intuitive physics knowledge that is strongest in V-JEPA at intermediate-to-late layers and depends on pretraining type and probe design.
GPT-4o achieves macro F1 scores of 0.89 for politician face recognition and 0.86 for person counting in election Instagram stories, outperforming FaceNet512, RetinaFace, and Google Cloud Vision.
New Zealand Reddit users link language to place and form contiguous speech communities with complex geographic alignment; Word2Vec embeddings reveal semantic variations and shifts in NZ English on a 4.26 billion word corpus.
The survey organizes mechanistic interpretability techniques into a Locate-Steer-Improve framework to enable actionable improvements in LLM alignment, capability, and efficiency.
Contextually-enhanced transformers integrating timetable and occupancy data achieve 26.6% and 56.3% average MAE reductions in railway and building energy forecasting respectively, outperforming prior methods.
A survey of argumentation mining techniques that reviews models from structured text to social media and proposes a flexible conceptual architecture framework for social media data.
citing papers explorer
-
Certified Robustness under Heterogeneous Perturbations via Hybrid Randomized Smoothing
A hybrid randomized smoothing method yields a closed-form certificate for joint discrete-continuous perturbations that generalizes prior Gaussian and discrete smoothing approaches.
-
Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning
Memorized facts in fine-tuned LLMs often sit off the mid-layer reasoning path; relocating those representations recovers most multi-hop generalization failures.
-
When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis
Fragility, the activation noise level causing probe accuracy collapse, reveals evolving lexical-to-compositional moral encoding, layer robustness gradients, and fine-tuning differences invisible to saturated probing accuracy.
-
SV-Detect: AI-generated Text Detection with Steering Vectors
Steering vectors from frozen LM layers enable a lightweight classifier to detect machine-generated text robustly across domains, source models, and editing attacks.
-
Brain-LLM Alignment Tracks Training Data, Not Typology
Training-language dominance, not English inherent properties, determines brain-LLM alignment across English, Chinese, and French, with additional independent effects from typological distance concentrated in syntactic brain regions.
-
Bounded Fitting for Expressive Description Logics
Bounded fitting can be extended to expressive description logics while retaining generalization guarantees and implemented practically via SAT solvers.
-
Reading Order Inference for Complex Document Layouts
Training-free graph method with LM edge scoring and max-regret path cover recovers 95% successor edges on Glossa wrap-around layouts vs 50% for XY-cut and 88% on OmniDocBench multi-column vs 75% XY-cut.
-
Steering Multirobot Behavior via Closed-Loop Affine Activation Editing
CLAE steers frozen multi-quadrotor policies at inference by training a sparse autoencoder on activations, probing for relevant latents, and learning an RL policy to apply state-dependent affine edits for velocity control, formation preservation, and camera avoidance.
-
A Geometric Measure of Linear Separability for Neural Representations
Introduces the directional linear separability measure (LSM) as an asymmetric diagnostic for one-sided affine separability of neural representations.
-
The Epi-LLM Framework: probing LLM behavioral priors through epidemiological agent-based models
Epi-LLM integrates LLMs as agents in ABM epidemic simulations, finding reduced peak infections, 58-65% quarantine compliance, and perceived severity as top predictor with pseudo-R² 0.055 comparable to human data.
-
How to Interpret Agent Behavior
ACT*ONOMY is a Grounded-Theory-derived hierarchical taxonomy and open repository that enables systematic comparison and characterization of autonomous agent behavior across trajectories.
-
Instructions Shape Production of Language, not Processing
Instructions trigger a production-centered mechanism in language models, with task-specific information stable in input tokens but varying strongly in output tokens and correlating with behavior.
-
Visual Fingerprints for LLM Generation Comparison
Visual fingerprints represent distributions of linguistic choices extracted from repeated LLM samples to enable direct comparison of behaviors under different generation conditions.
-
Dual Alignment Between Language Model Layers and Human Sentence Processing
Later LLM layers align better with human cognitive effort in syntactic ambiguity than early layers do, indicating dual processing modes and complementary benefits from multi-layer probability updates.
-
Prompt Injection as Role Confusion
Prompt injection works because models internally treat text that sounds like a trusted role as if it were tagged as that role, and this confusion can be measured before generation.
-
How Large Language Models Systematically Misrepresent American Climate Opinions
LLMs compress U.S. climate opinion diversity and apply uniform gender assumptions that match reality for White and Hispanic Americans but misrepresent Black Americans.
-
Who Determines the Meaning of an Emotion? Affective Sovereignty as an Epistemic Consequence of Measurement Limits
Emotion AI encounters an epistemic gap preventing recovery of individual emotion meanings from annotator distributions, supporting the norm of affective sovereignty.
-
The Shift Toward Open and Reproducible AI Research
Longitudinal study of 56,800 AI papers finds sixfold increase in code+data sharing from 2014-2024 with inferred reproducibility rising from 28% to 64%.
-
Do Video Foundation Models Understand Intuitive Physics? A Layerwise Probing Analysis
Video foundation models encode intuitive physics knowledge that is strongest in V-JEPA at intermediate-to-late layers and depends on pretraining type and probe design.
-
Seeing Candidates at Scale: Multimodal LLMs for Visual Political Communication on Instagram
GPT-4o achieves macro F1 scores of 0.89 for politician face recognition and 0.86 for person counting in election Instagram stories, outperforming FaceNet512, RetinaFace, and Google Cloud Vision.
-
Language, Place, and Social Media: Geographic Dialect Alignment in New Zealand
New Zealand Reddit users link language to place and form contiguous speech communities with complex geographic alignment; Word2Vec embeddings reveal semantic variations and shifts in NZ English on a 4.26 billion word corpus.
-
Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models
The survey organizes mechanistic interpretability techniques into a Locate-Steer-Improve framework to enable actionable improvements in LLM alignment, capability, and efficiency.
-
Integrating the Expected Future in Load Forecasts with Contextually Enhanced Transformer Models
Contextually-enhanced transformers integrating timetable and occupancy data achieve 26.6% and 56.3% average MAE reductions in railway and building energy forecasting respectively, outperforming prior methods.
-
The evolution of argumentation mining: From models to social media and emerging tools
A survey of argumentation mining techniques that reviews models from structured text to social media and proposes a flexible conceptual architecture framework for social media data.
-
Formal Constraints on Dependency Syntax
Various formal constraints on dependency trees are explored as a middle ground between projectivity and fully unrestricted structures to better model real language.
-
Bridging the Linguistic Divide: A Survey on Leveraging Large Language Models for Machine Translation
A literature survey that organizes prompting, fine-tuning, preference optimization, and context-aware techniques for LLM-based machine translation with emphasis on low-resource languages.
- How Many Human Survey Respondents is a Large Language Model Worth? An Uncertainty Quantification Perspective