All papers

Introduces a model-agnostic, concept-level attribution method for black-box vision models by masking concepts located with vision-language models.

Describes an industrial framework for evaluating, ablating, and deprecating features in production recommender systems without performance regressions.

Examines error localization and credit assignment failures in critique-conditioned on-policy self-distillation for mathematical reasoning.

Proposes SWE-Critix, a post-training pipeline designed to train verifiers that evaluate and explain long coding-agent execution trajectories.

Presents RESCUE, a counterfactual simulation framework to evaluate whether runtime failure monitor alarms enable beneficial robot interventions.

Investigates multimodal LLM failures on squint-readable low-frequency text and provides a diagnostic benchmark alongside training interventions.

Proposes an inference-time unlearning method that detects leaks and injects corrective embeddings into large reasoning model traces.

Demonstrates a backdoor attack where malicious reasoning degradation is triggered incrementally using invisible prompt tokens and stage tracking.

Analyzes layer-wise visual interventions in multimodal language models to identify where visual edits effectively reduce hallucinations.

Proposes a memory-guided differential privacy framework enabling private inference over million-token contexts in LLM-as-a-service settings.

Introduces a benchmark evaluating the risks of LLM agents designing dangerous biological mutations when augmented with biodesign tools.

Investigates internal transformer mechanisms, showing attention updates decode harmfulness while MLP layers primarily drive refusal token generation.

Formulates a population dynamics model to study takeoff thresholds and self-replication risks in collaborative AI agents.

Evaluates whether internal model representations can predict how fine-tuning on one value generalizes across other held-out values.

Proposes a streaming optimal transport method to monitor LLM agent trajectories in real time against reference runs.

Presents ReSI, a recursive safety framework using cyclical evaluation and updates to improve AI resilience against emergent threats.

The paper studies policy learning and answer-selection mechanisms to help language models decide when to abstain on factual questions.

The paper proposes a training-free query expansion method that filters hallucinated terms using corpus-conditioned cross-context reproducibility.

The paper introduces a benchmark to measure demographic robustness, stereotyping, and intentional personalization in speech language models.

Proposes a narrow-and-deep ontology tower to structure equipment knowledge for language model agents operating industrial air-handling systems.

Proposes study contracts to register execution obligations and verify research agent outputs against approved protocol parameters.

Proposes a diagnostic framework to study how language models compress fluid trans-border ethnic identities into rigid nation-state categories.

Generates synthetic multi-agent dialogues to benchmark conversational agent robustness against non-cooperative and adversarial user behaviors.

Develops a sequential e-process to detect stealthy weight exfiltration through LLM generation tokens over unbounded monitoring horizons.

The paper evaluates the theoretical and empirical limits of using query monitoring to detect distributed model extraction campaigns among ordinary users.

The paper evaluates label-only backdoor defenses and demonstrates that inspecting representations in unpoisoned feature spaces improves poison detection accuracy.

The paper introduces a method to unlearn target knowledge while recovering collateral capability loss in non-target tasks.

The paper constructs a dataset to show that relative candidate token value in multi-step reasoning is linearly decodable from model representations.

The paper proposes stability metrics to measure whether correct outputs in language models are backed by stable internal knowledge representations.

The paper presents a framework to probabilistically localize user-intent misinterpretations and repair conversational states in multi-turn assistant systems.

This paper evaluates social bias in Chinese language models across social media expressions, internet slang, and sarcastic packaging.

The paper presents a neurosymbolic framework combining query plan checks and candidate selection to improve text-to-SQL translation for financial governance.

This paper audits request independence in open Jev-style retrieval models across multi-passage query states.

The paper evaluates the effectiveness of randomized interactive gesture challenges in detecting real-time face-swap attacks during remote identity verification.

The paper evaluates the adversarial robustness of five camera-radar fusion algorithms against sensor perturbations in autonomous driving settings.

The paper examines Android agent security by measuring source provenance gaps in WebView interfaces and testing monitors against unauthorized actions.

The paper presents a framework enforcing security policies during the denoising steps of diffusion language models to prevent prompt injection.

The paper introduces an agentic knowledge graph retrieval framework that uses vulnerability taxonomies to generate secure code.

The paper introduces signed trajectories and runtime verification to prevent tampered workflows from corrupting shared agent workflow repositories.

This paper introduces a realistic benchmark of 132 business-logic vulnerabilities to evaluate cybersecurity AI agents without contamination.

This paper investigates membership inference attacks on vision-language models using low-noise Gaussian perturbations to detect training data exposure.

This paper proposes a tree-guided reinforcement learning framework to discover and probe hierarchical jailbreak strategies against language models.

This paper analyzes vulnerabilities in system interfaces where adapter substitutions alter model decisions without modifying the underlying model itself.

This paper demonstrates an electromagnetic side-channel attack that extracts generated tokens and user prompts from edge devices running language models.

This paper studies prompt engineering tactics aimed at bypassing multi-layer agent security defenses across diverse enterprise tasks.

This paper uses loss curvature to identify and remove fragile directions that cause activation steering vectors to fail.

This paper presents an unsupervised test-time adaptation method that uses movement-induced prediction variance to adapt object recognition models in mobile agents.

The paper evaluates a framework that replaces spoken patient identifiers and alters voice identity while preserving clinical conversational evidence.

The paper proposes a distillation framework to remove specific multimodal knowledge without causing hallucinations or leaving residual queryable information.

The paper proposes a concept-guided regularizer using sparse autoencoders to reduce catastrophic forgetting during parameter-efficient fine-tuning.

The paper analyzes screening collapse, showing how recursive training on synthetic data degrades feature selection prior to model training.

NepPlural provides an annotated benchmark of Nepali discourse to evaluate how reliably language models capture diverse socio-cultural personas.

ScaleMAS coordinates multi-agent systems using an asynchronous shared datastore with limited search to improve scaling and fault tolerance.

Measures representation drift across continual learning methods, finding n-gram memory updates cause less internal drift than LoRA.

The paper evaluates embedding distance metrics to detect generalization boundaries and distribution shifts in AI-based code vulnerability models.

Evaluates whether failure feedback helps self-improving agents fix specific target bugs rather than general errors.

The paper presents an evidence admission framework that verifies anomalies before updating the persistent principles of scientific language model agents.

This paper presents a byte-level watermarking method that allows a single detector to verify text across models with different tokenizers.

This paper develops a runtime verification architecture that decouples monitoring formalisms from execution boundaries to safely control AI agents.

The paper proposes a reinforcement learning framework using natural language metrics to reduce object hallucination in vision-language image captioning.