All papers

This paper predicts human annotator disagreement on harmful content across multiple languages by extracting interpretable semantic question features with models.

The paper trains language models to report internal states through supervised fine-tuning and on-policy learning.

The authors introduce a calibration harness using structured domain knowledge to safely delegate Earth science model parameterization to language model agents.

This paper introduces an instruction-preserving decoding strategy to prevent visual contrastive decoding from discarding valid abstention responses.

This paper introduces an evaluation index to measure whether machine translation of emergency communication preserves critical safety actionability.

This paper presents a label differential privacy framework to protect survey-derived ordinal preference data during language model alignment.

The paper examines conceptual parallels between measurement science and machine learning uncertainty evaluation to propose a more unified assessment framework.

This paper compares observational feature discovery scores in sparse autoencoders against actual causal intervention effects on model outputs.

The paper evaluates how commercial language models value human time by measuring recommended delay compensation amounts across 193 different countries.

Demonstrates that discrete graph diffusion models exhibit conditional memorization when reverse trajectories are conditioned on corrupted training graphs.

Employs a bilevel Stackelberg formulation to jointly optimize classification accuracy and confidence calibration without separate calibration datasets.

Studies how fine-tuning helps tool-using agents learn to withdraw and repair actions when environment assumptions change.

Introduces MAGMA-Bench to evaluate long-horizon embodied agent reasoning during dynamic task execution failures and multi-robot coordination.

Introduces an execution-grounded dataset and method to repair geometric and engineering constraint violations in LLM-generated industrial scene programs.

Audits audio-language model hallucination mitigations by testing localized acoustic evidence ablation against baseline methods across model architectures.

Proposes a benchmark and probing methodology based on attribution theory to examine how language models assign responsibility and blame.

Proposes an adaptive multi-source steering decoding method to mitigate hallucinations in vision-language models by balancing visual and textual context contributions.

Introduces a training-free inference method using Gaussian optimal transport to reduce object hallucination in vision-language models without suppressing valid representations.

Analyzes interaction failures between KL regularization and rewards in group policy optimization and proposes a calibrated surrogate objective.

Presents a hypothesis testing framework using shadow models to audit whether a language model was distilled from a proprietary teacher.

This paper introduces a counterfactual benchmark to evaluate honorific register allocation bias across gender and occupations in English-to-Bengali translation.

This paper audits Chinese language models for bias and error propagation across formal and ironic social registers.

The paper proposes an architectural ML Plane to expose and audit privacy risks of federated learning in operational technology intrusion detection systems.

The paper proposes a framework that embeds keyed Fourier-phase signals into face images to enable verification against publisher release records.

The paper introduces confidential error-correcting codes, providing a theoretical framework and concrete construction balancing message privacy with reliable decoding.

The paper evaluates representation-based defenses against backdoors in federated intrusion detection systems using disjoint server data.

The paper introduces a decision-layer watermarking method for multi-agent routing to protect routing policies against unauthorized cloning.

This paper introduces a framework combining zero-knowledge proofs and serving certificates to verify model selection and query rewriting in routers.

This paper evaluates whether existing deepfake detection models can identify physical humanoid face robots performing speech-synchronized lip movements.

This paper establishes mathematical conditions to guarantee long-term performance and low counterfactual regret for reflective language agents under resource constraints.

The authors introduce a parameter-free conformal martingale method modeled as a universal portfolio to quickly detect distribution shifts at unknown times.

The paper analyzes the theoretical detection limits and sum-test efficiency of textGrain watermarking under human text edits.

The paper proposes an inference-time alignment method that optimizes expected costs while controlling tail risk in diffusion models.

This paper establishes authorization models and integrity proofs to prevent narrative model history from becoming unauthorized actions.

The paper proposes an embedding-based graph kernel entropy method for computationally efficient hallucination detection in language models.

The paper presents a framework using Wasserstein geometry and Shapley values to audit and decompose intersectional fairness disparities across multiple attributes.

The paper develops an adaptive computation framework offering conformal uncertainty guarantees for processing large scientific and medical imagery efficiently.

The paper introduces a federated multi-agent framework that guides agent selection and query refinement under dynamic privacy constraints.

The paper compares mental health stigma representations in large language models against human survey data across different gender groups.

The paper introduces a two-stage reasoning framework to generate personalized, safe, and explainable medication recommendations using language models.

The paper introduces a benchmark to measure vision-language-action model robustness against hardware, optical, and physical robot setup variations.

Shows that evil persona vectors extracted from low-resource languages transfer poorly and fail to steer model behavior effectively across languages.

The paper introduces a test-time wrapper that distorts exposed output probabilities to defend black-box machine learning models from query-based adversarial attacks.

Introduces HelpMeOut, an embodied AI benchmark for diagnosing human execution errors across household activity trajectories motivated by neuropsychology research.

The paper analyzes forty legal instruments to examine how copyright governance protects scholarly authors whose works train generative AI models.

The paper formalizes transparency as role-dependent answerability using information flow networks to evaluate AI-supported decision systems.

The paper proposes an unsupervised drift monitor using internal embedding distances to flag unfamiliar inputs in deployed deepfake speech detectors.

This paper presents an imitation framework that enables agents to identify violated contracts, extract skills, and repair execution failures.

The paper analyzes visual attention sinks in vision-language models and shows they stem from alignment with language text sinks.

This paper presents an adversarial self-play framework using multi-perceptual views to improve language model detection of obfuscated harmful text.

The paper introduces TRACE to monitor harmful response risks during generation, edit token activations selectively, and maintain auditable decision traces.

The paper develops an attack framework to generate physical adversarial textures that deceive video models across varying trajectories and viewpoints.

This paper proposes a watermarking method embedded during speech synthesis to trace and verify generated audio.

This paper fine-tunes language models to selectively correct speech recognition transcripts using curated candidate pools.

This paper analyzes prediction errors in spectral saliency when deleting singular components to selectively edit language model capabilities.

The paper estimates reasoning confidence by segmenting chain-of-thought traces into argument graphs evaluated through natural language inference.

The paper studies Byzantine robustness in fair federated learning and proposes an update screening method using client trajectory consistency.

Combines language model review with mechanical numerical checks to verify mathematical formalizations and catch errors missed by automated judges.

The paper presents a benchmark evaluating how reliably language models correct and maintain formatting requirements across changing instructions.

The paper presents an empirical framework measuring neural network backdoor stealthiness across input, feature, and parameter spaces.