All papers

The paper derives mathematical bounds to evaluate the calibration of query-based probabilistic models using logical consistency without ground-truth labels.

The paper investigates data poisoning attacks targeting pre-training and retrieval corpora of protein language models and generative design systems.

Presents a bound-propagation verification framework to bound worst-case reconstruction errors for deep joint source-channel coding decoders under noise.

This paper investigates internal model representations that separate assistant actions from externally provided text across conversational layers.

Introduces a Ukrainian meme dataset and evaluates vision-language models on detecting persuasive and propaganda techniques.

Analyzes how prompt variations redistribute internal feature activations in sparse autoencoders and traces their impact on downstream model decisions.

The paper evaluates the impact of sarcasm on Turkish hate speech detection models using a targeted external test dataset.

Develops a closed-loop distillation pipeline where student feedback refines teacher models to improve adversarial robustness.

The paper presents a metric and benchmark measuring how quickly environment dynamics can change before agent decision-making fails.

The paper proposes an algorithm using betting-based sequential tests to monitor supermartingale certificate violations during runtime execution of stochastic systems.

The paper presents a neural-symbolic method for multi-robot systems to evaluate robot and object hypothesis trust purely via inter-robot observation consistency.

Introduces an imprecise Dirichlet model method to calculate interval-valued Shapley values and analyze attribution robustness in tree-based models.

Develops a personalized federated preference optimization framework maintaining local LoRA adapters atop a globally synchronized model.

Develops robust estimators and optimization algorithms to learn linear dynamical systems from trajectories corrupted by adversarial outliers.

The paper presents an evidence-driven multi-agent framework for collaborative vulnerability understanding and automated repository-level patch generation.

The paper introduces a logit adjustment method based on pairwise comparisons between sampled reasoning paths to improve test-time answer selection in language models.

Presents an open-source library that unifies training, evaluation, reinforcement learning, and distillation workflows for tool-using language model agents.

Introduces a low-latency guardrail model that routes incoming agent actions through editable policy trees for safety auditing.

Proposes dynamic activation steering using preceding denoising predictions to reduce hallucinations in diffusion language models.

The paper presents a multi-stage automated labeling framework using multimodal models and agents to annotate short videos for content moderation.

Investigates how changing tasks redistributes attention head influence and alters internal activations in open-weight large language models.

Decomposes joint temporal logic constraint violations to allocate credit among agents in cooperative reinforcement learning.

Introduces a training-free semantic text watermarking method that aligns cumulative sentence embeddings with a secret Gaussian vector.

Presents a local agentic framework using smaller language models to automatically synthesize edge-case testing scenarios for autonomous systems.

Introduces representation geometry regularization to prevent capability degradation during on-policy self-distillation in language models.

Develops a low-rank generalized Laplace approximation over active curvature directions for scalable neural network uncertainty quantification.

Presents a stateless service and library providing isolated, semantics-preserving text perturbations for language model robustness testing.

Implements encrypted WebAssembly execution on Arm TrustZone to protect proprietary AI models on edge hardware.

The paper presents an adversarial perturbation defense that disrupts the value vector diversity in image-to-video diffusion models to prevent unauthorized generation.

The paper designs a quantized gradient checksum interrogation scheme for cross-silo federated learning to detect active inference attacks under floating-point variations.

The paper proposes embedding multi-scale wavelet-domain watermarks into initial diffusion noise to enable robust provenance tracking under image removal attacks.

The paper proposes random allocation batching and GPU-accelerated accounting to improve accuracy and efficiency in differentially private model training.

The paper introduces a hypergraph autoencoder framework to filter malicious model updates and preserve Byzantine resilience in federated large language models.

Presents SADDER, an inference-time text diffusion safety framework using directional embedding signals to steer unsafe tokens toward benign distributions.

This paper proposes a preference optimization algorithm grounded in uniform information density theory to align language model outputs.

The paper proposes design requirements and evaluates defenses against model-side protection removal in open-weight image diffusion models.

The paper surveys security vulnerabilities and defenses in edge-deployed audio language model agents across the sensing-to-action pipeline.

The paper reports experiences using iterative AI agents to discover vulnerabilities across operating system kernels, drivers, and web applications.

Analyzes activation steering interventions in language models by isolating the effects of covariance filtering from target gain and displacement.

This paper introduces a learnable probabilistic instance sampler to handle severe class imbalance and prioritize critical diagnoses in multilabel learning.

The paper presents a feature-proxy method to selectively erase combined object-style concepts from diffusion models without degrading general generative performance.

The paper refines a high-level spatial AI orchestration framework into an implementation-focused low-level architecture for governed agent execution.

The paper evaluates how objective functions and candidate constraints affect decision conversion when generating radiology reports from sampled evidence.

The paper establishes theoretical bounds on interval lengths for split conformal prediction under Huber data contamination.

This paper extends backward conformal prediction to covariate shift settings using importance weighting derived from sorted score vectors.

Presents a taxonomy and research roadmap for measuring algorithmic fairness over time in systems with feedback loops.

This paper proposes a post-hoc framework based on Linearized Laplace approximation to quantify uncertainty across downstream tasks in contrastive learning.

This paper proposes a framework that dynamically steers internal language model neurons based on query-induced activation patterns.

The paper proposes a probabilistic decomposition of neural models that separates predictable features from an input-dependent reliability score.

The paper develops a theoretical Rényi-divergence framework to measure inherent differential privacy guarantees obtained directly from Bayesian posterior sampling.

The paper develops quality functions based on distance metrics to assess the diversity of generated test scenarios for automated driving verification.

The paper applies a lightweight post-hoc classifier to evaluate confidence and filter out incorrect next-location predictions from recurrent models.

The paper introduces a framework providing analytical bounds for counterfactual reasoning in sequential decision processes under statistical and causal uncertainty.

The paper introduces an agentic framework that removes peripheral visual clues to protect geolocation privacy in images without adding visible artifacts.

The paper introduces a compiler framework that transforms defective natural-language protocols into verifiable laboratory execution workflows.

The paper presents a subspace construction method for continual low-rank adaptation to mitigate forgetting under differential privacy constraints.

The paper presents a byte-level transformer model designed to detect malware directly from raw executable byte sequences.

The paper proposes an artificial neural network framework designed to reject uncertain predictions to prevent misclassification errors.

The paper evaluates an eight-qubit variational circuit bottleneck to improve confidence calibration in synthetic speech detection.

The paper presents a method for robots to generate targeted preference queries to learn underspecified rewards from humans.