All papers

Introduces a self-supervised consistency loss across varied tool interfaces to make language model agents robust against notation and naming changes.

The paper proposes a runtime monitoring method to detect and correct language model agents that deviate from intended procedures without reporting it.

The paper introduces a benchmark for medical knowledge poisoning and a routing controller that selectively avoids unreliable retrieved evidence during clinical reasoning.

This paper proposes a multi-agent system designed to reduce clinical medication errors through dynamic collaborative coordination.

This paper presents a training-free contrastive decoding framework using attention-guided perturbations to reduce hallucinations in video language models.

The paper presents a Lean 4 runtime allowing self-extending agents to run verified computations across tools without constant human oversight.

The paper proposes a likelihood-based attribution method to identify the source model of machine-generated text under distribution shifts.

The paper introduces a training-free controller that uses token entropy to target and regenerate faulty reasoning steps at test time.

The paper introduces an internal probing method to detect whether safety-aligned language models still retain accessible representations of hazardous concepts.

The paper analyzes Best-of-N sampling as an extreme-value competition between tails of correct and incorrect responses to explain reward hacking.

The paper evaluates the reliability, score variance, and consistency of automated peer reviews generated by large language models.

The paper introduces a framework to audit cultural, geographic, and user-representation biases embedded within popular machine learning benchmarks.

The paper develops an analytical framework to interpret and reverse-engineer incentives and cooperative behaviors in multi-agent artificial intelligence systems.

This paper introduces MolLean, a verifiable language and tool interface that grounds chemical reasoning in formal state transitions and explicit checks.

This paper proposes Claim Cards to make agent interpretations explicit when verifying ambiguous natural-language scientific claims.

This paper surveys 138 studies on sparse autoencoders in NLP, examining dictionary design, feature recoverability, and practical utility.

The paper formalizes clarification criteria to estimate model-dependent input ambiguity based on accessible and valid answers.

The paper introduces a lightweight gray-box hallucination detector combining token probabilities with vector representations of token identities.

The paper systematizes grading requirements for open-text legal exams and compares eight evaluation paradigms using human and model judges.

The paper introduces a threat modeling framework examining adversarial query demotion and graph edit attacks on graph-based RAG systems.

This paper constructs an empirical graph of 62 private transformer inference systems to evaluate comparability, cryptographic progress, and scaling behavior.

Analyzes why feature-based test-time adaptation fails to recover accuracy on corrupted inputs in generative multimodal language models.

This paper analyzes the mathematical limits of statistical safety certification for embodied and agentic AI facing adaptive adversaries.

Studies statistical calibration techniques to bound risk when fast language models make automated decisions under uncertainty.

The authors propose a learned steering agent that monitors execution context and intervenes to prevent goal drift in tool-using language model systems.

The authors present a random projection method to protect sensitive datastore embeddings against extraction attacks in retrieval-augmented generation systems.

The paper builds state graphs across multiple agent rollouts to evaluate behavioral consistency and predict execution failures.

The paper separates task generation from model update inheritance to avoid compounding errors in self-evolving language model loops.

The authors evaluate how easily item identities and user profiles can be recovered from language embeddings in recommender systems.

The authors introduce a framework to benchmark language model agent robustness under interaction perturbations across diverse tasks.

Investigates how partial interventions in concept-based models can unexpectedly reduce downstream task performance due to information leakage.

The paper analyzes multi-reward reinforcement learning failures where auxiliary objectives degrade primary performance and introduces a trust-anchored gating method.

The paper proposes a test-time aggregation method using bias control and injection to resolve disagreements among reasoning language model agents.

Estimates language model ensemble member reliability from unlabeled agreement patterns using an expectation-maximization jury formulation.

Proposes an adversarial meta-prompt evolution method to improve model-generated evaluation rubrics for open-ended tasks.

Audits how persistent conversational agents remain with therapeutic crisis referrals after users repeatedly refuse help for suicidal thoughts.

Presents an evaluation system that tracks state and evidence to ensure autonomous agents fulfill obligations across long-horizon executions.

The paper proposes a multi-tiered governance model regulating AI system access based on user purpose and model capability.

Presents an adversarial training approach designed to improve language model honesty across diverse benchmarks without degrading core capabilities.

The paper presents a framework and dataset for learning to detect and repair errors during multi-agent system execution.

The paper introduces an expert-guided benchmark to measure the impact of generative language models on user wellbeing.

Introduces a benchmark to test model generalization under systematic convenience sampling biases in tabular data.

The paper introduces QuasiLeak-Bench to evaluate whether privacy detectors can identify distributed quasi-identifiers governed by disclosure policies.

Investigates how canary placement and training context affect secret memorization and extraction in fine-tuned language and multimodal models.

Surveys privacy accounting frameworks and differential privacy guarantees across single-query and multi-query retrieval-augmented generation systems.

This paper categorizes failure modes when deploying neural networks in closed-loop control systems and proposes online adaptation policies to counter distribution shifts.

This paper introduces a stateful testing method to detect bugs in deep learning compilers caused by retained state across invocations.

The paper applies game theory to analyze market incentives for AI safety investments across upstream foundation model providers and downstream firms.

The paper introduces a training-free weight-editing method to reduce large language model compliance with harmful instructions such as lying and sexism.

The paper introduces a causal evaluation method for healthcare language model agents to iteratively select and retain useful task skills.

The paper introduces an update-matched evaluation reference to balance unlearning effectiveness with downstream learning ability in language models.

The paper introduces a watermarking method to verify ownership and trace unauthorized copying of agent skill instructions.

This paper introduces an embedding perturbation technique to improve exploration diversity during reinforcement learning with verifiable rewards for language models.

Evaluates explainability methods across CNN, ViT, and Mamba models using eye-tracking gaze data as ground truth.

Introduces an information-theoretic framework using rate-distortion theory to establish computable lower bounds on feature reconstruction under DP-SGD.

Describes the Spec Growth Engine, which uses deterministic validation and LLM agents to prevent specification-code divergence.

The paper develops a framework explaining why large rollout batches in reinforcement learning reduce sample efficiency despite maintaining drift thresholds.

This paper investigates whether cloning a familiar voice systematically transfers personality impressions to an embodied virtual character.

Introduces an adaptation framework for frozen vision-language-action policies using residual reinforcement learning and online capability inference during joint faults.

Proposes a reliability-aware conditioning method that mitigates harmful temporal misalignments when using generated video futures for robot manipulation.