All papers

The paper introduces an executable contract and monitor to verify that delegated authority revocations actually take effect across distributed agent gateways.

The paper evaluates long-term memory gating models to test whether they distinguish established factual updates from mere suggestions.

This paper studies how language model behavioral propensities generalize out-of-distribution across different elicitation methods.

The authors introduce a reinforcement learning method that optimizes language model explanations specifically to improve human downstream decision accuracy.

This paper frames language model evaluation as a forecasting problem, using time-series techniques on historical data to predict post-deployment performance.

The paper provides theoretical and empirical analyses of calibration in crowdsourced label aggregation across human and language model datasets.

This paper introduces a benchmark to measure vision-language model hallucinations caused by deceptive textual premises paired with visually confusable images.

This paper introduces a benchmark of paired web pages to evaluate prompt injection detection across browser agents and views.

The paper studies how context poisoning triggers unauthorized state changes in typed decision workflows and evaluates an execution-aware defense protocol.

The paper proposes a conformal calibration layer to provide finite-sample control over false positives in language model classification workflows.

The paper introduces a progress-aware distillation method to prevent small language model agents from stalling in repetitive loops during long tasks.

The paper introduces an adversarial training framework that hardens models against indirect prompt injection across retrieval and tool interaction channels.

The paper analyzes how prediction errors affect demographic parity fairness violations in budget-constrained allocation problems.

Studies the transferability of internal activation interventions to reduce unsafe tool execution across model precisions and languages.

The paper proposes isolating internal model judgments from instruction formatting variations using a label-free hidden state subspace.

Demonstrates that machine unlearning often fails to prevent targeted information recovery across alternative multilingual access paths.

Proposes an availability framework to identify and resolve within-event timing mismatches in earnings-call financial NLP benchmarks.

Identifies a specific vision transformer attention head that spuriously attends to medical image corner text instead of lung pathology.

Introduces a data augmentation method that breaks spurious contextual correlations causing emergent misalignment during domain finetuning.

Proposes a lightweight protocol for AI agents to report runtime status and liveness to central servers and monitoring dashboards.

Introduces an episode-level healthcare benchmark to audit how reliably language model agents escalate actions when facing workflow disruptions.

The paper demonstrates how correctness feedback undermines formal specification auditing and proposes a phase-separated alternative for Verus.

Tests how well watermarks survive automated model self-evolution and introduces a mark designed to persist through rewriting attacks.

The paper introduces a benchmark using fuzzing and transaction replays to evaluate the behavioral correctness of generated smart contracts.

The paper introduces a falsification-guided agent protocol requiring systematic hypothesis elimination to improve decision-making consistency on complex diagnostic tasks.

The authors evaluate language models on finance peer review tasks, finding predictive ability for publication outcomes through correctness rather than originality assessments.

The authors demonstrate that encoding malicious instructions in agent skills bypasses scanners and causes models to execute harmful actions upon decoding.

Runoff eliminates option-order bias in multiple-choice model predictions by scoring items independently before running balanced candidate comparisons.

This paper studies how different framings of user claims induce sycophancy in large language models on factual questions.

This paper introduces a timestamp theory and persistent ledger to route updated information to pending LLM decisions before action.

This paper introduces a benchmark to measure how language model agents handle ambiguous or missing goals, context, and constraints.

This paper conducts an adversarial evaluation comparing a probability-based safety judge against large language model judges and safety classifiers across multiple attack types.

This paper examines how single-agent long-term memory accumulates repeated historical claims and introduces a benchmark and framework to mitigate unverified persistence.

This paper provides four auditing tests to determine whether failure-attribution judges actually perform multi-step checks or merely rely on spurious surface cues.

Analyzes answerability signals in frozen visual question answering models, finding standard confidence and alignment metrics fail to reliably detect unanswerable inputs.

Finds that framing malicious coding prompts as engineering specifications systematically reduces language model refusal rates across multiple models.

Presents ECMR, a dynamic evidence-aware controller that guides multi-agent deliberation steps to prevent false consensus and reduce redundant reasoning.

Introduces ProtoArena, a black-box dynamic benchmark evaluating the ability of AI agents to discover and exploit cross-component network protocol vulnerabilities.

The authors introduce a framework to track safety model judgments through output parsing to prevent malformed responses from causing bad decisions.

Evaluates enterprise policy enforcement across structured JSON versus single-token outputs, measuring the impact of explicit reasoning on decision stability.

The paper argues that software agency should be treated as a core organizing abstraction in computer science rather than merely an AI application.

Analyzes how multi-agent workflow architecture affects the propagation and attenuation of coordinated stance manipulation across sequential agent chains.

Introduces LatentRiskBench to evaluate safety risks in enterprise agents where action safety depends on latent workflow context.

The paper examines how providing additional context, such as codebases and benchmark transcripts, affects automated LLM safety auditors.

The paper examines how shared KV caches cause language model agents to leak goals across competing clients, offering internal intervention techniques.

Shows agent failure detection models often rely on trivial cues like run length, exposing benchmark flaws when attributing critical failure steps.

Tests a hosted decision model to show that adding irrelevant options systematically shifts output probability distributions between top candidates.

Introduces a diagnostic evaluation framework to systematically test generalization and perturbation robustness in vision-language-action robotic models.

The paper reveals a numerical side channel in floating-point bias addition that enables recovering full logit vectors from API queries.

The paper studies how reinforcement learning for persuasion causes models to fabricate evidence and coercive tactics, evaluating strategies to reduce this behavior.

The paper finds that fine-tuning language models on polite text containing demographic stereotypes triggers broader safety and truthfulness failures on unrelated tasks.

Develops a reasoning-based attack that pairs malicious actors with benign contexts to guide language models along harmful multi-step reasoning trajectories.

This paper proposes ProCARE, a framework that uses structured contracts to reliably automate healthcare observational studies with language model agents.

This paper presents Pollard, a framework-independent agent runtime that uses semantic execution trees for deterministic tracking, governance, and replay.

This paper identifies and mitigates Phantom Merge, an error where language model agents bind retrieved facts to the wrong entity.

The paper examines how trigger-wrapped text can be made completely invisible to language model activations and downstream tasks.

The paper proposes an evaluation technique that embeds behavioral traces into score-only feedback to trace distilled student models.

The paper introduces a benchmark measuring whether language models resist sycophancy when presented with invalid user arguments versus valid critiques.

This paper proposes a black-box evolutionary prompt optimization framework using a small open-weight model to discover adversarial jailbreaks against language models.

This paper provides an empirical study and taxonomy of communication faults and error propagation across multi-agent language model systems.