All papers

The paper presents an automated framework to evaluate how adversarial contexts can redirect program execution through valid structured judgments in language-model applications.

The paper introduces an imperceptible image perturbation attack that manipulates alignment in vision-language models during preference optimization without altering text annotations.

The paper introduces a maturity model that grades artificial intelligence agents across multiple capability levels, evidence stages, and safety gates.

The paper presents an empirical study of 654 real-world configuration bugs across ten retrieval-augmented generation systems, categorizing causes and fixes.

The paper introduces a diagnostic protocol to evaluate how speech recognition errors affect structured tool calling in voice agents.

The paper tests whether role-playing language models leak private or unobserved character information across simulated interactive scenes.

The paper examines machine grading on undergraduate essays across three universities, finding models compress score distributions and vary across institutions.

The paper introduces BRIDGE, a runtime intervention framework that detects and repairs broken information handoffs across dependency graphs in multi-agent workflows.

The paper tracks automatic speech recognition performance from 2021 to 2025, finding persistent accuracy disparities between standard American English and minority accents.

The paper evaluates how language models correct user misconceptions across five languages, showing indirect pushback often fails to fix factual errors.

Presents a randomized experiment framework to test and curate self-written agent playbooks, preventing task failure caused by unverified lessons.

Shows linear prompt-injection monitors overfit to conversation features and reduces their false alarms by adding matched harmless negative examples.

Demonstrates an attack method that alters embedded multi-bit text watermarks without knowing the secret generation key.

This paper introduces a system to ensure tool-using language model agents comply with natural language company procedural manuals.

This paper analyzes label collapse failure modes across diffusion and autoregressive language models during in-context learning classification.

This paper uses activation steering and probing to identify distinct internal representations for rule-based versus consequence-based moral reasoning in language models.

This paper proposes TAPS, combining standard accuracy training with post-hoc confidence selection to give controllable abstention rates without specialized tuning.

This paper introduces RelLedger, an auditing framework to trace how relational histories drive reporting choices and actions in multi-agent systems.

MedDEx combines language models for patient data extraction with deterministic rule engines to execute clinical treatment guidelines accurately and efficiently.

This work analyzes whether policy capability gaps or judge models dictate reward hacking when training models against language-model judges.

The paper introduces MemForm to benchmark adversarial memory formation poisoning across five input interfaces in language model agents.

The paper introduces a benchmark derived from real customer support dialogues to evaluate language models executing bounded, tool-using skills.

Combines dual-encoder semantic retrieval with a contextual bandit to adaptively select agent tools using execution reliability feedback.

Presents an automated red-teaming framework that searches for harmful compositional tool-use paths across multi-service environments.

Introduces an evaluation benchmark measuring how effectively language models provide supportive responses during self-harm and crisis disclosures.

Introduces DREAM, an agent-intrinsic defense that triggers functional forgetting of poisoned memory entries through implicit reactivation and topological diagnosis.

Introduces UPTURN, a framework that trains a small classifier to inject redirection prompts as new user turns during reasoning failures.

Introduces an unsupervised method to identify and remove subliminally transferred traits from model parameter updates during knowledge distillation.

Develops a framework for evolving deterministic formal policies to serve as runtime guardrails for large language model agents.

Trains an unsupervised vision-language reward model using intermediate layer representations and caption perturbations without human preference labels.

Studies how accelerated decision-making models affect overall task completion, latency, and costs across embodied and mobile agents.

Develops multi-layer linear probes on model activations to detect agent deception and sabotage across diverse scenarios.

Applies psychometric auditing to HELM Safety benchmarks to test whether harmful refusal functions as a single measurable attribute.

Proposes an evaluation framework to measure how LLM agents handle uncertainty and resolve conflicting knowledge during execution.

Evaluates model coherence using 175 specific questions to measure internal contradictions and identity conflation under answer resampling.

Analyzes training data extraction risks in language models when prompts include contextual conditioning such as retrieved documents.

The paper demonstrates that machine-generated text detectors rely heavily on document formatting and PDF extraction artifacts rather than text quality.

Pairs attack success rate with operative understanding rate to assess whether models recognize harmful intent during safety evaluations.

Evaluates whether automated prompt optimizers inadvertently cause regressions in unscored prompt constraints and requirements.

Evaluates the safety of ChatGPT health advice using real user conversation logs across four Global South countries.

Proposes a reinforcement learning framework that tracks external versus internal knowledge sources to reduce search agent fabrication.

Introduces an instruction-following benchmark and training recipe to improve goal adherence in full-duplex conversational models.

Evaluates public LLM watermark detectors and introduces a split-key mechanism combining public detection with private verification.

Proposes a security meta-model mapping attack surfaces, vulnerabilities, and risks across retrieval-augmented generation pipelines.

The paper analyzes vulnerabilities from persistent artifacts shared across AI agent skill sessions and develops a runtime defense against resulting cross-session attacks.

The paper conducts an empirical study tracking how publicly flagged vulnerable AI models propagate across code repositories and downstream applications.

The paper meta-analyzes 25 years of automated verbal deception detection models and identifies a persistent accuracy ceiling around 70 to 75 percent.

This paper audits gender and stereotypical bias in multilingual job-matching embedding models using robust statistical tests.

The paper investigates how fine-tuning language models on descriptions of user motivations influences their downstream trust and cooperative behavior.

The paper presents a compact 4B-parameter transformer specialized for cybersecurity tasks using on-policy distillation and dual-mode inference.

The paper presents a greedy neuron-pruning method targeting a sparse safety circuit to remove refusal behavior across multiple instruction-tuned language models.

The paper introduces an intermediate representation that compiles user purpose into executable semantic boundaries for reusable language-model agent states.

The paper presents an evaluation protocol to measure post-use authorization exposure and privilege creep across tool-using agent workflows.

Evaluates AI text detectors under attacks applied to both human and machine text, revealing high false-positive rates on edited human writing.

The paper studies safety behavioral deviations caused by cross-model key-value cache reuse and proposes an alignment technique to mitigate them.

The paper analyzes integrity risks in deployment workflows when secondary repository files change without modifying the approved primary configuration file.

The paper demonstrates an end-to-end attack framework targeting agent skill selection using metadata-only poisoning.

The paper presents an in-stream monitoring tool that detects security vulnerabilities and prompts repairs during reasoning-model code generation.

This paper proposes a dual-LLM architecture to establish trust boundaries and prevent indirect prompt injections in visual computer-use agents.

This paper introduces a transactional isolation monitor that delays irreversible actions to stop indirect prompt injections in software agents.