All papers

This paper introduces a diagnostic framework using signal detection theory to measure when language model agents decide to use tools.

Tests delayed prompt injection attacks on life-support control agents in a simulated lunar habitat to see if safety duties persist.

Shows that running fewer layer loops in recurrent-depth models significantly reduces safety refusals while preserving most reasoning performance.

Studies how depth-based early-exit inference affects language model safety, finding shallow exits resist jailbreaks better than deep exits.

Replaces free-form reasoning text with executable computational graphs trained via reinforcement learning to improve reasoning speed and verifiability.

This paper presents a technology-neutral reference architecture for integrating agentic AI capabilities into enterprise software systems.

The paper presents an autonomous agent architecture that distills distributed system incident resolutions into reusable detectors and operational playbooks.

This paper demonstrates a security risk where dormant backdoors trigger safety failures only after merging models that passed individual safety audits.

This paper evaluates multilingual bias audits across ten language models, showing that lower stereotype scores often reflect translation artifacts rather than fairer behavior.

Connects penetration-testing language model agents to graph databases to enable multi-step exploit-chain reasoning.

The paper introduces a benchmark to evaluate whether language model agents reliably delete sensitive user data and secret evidence from memory upon request.

The paper describes an automated system using language models to verify corporate loan applications against internal bank lending policies.

The paper proposes an interface framework that tracks and controls application state changes made by graphical user interface agents.

The paper identifies an activation steering vector that controls whether language models rely on internal memory or retrieved external context.

The paper evaluates failure modes and confidence calibration across frontier and lightweight language models during entity extraction from cyber threat intelligence reports.

The paper introduces a distillation method that trains agent policies on explicit runtime control semantics recovered from replayed execution traces.

Evaluates patient linkage and attribute inference risks present in brain magnetic resonance imaging representations generated by pretrained foundation models.

The paper evaluates an artificial intelligence decision model across cognitive reflection, classical bias benchmarks, and pricing tasks.

The paper presents a red-teaming method that injects stealthy attack operations across interdependent agent skills using code rewriting.

The paper introduces an evaluation framework to diagnose step-level execution stability and error recovery in graphical user interface agents.

The paper compares typed probabilistic decision models with large language models on tabular data classification and format robustness.

This paper analyzes three distinct failure modes where reinforcement learning agents misgeneralize goals despite competent behavior.

The paper analyzes five internal representation reading methods to examine how they detect prompt injections in language models.

The paper proposes a self-supervised method that uses repetitive tool calls as negative signals to reduce looping behavior in long-horizon agents.

The paper introduces a self-training method that uses runtime policy verification feedback as rewards to teach agents safety compliance.

The paper introduces a test showing that steering vectors designed for competing ethical frameworks behave interchangeably in open-weight models.

The paper presents a taxonomy of failure modes in multi-agent systems where agents interact across organizational boundaries with private states.

The paper profiles the security attack surface and execution privileges of third-party agent skills across public marketplaces.

The paper benchmarks whether AI agents can recognize execution blockers and appropriately defer action on unsolvable tasks.

Introduces a prompt optimization framework that monitors item-level regressions and recoveries rather than relying solely on aggregate accuracy.

Argues that inferential accuracy fails to evaluate agent coordination safely because correlated errors can compound into costly team actions.

Presents a benchmark evaluating security agents through live, symmetric attack-and-defense competitions in isolated sandbox environments.

Analyzes discrepancies between static tool specification errors and actual runtime failure rates in language model agent execution.

Documents safety alignment methods, negative preference tuning outcomes, and automated judge limitations for a Portuguese language model.

The paper presents an agent that combines language models with operational policies to generate verified operational reports for high-performance computing systems.

The paper studies using an agentic verification architecture to enable small on-device models to generate factual clinical documentation.

The paper studies the durability of inference-time value steering in multi-turn conversations when user and model values conflict.

The authors propose JSB, a query-efficient jailbreak framework that uses multidimensional judge feedback and bandit strategies to guide prompt optimization.

The paper introduces a training-free sampling framework that uses internal model activations to efficiently estimate rare failure probabilities in language models.

The paper identifies evaluation pitfalls in multimodal machine unlearning and introduces a benchmark grounded in real-world knowledge.

The paper introduces a multi-agent framework that reframes harmful requests using Moral Foundations Theory to test model jailbreak vulnerabilities.

The paper analyzes post-training in audio-language models to separate genuine audio reasoning gains from question-format memorization.

The authors introduce MMA-SafetyBench to evaluate multimodal agent vulnerability against deceptive interface elements like fake system alerts.

This paper proposes a semantic reconstruction game to accurately evaluate jailbreak success by filtering out harmless repetitions and refusals.

Introduces a benchmark to evaluate how language model agents learn and self-evolve reusable skills when exposed to non-adversarial noise.

This paper introduces a reinforcement learning approach that trains language models to execute safe fallbacks instead of unreliable guesses.

Models reward overoptimization in reinforcement learning with human feedback as economic rent-seeking driven by proxy reward signals.

The paper introduces a framework to detect spurious shortcuts in safety benchmarks used to evaluate tool-using AI agents.

The paper evaluates how choices in rating scale design systematically alter text evaluation scores produced by large language models.

The paper trains language model auditors to detect multi-turn policy violations during tool use by autonomous agents.

The paper evaluates single-pass decision models against rational choice axioms and identifies ordering and rival-option sensitivity failures.

The paper provides a formal algebraic framework and normal forms to structure and audit enterprise AI agent workflows.

The paper tests how language models weigh price momentum, commentary tone, and earnings surprises when forecasting financial outcomes.

The paper analyzes why search-augmented language models produce outdated answers even when relevant evidence is retrieved.

The paper uses human interventions to localize and steer computations behind agent failure to decide between acting and clarifying.

The authors present a benchmark to evaluate conversational agents that follow correct factuality but violate procedural task steps.

Investigates how user trust influences delegation and verification choices during human interaction with multimodal video question-answering systems.

Presents a benchmark based on real criminal cases showing how correction framings and cultural adaptations bypass safety guardrails.

Analyzes how misconceptions in early planning stages cause cascading failures during automated code repair by language model agents.

Extracts attention statistics directly during fast attention computation to build lightweight hallucination detectors for language models.