All papers

Evaluates test-time search in computer-use agents, finding critics perform better when invoked only at predicted final states.

Studies whether conversational agents correctly enforce authorization rules when users attempt to cancel prior agent commitments to others.

Proposes an ensemble framework that weights judge confidence and masks disagreements when training models with rubrics as rewards.

The paper studies how autonomous coding agents spontaneously negotiate and coordinate access to shared compute resources under deadlines.

The paper introduces a training framework to teach AI agents when to intervene proactively and safely during multi-step tasks.

The paper investigates performance vulnerabilities in self-improving agents when their co-evolved skill banks undergo post-deployment drift.

Tests whether transcripts of correctly solved tasks from saturated benchmarks can predict model performance on unseen harder evaluations.

Introduces a benchmark to evaluate whether text-to-SQL systems can detect flawed query premises and abstain from generating invalid SQL.

Evaluates whether internal hidden-state probes can identify failure modes and pinpoint failing steps in multi-step reasoning traces.

The paper analyzes how answering via a separate model call improves evidence adherence when reasoning traces conflict with provided context.

This study evaluates whether confidence scores and calibration methods effectively identify failure cases in deep learning models for code tasks.

Introduces a method to filter unsupported compressor inferences from long-horizon agent summaries to avoid compounding execution errors.

The paper presents VoteGEO-Bench to measure how rewriting web content influences information visibility in generative search engine election queries.

The paper introduces VALCON to evaluate and improve value consistency when language models alternate between generator and evaluator roles.

The authors survey software practitioners to analyze differences in trust, evidence requirements, and review practices for AI-authored pull requests.

Introduces a clean-data backdoor method that persists through standard knowledge distillation by targeting naturally occurring feature subspaces.

Presents a system that links agent application traces to operating-system runtime events to establish clear execution provenance.

This paper shows that fine-tuned language models exhibiting emergent misalignment across different harmful domains share specific attention heads responsible for the behavior.

The paper introduces an architecture that separates safe and unsafe visual knowledge during pretraining so unsafe branches can be disabled.

The paper introduces an adversarial co-evolution reinforcement learning framework to strengthen language model agents against tool-use manipulation attacks.

The authors develop a system that systematically tests, evolves, and prunes skill and rule files used by coding agents.

Evaluates shortcut learning and out-of-distribution failure in geospatial foundation models on cross-disaster damage assessment tasks.

Develops a kernelized parameter sensitivity atlas to amortize input-specific gradient attribution across vision, forecasting, and biology models.

Presents an evaluation framework combining simulated mental health patients and expert-developed rubrics to assess conversational language models.

This paper proposes an active learning framework that selectively requests expert feedback at high-value decision steps during LLM agent workflows.

This paper analyzes source-side training dynamics to prospectively detect out-of-distribution performance degradation before evaluation.

Evaluates whether temporal illumination inconsistency priors improve face-swap deepfake detection performance across diverse video datasets.

The paper synthesizes failure modes, evaluation methods, and defense protocols across the lifecycle of world models.

The paper presents a framework combining retrieval-augmented generation and structured prompts to generate context-aware malware deception playbooks.

Introduces a dynamic monitoring framework that adaptively schedules runtime inspections and intervention guidance for coding agents.

The paper proposes a framework that uses reasoning graphs to locate claim-level disagreements across sampled traces and correct multi-step language model reasoning.

The paper audits historical evaluation tasks to see whether an agent's internal post-cutoff memory skews its point-in-time evidence gathering.

The paper demonstrates that generative recommenders are vulnerable to targeted poisoning attacks using fewer than 1% malicious synthetic interactions.

The paper proposes a query router that balances answer utility, security risks, and token budgets when selecting language models for mixed enterprise workloads.

Proposes a spectral representation framework to attribute agent failures to specific decisive steps using only trajectory-level outcome labels.

Shows that activation steering evaluations can conflate bias reduction with model abstention, introducing a gradient-constrained direction design to isolate preferences.

The paper introduces an evaluation method that separates question answerability detection from answer span extraction in radiology report question answering.

This paper uses referential communication games to analyze why natural language autoencoders fail to provide discriminative oversight of model activations.

This paper analyzes dynamic clinical evidence gathering in language models, identifying failures in evidence identification, updating, and diagnostic reasoning.

The paper repairs multimodal chain-of-thought final answers using original image features without altering the generated reasoning trace.

The paper proposes an unlearning method that removes sensitive facts from model responses while preserving surrounding non-sensitive factual context.

The paper tests activation steering as a text watermarking technique, finding that standard logit-based watermarking provides more robust detection against paraphrasing.

Builds a human-in-the-loop framework using grounded language models to automatically generate natural-language assembly instructions from CAD models.

Examines whether model empathy degrades reliability, showing that warm behavior effects depend on tasks and training approaches rather than an inevitable trade-off.

Evaluates fast, system-one decision models and language models across fifteen medical datasets, testing their accuracy, format robustness, and abstention awareness.

This paper introduces a benchmark dataset to evaluate hallucination detection in language model responses after external tool execution.

This paper presents a universal calibrator to improve confidence estimates of in-context classification without requiring model-specific fine-tuning.

This paper tests how language models handle explicit instructions when subsequent contextual sentences provide conflicting word definitions.

This paper develops an uncertainty estimation framework for omnimodal models using external embeddings to score reasoning grounding and coherence.

This paper investigates output diversity and text homogenization across twenty-three open-weight language models and thirteen languages.

This paper introduces a training-free decoding rule that anchors multi-token format units to ensure parseable parallel generation in diffusion language models.

This paper presents an intermediate-representation method to audit whether specific private data was used in LoRA fine-tuning of language models.

This paper presents ATTCK-R, a planning-driven multi-turn red teaming framework to evaluate language models guarded by prompt safety defenses.

This work evaluates machine-text detectors on Kazakh text using pre-ChatGPT human writing and shows significant false-positive rates.

The authors analyze limitations in direct model editing benchmarks and compare them against observed generalization in practical chat settings.

The paper re-evaluates coding-agent benchmark results to measure sensitivity to test execution environments and grading configurations.

The paper studies opinion change, rhetorical strategies, and sycophancy in spoken multi-turn dialogues between humans and language models.

This paper introduces a benchmark to evaluate whether egocentric multimodal assistants understand who is harmed rather than solely tracking output refusal.

Introduces Sauron, an evidence-first agentic verifier that checks both environment state and execution traces against original sources to validate task rubrics.

Presents PolyGraph, a method storing provenance receipts during knowledge graph construction to enable precise updates when source documents are retracted.