All papers

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

The paper presents an adaptive, training-free speculative decoding method to speed up inference for vision-language models.

Introduces an explainability framework and weakly supervised video-text corpus to evaluate causal and temporal facial expression understanding in multimodal models.

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

The paper introduces a framework to model evolving social dynamics and update personas in multi-agent language model teams.

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

This paper introduces CACTUS, an ADMM-based optimizer that speeds up curvature-aware pruning and preserves influence functions in compressed language models.

Presents an embodied agent framework combining streaming spatial memory with active exploration for visual reasoning in physical environments.

The paper evaluates five language models on extracting complete interval networks of relative temporal events from narratives.

The paper presents a runtime system that structures long-horizon agent memory around reusable progress objects rather than raw interaction history.

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

This paper presents a flow matching method to distill diverse continuous latent reasoning distributions from multiple chain-of-thought rationales.

This paper introduces a vision-language model trained on paired microscopy image-text data for open-set biological image classification and retrieval.

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

The authors formulate a meta-game framework to dynamically route and select language models during strategic multi-agent interactions.

The paper proposes an interface for typed decision models that selectively invokes and interrupts reasoning to balance accuracy and latency.

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

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

This paper presents a knowledge distillation framework that reconstructs dense teacher visual tokens to train efficient compressed student vision-language models.

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

Investigates why looped transformers fail to retain or reuse earlier correct intermediate states when extended past training depth.

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

EgoProbe benchmarks long-range retention and event-specific recall under interference for streaming egocentric video models.

Introduces a benchmark evaluating how reliably software coding agents extract and digitize complex unstructured clinical trial protocol documents.

Proposes Web$hell, a unified interface combining broad information gathering, browser interactions, and evidence management for web-based agents.

The paper presents a tiny learned policy controller to optimize token unmasking and remasking schedules in diffusion language models.

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

Introduces a training-free open-vocabulary 3D segmentation approach that maintains per-primitive label distributions for clean object boundaries.

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

Presents an environment scaling framework that alters dependencies to scale agent benchmark difficulty while preserving task objectives.

The paper proposes filtering high-drift responses to mitigate off-policy degradation in mixture-of-experts reinforcement learning.

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

Introduces a corpus of vision-language model generated importance scores to improve temporal sampling in video action recognition.

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

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

The paper introduces a method that accelerates on-policy knowledge distillation by truncating student rollouts and applying inverse-probability weighting.

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

The paper introduces a training-free visual token pruning framework using information-density trees to speed up video large language model inference.

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

SHELDON introduces a dual-search reasoning framework that independently explores hypotheses and observation spaces to avoid confirmation bias.

Proposes a fractional power-of-two quantization grid to enable efficient multiplier-free integer inference for large language models.

Equips multimodal vision-language models with video world model operations as tools to support dynamic spatiotemporal visual reasoning.

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

The paper introduces a stateful commit-control framework with an evidence blackboard to manage revisions during agentic video reasoning.

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

The paper uses latent eigenvectors from CLIP to select smaller, representative training datasets for diffusion models.

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

The paper presents a training-free semantic feedback framework using vision-language models to iteratively detect and explain video anomalies.

This paper introduces an environment benchmark and exploration framework to help graphical user interface agents learn unfamiliar interaction dynamics.

Proposes a policy mechanism enabling language models to decide when, how, and which test-time training procedures to execute.

Studies how using UTF-16 byte order improves boundary prediction and representation efficiency in multilingual byte-level language models.

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

The paper proposes an adaptive image tokenizer that balances pixel reconstruction and semantic features to prevent discarding critical semantic content during compression.

This paper introduces a multi-agent sandbox benchmark to evaluate language models on state inference, forecasting, and intervention prediction.

The paper introduces a 30B tool-use model and an execution-grounded dataset to scale agentic tool supervision with verifiable environments.

The paper surveys memory-centric compression techniques for managing inference resources and context windows in large language models.

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

The paper develops an automated method that measures semantic saliency by evaluating how removing individual objects alters multimodal model scene descriptions.

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

Presents a benchmark evaluating agentic large language models on long-horizon strategic reasoning in the game Slay the Spire.