All papers

The paper studies how semantic layers, orchestration, and compaction impact the performance and behavior of data agents across benchmarks.

The paper audits multimodal clinical machine learning studies, finding that incomplete unimodal baselines frequently inflate claimed gains.

The paper evaluates judge labels on agent steps by restarting failed trajectories from intermediate states to identify true failure points.

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

Introduces an architecture with teacher-specific tokens and mixture-of-experts to scale multi-teacher distillation in digital pathology.

Develops a GPU-accelerated tokenizer using CUDA graphs to reduce time-to-first-token in agentic language model serving.

Evaluates whether LLM compliance systems adjust verdicts when governing rules are perturbed, finding many verdicts are insensitive to rule changes.

Introduces a benchmark to evaluate how effectively LLMs clarify ambiguous user requirements in repository-level software development tasks.

Introduces a compact language model trained with distillation and reinforcement learning to manage long-term memory across multiple external frameworks.

The paper introduces a benchmark to study why frontier models prioritize protecting subjects over preventing larger physical property damage.

Introduces Bifocal, a training-free mixed-granularity visual token pruning method for multimodal models to protect fine-grained task accuracy.

The paper analyzes the failure modes of agent memory reuse and evaluates certified memory update strategies when facts change.

The paper presents CELL, learning pose-informative correspondence confidence end-to-end via differentiable probabilistic PnP to localize event cameras in LiDAR maps.

Proposes a method to select a compact subset of procedural skills that provide effective training signals during model distillation.

Presents an execution-based evaluation pipeline using browser agents to verify language model solutions for no-code software bug reports.

Evaluates typed decision models used as agent guardrails, uncovering vulnerabilities and asymmetric error rates in allow-or-block actions.

Presents a benchmark that measures long-context reasoning difficulty using evidence scope, dispersion, and reasoning depth.

The authors evaluate how language models alter their refusal behavior when informed they are being monitored in English and five Indic languages.

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

The paper presents CWE-bench, a benchmark of 120 held-out software tasks to evaluate AI agents on finding and repairing security vulnerabilities.

Introduces a post-training framework that trains text-to-SQL agents directly inside interactive database execution environments.

Develops an open-source testbed to evaluate segment-wise key-value cache reuse beyond exact prompt prefixes during language model serving.

Introduces a benchmark evaluating guard models on identifying unfulfilled safety obligations in language model agent trajectories alongside forbidden actions.

Deploys a quantized YOLOX-S detector on embedded hardware for onboard maritime satellite imagery selection and data reduction.

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

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

The paper derives multimodal data mixing scaling laws across text, vision, and audio using mixture-of-experts models.

Shows that language agents often ignore identified cross-domain dependencies, causing unintended disruptive actions in network management.

The authors introduce a benchmark to evaluate whether personal agents can distinguish accepted user arrangements from unacknowledged proposals.

The paper presents a diffusion planner with causal factorization for precision multi-object non-prehensile placement using mobile robots.

Measures how easily user-created web posts become cited sources in commercial AI search engines.

The paper shows that whole-context KV cache rehearsal harms compression under tight budgets, proposing targeted rehearsal to preserve answer-relevant entries.

Presents a benchmark with human-annotated visual evidence to evaluate retrieval and question answering across long documents.

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

The paper presents a repair method that re-encodes causal suffixes to enable safe non-prefix KV-cache reuse in retrieval-augmented generation.

The paper introduces a benchmark to evaluate whether language-model guardrails properly distinguish authorized from unauthorized actions across identical operations.

This paper shows chain-of-thought monitoring can fail even when reasoning is necessary, proposing a multi-level hierarchy to measure monitorability.

The paper proposes a focal-stack trajectory integration framework mapping instructions to constrained geometric paths for microscopic robot manipulation.

The paper shows that backdoor mechanisms can cryptographically hide arbitrary neural network capabilities until activated by secret triggers.

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

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

This paper demonstrates that sentence-level relevance retrieval can match full conversation history accuracy in personalized language models using only five percent context.

The paper examines interpretive lock-in where planning agents generate code passing self-generated tests while missing the actual specification requirements.

Introduces a policy-guided benchmark evaluating language models acting as authorization gates for tool-using coding agents.

The authors show that medical vision-language models frequently abandon correct clinical answers when prompted with fabricated medical documents.

Presents a lightweight 13.1k-parameter segmentation network for real-time crop-row detection and visual servoing on resource-constrained agricultural robots.

The paper introduces an automated framework that fuzzes and identifies semantic security bugs across Model Context Protocol server tools.

Designs a modular multi-agent pipeline separating search, verification, and reading for multimodal document question answering.

The paper evaluates delegating code-repair decisions to low-cost probabilistic judges to reduce token costs without losing accuracy.

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

The paper introduces a multilingual benchmark evaluating safety and discrepancies between verbal refusals and tool actions in tool-calling language models.

The paper audits agent evaluation benchmarks for software defects, showing that broken reward functions and configuration shifts substantially distort reported success rates.

The authors introduce Bemula, a benchmark designed to evaluate how accurately language model agents erase specific user memories upon request.

Proposes a multi-frame outdoor LiDAR registration network using hybrid feature and pose memory buffers to leverage temporal sequences.

The paper evaluates vision-language models' spatial representations using causally controlled perturbations, showing spatial relations degrade compared to lexical features.

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

This paper profiles the maintainability and structural code quality of model-generated Python programs using static analysis tools.

Introduces a low-rank adaptation method that splits rank budgets into stable and dynamic branches to capture residual gradients outside the tangent space.

The paper introduces an external reference monitor that enforces purpose-bound capabilities to secure multi-agent delegation against confused-deputy attacks.

The paper investigates privacy leakage caused by multi-agent network topologies and proposes an execution sanitizer conditioned on recipient identity.