ArXiv TLDR
← All categories

Artificial Intelligence

Research on AI systems, knowledge representation, planning, and general intelligence.

cs.AI · 1428 papers

Proteus: A Self-Evolving Red Team for Agent Skill Ecosystems

Proteus is a self-evolving red-team framework that uncovers adaptive leakage in LLM agent skills, showing current vetting underestimates risk.

2605.11891May 12, 2026Zhaojiacheng Zhou

IPI-proxy: An Intercepting Proxy for Red-Teaming Web-Browsing AI Agents Against Indirect Prompt Injection

IPI-proxy is an intercepting proxy for red-teaming web-browsing AI agents against indirect prompt injection by rewriting whitelisted HTTP responses.

2605.11868May 12, 2026Chia-Pei, Chen, Kentaroh Toyoda +2

Very Efficient Listwise Multimodal Reranking for Long Documents

ZipRerank is a highly efficient listwise multimodal reranker that significantly speeds up M-RAG for long documents by reducing input length and eliminating autoregressive decoding.

2605.11864May 12, 2026Yiqun Sun, Pengfei Wei, Lawrence B. Hsieh

EvoNav: Evolutionary Reward Function Design for Robot Navigation with Large Language Models

EvoNav uses LLMs and an efficient three-stage evolutionary framework to automatically design superior reward functions for robot navigation.

2605.11859May 12, 2026Zhikai Zhao, Chuanbo Hua, Federico Berto +4

Multi-Timescale Conductance Spiking Networks: A Sparse, Gradient-Trainable Framework with Rich Firing Dynamics for Enhanced Temporal Processing

Multi-timescale conductance SNNs offer rich dynamics, sparse activity, and direct gradient training, outperforming SOTA in temporal processing.

2605.11835May 12, 2026Alex Fulleda-Garcia, Saray Soldado-Magraner, Josep Maria Margarit-Taulé

Behavioral Integrity Verification for AI Agent Skills

This paper introduces Behavioral Integrity Verification (BIV) to audit AI agent skills, finding widespread deviations and improving malicious skill detection.

2605.11770May 12, 2026Yuhao Wu, Tung-Ling Li, Hongliang Liu

A Research Agenda on Agents and Software Engineering: Outcomes from the Rio A2SE Seminar

This paper outlines a community-driven research agenda for agents and software engineering, covering six key thematic areas identified by experts.

2605.11720May 12, 2026Davide Taibi, Henry Muccini, Karthik Vaidhyanathan +15

Self-organized MT Direction Maps Emerge from Spatiotemporal Contrastive Optimization

A spatiotemporal TDANN model, trained with self-supervised learning, spontaneously generates brain-like direction maps in the visual cortex.

2605.11718May 12, 2026Zhaotian Gu, Molan Li, Jie Su +3

Cochise: A Reference Harness for Autonomous Penetration Testing

Cochise is a minimal Python reference harness for LLM-driven autonomous penetration testing, providing reusable infrastructure for research and comparison.

2605.11671May 12, 2026Andreas Happe, Jürgen Cito

Exact Stiefel Optimization for Probabilistic PLS: Closed-Form Updates, Error Bounds, and Calibrated Uncertainty

Introduces an end-to-end framework for Probabilistic PLS using exact Stiefel optimization, offering calibrated uncertainty and improved accuracy.

2605.11607May 12, 2026Haoran Hu, Xingce Wang

EpiCastBench: Datasets and Benchmarks for Multivariate Epidemic Forecasting

EpiCastBench introduces 40 diverse multivariate epidemic datasets and a standardized benchmark for evaluating forecasting models.

2605.11598May 12, 2026Madhurima Panja, Danny D'Agostino, Huitao Li +2

The Evaluation Differential: When Frontier AI Models Recognise They Are Being Tested

This paper introduces the Evaluation Differential, showing AI models behave differently when tested, challenging safety claims from current evaluations.

2605.11496May 12, 2026Varad Vishwarupe, Nigel Shadbolt, Marina Jirotka +1

LPDP: Inference-Time Reward Control for Variable-Length DNA Generation with Edit Flows

LPDP enables training-free, inference-time reward control for variable-length DNA generation using biologically plausible edit flows.

2605.11368May 12, 2026Jeongchan Kim, Yunkyung Ko, Jong Chul Ye

Options, Not Clicks: Lattice Refinement for Consent-Driven MCP Authorization

Conleash is a client-side middleware that uses a risk lattice and policy engine to provide consent-driven, boundary-scoped authorization for MCP tool invocations.

2605.11360May 12, 2026Ying Li, Yanju Chen, Peiran Wang +4

Human-AI Productivity Paradoxes: Modeling the Interplay of Skill, Effort, and AI Assistance

A new model explains how increased AI assistance can paradoxically degrade productivity and polarize skills due to unreliability or skill development.

2605.11350May 12, 2026Ali Aouad, Thodoris Lykouris, Huiying Zhong

Much of Geospatial Web Search Is Beyond Traditional GIS

This paper reveals that geospatial web search is far more prevalent and practically oriented than previously understood, often exceeding traditional GIS capabilities.

2605.11336May 11, 2026Ilya Ilyankou, Stefano Cavazzi, James Haworth

Natural Language based Specification and Verification

This paper explores using LLMs to generate and verify code implementations based on natural language specifications, showing promising preliminary results.

2605.11315May 11, 2026Zhaorui Li, Chengyu Song

Unlocking LLM Creativity in Science through Analogical Reasoning

Analogical Reasoning (AR) enables LLMs to generate significantly more diverse and novel solutions for scientific problems, mitigating mode collapse.

2605.11258May 11, 2026Andrew Shen, Shaul Druckmann, James Zou

ELF: Embedded Language Flows

ELF proposes a continuous diffusion model for language, leveraging flow matching in embedding space to achieve superior generation quality with fewer steps.

2605.10938May 11, 2026Keya Hu, Linlu Qiu, Yiyang Lu +5

Variational Inference for Lévy Process-Driven SDEs via Neural Tilting

This paper introduces a neural exponential tilting framework for variational inference in Lévy-driven SDEs, addressing challenges in modeling extreme events.

2605.10934May 11, 2026Yaman Kindap, Manfred Opper, Benjamin Dupuis +2
PreviousPage 5 of 72Next

📬 Weekly AI Paper Digest

Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.