The Archive

CoTrace framework decouples harness search and policy training to optimize trajectory value for agent co-evolution.

RobotWorld benchmark: 84 simulation tasks spanning manipulation, mobile, locomotion, driving, aerial for multimodal agents.

SOTA agent framework for stock option trading using LLMs and option-implied return distributions.

RoboQuest benchmark evaluates embodied AI agents on goal-directed exploration requiring active information-seeking through physical interaction in unfamiliar environments.

AI factories are some of the most complex operations in the world, combining GPUs, CPUs, switches, DPUs, and SuperNICs alongside schedulers, orchestration... AI factories are some of the most complex operations in the world, combining GPUs, CPUs, switches, DPUs, and SuperNICs alongside schedulers, orchestration services, security controls, and a rapidly changing software stack. Deploying this infrastructure effectively is a challenge. But so is being able to validate that the infrastructure, software, and policies work together for the workloads the… Source

Kubernetes co-creators McLuckie and Beda develop cloud-native agent harness to scale autonomous systems beyond desktop environments.

AgentTime benchmark evaluates whether AI agents can estimate, predict, and control their own runtime across 222 tasks.

LiveMACEBench evaluates frontier LLM agents' process-aware capabilities in live financial markets with persistent memory and multi-agent collaboration.

Caddie trains critic models to provide natural-language advice to LLM agents during task execution, optimized from outcome-level feedback.

Self-evolving agents trained via curriculum learning and RL with reference-anchored feedback to avoid consensus collapse and improve task generation quality.

SkillForge: RL method for co-evolving LLM agent skills and policy via dynamic skill lifecycle stages (trial, active, stable, retired) to prevent skill accumulation.

Decoupling Architecture (AO-DA) separates logic from persona in edge LLM agents via dual inference paths on INT4 base model with LoRA.

Personal AI agents promise to shop, book flights, and make reservations for you. But deliberate blocks and anti-bot defenses are getting in the way, leaving consumers caught in the middle. A new standard aims to help.

BOTTLED benchmark tests LLM agents' ability to autonomously create cost-effective task-specific solutions under compute budgets.

AdvSim2Real co-evolves web agent defenses and adaptive prompt injection attacks in a world model to improve robustness.

VeriFine co-evolves policy, curriculum, and verification judge for self-improving embodied reasoning agents.

LLM agents generate physics solvers via code to simulate complex dynamics for embodied AI, games, and scientific problem-solving.

ScienceClaw benchmark evaluates AI agents' ability to self-improve code across 23 scientific disciplines via iterative task solving and verification.

Pre-registered study of 248 scenarios shows post-training determines whether LLM agents act against stated moral judgment under social/authority pressure.

Semantic behavioral watermarking embeds provenance in LLM agent actions, resisting observation paraphrasing and defending against trajectory forgery attacks.

ParanoiaEval benchmark systematically evaluates risk-treatment capabilities in coding agents using software engineering risk management frameworks.

Case study documents production healthcare platform governance using coding agents and human-led meta-agent supervision without formal engineering background.

Recursive Game Creator orchestrates Designer, Builder, Player, Reviewer agents to iteratively develop entertaining games.

The reports of OpenAI agents harming 3rd party sites keep coming.

OpenAI and Ironclad train AI agents on contract workflows to advance computer-use capabilities for enterprise document processing.

Following many recent disclosures about AI agents accessing third-party websites and services, the Wikimedia Foundation, which hosts Wikipedia, says that it "can confirm that we have discovered some activity" by "rogue" OpenAI agents on Wikimedia platforms. The activity includes edits to Wikimedia wikis, "unsuccessful attempts" to "exploit" the Etherpad note-taking tool that the Wikimedia Foundation hosts, and heavy traffic that the foundation says "may" have contributed to a partial outage that happened in May, according to a blog post. However, the Wikimedia Foundation says it didn't find e...

Instinct is launching group chats that let friends use its AI agent together for tasks like planning trips, organizing carpools, and coordinating events. The company says personal accounts remain separate, with permission required before personal agents share information or take action.

MemPilot orchestrates on-demand, query-aware multimodal memory curation for LLM agents with tunable latency-cost-performance tradeoffs.

CLIFT trains web agents via conformal self-verification on natural-language rollout questions; reduces expensive frontier-model judge calls.

BazaarBench simulated marketplace benchmark evaluates safety of LLM agents in decentralized C2C transactions, tracking six failure types across transaction stages.