🚨🚨 Off-peak PR!
6 papers at
#ICML2026! Here's the lineup 👇
🤖 OpenSage: Self-Programming Agent Generation Engine: The FIRST ADK that allows agents to build agents by themselves without human-defined workflows or topologies. Given a task, OpenSage allows agents to program, assemble, and evolve by themselves. Agents built by OpenSage have appeared at the top of multiple benchmarks at the time of release, including CyberGym, TerminalBench, and SWE-Bench-Pro. Recently, we constructed SageCTF with OpenSage, which outperformed more than 90% of the teams in DECCON CTF Qualification 2026, the most difficult CTF competition in the world.
🔗
opensage-agent.ai/
🧠 rePIRL: Learn PRM with Inverse RL for LLM Reasoning: A new exploration of intermediate rewards: learning process reward models with inverse RL — a principled route to dense, step-level signal that makes LLM reasoning stronger and trains more stable and more efficient.
🔗
arxiv.org/pdf/2602.07832
🛡️ BlueCodeAgentA: a blue-teaming agent powered by automated red teaming — turning attack knowledge into defense for safer CodeGen AI.
🔗
arxiv.org/abs/2510.18131
🌐 CyberCycle: Scalable Real-World Benchmark for AI Agents' End-to-End Cybersecurity Capabilities.
🔗
icml.cc/virtual/2026/poster/…
⚔️ Position: To Defend Against Cyber Attacks, We Must Teach AI Agents to Hack.
🔗
arxiv.org/abs/2602.02595
🔁 Position: AgentBeats: Agentifying Agent Assessment for Openness, Standardization, and Reproducibility: let agents evaluate agents.
🔗
arxiv.org/abs/2606.13608
#ICML2026 #AIAgents #LLMReasoning #AgentSecurity