FrontierCS 2 benchmark makes agents beat original methods from code with the paper's contribution deleted
berkeley_ai · x · 2026-10-24
Ion Stoica boosted the release of FrontierCS 2, a new benchmark for CS research & development by Qiuyang Mang. Each task is built from a research paper: the paper's contribution is deleted from the released code, and agents must outdo the original method—judged by the authors' chosen metrics—without access to the paper or the web.
Related event: FrontierCS 2 Launches: Top LLM Agents Pass Only About Half of CS R&D Tasks(7 posts)→
More from coding & agent
- Factory tested Reflection's Beam early via Together AI on real coding and agentic workflows — togethercompute · 2026-10-24
- Seroter's Reading List: CLIs Save Tokens, Agent Harness Explained, When to Use Diffusion LLMs — rseroter · 2026-10-24
- Dev's workflow: pairing Grok bot with open-source Hermes Agent side by side — alexcovo_eth · 2026-10-24
- Anthropic's Claude Duet: a proactive agent that keeps working after you close the app — edwinarbus · 2026-10-24
- Workato's Google Workspace MCP uses one tool, 1,000+ AI-reviewed self-updating actions — adamse · 2026-10-24
- Infinity: a Rust agent runtime that fits 75k agents in a Raspberry Pi's RAM — ShadajL · 2026-10-24