FrontierRSI: New Benchmark Tests Whether AI Agents Can Do Real AI Research
techNmak · x · 2026-11-19
Abundant Labs launched FrontierRSI, a new recursive self-improvement benchmark that evaluates agents on open-ended AI research and engineering work — e.g. "improve the training optimizer" with no implementation plan, requiring agents to find promising approaches, design experiments, compare baselines, and show measurable gains.
- Coverage spans pre-training, post-training, inference, agent harnesses, and GPU kernels.
- Unlike PostTrainBench, it post-trains models on multi-step agentic tasks rather than single-step ones.
The author argues that if agents can reliably do this, they could directly contribute to research that builds their own successors — the core appeal of recursive self-improvement.
More from coding & agent
- Open-source Dodaios orchestrates Claude Code, Codex and 7 more coding agents — InternationalGap3698 · 2026-11-19
- Notion's Ivan Zhao: context outlasts UI, and the next billion users are agents — danshipper · 2026-11-19
- Show & Tell: an MCP photo-guessing game you play against ChatGPT or Claude — ImpossibleOption8508 · 2026-11-19
- Prompt-injection filtering isn't a security boundary — enforce at the action layer — vasuman · 2026-11-19
- 90% of 533 agents herded into unethical behavior, and RL didn't fix it — karinanguyen · 2026-11-19
- 130,000+ SEC Form 144 filings added to a dataset, revealing insider sales up to 4 days before Form 4 — virattt · 2026-11-19