Alibaba's Recursa Pushes Recursive Self-Improvement, Lifting Firm Benchmark Pass Rate to 55%
alibabagroup · hf · 2026-11-04
Alibaba researchers present Recursa, a foundation model system that adapts to new domains by recursively learning from its own experience, without human experts or stronger teacher models.
Key ideas:
- The system comprises model, instructions, skills, memory, and a harness — domain adaptation means learning the whole system, not just the model
- A knowledge repository stores domain skills, when to apply them, and reflections on past tasks
- Two gaps: knowledge gap (missing terminology/where materials live) closed by studying domain materials; behavior gap (having information but using it wrong) closed by working tasks and reflecting on feedback
- Parallel learners execute tasks and synthesize local improvements, validated and merged into the repository for recursive self-improvement
On Harvey LAB's Firm benchmark, Recursa raises the full-pass rate of DeepSeek-V4.1-Flash from 30.0% to 55.0%, and from 25.4% to 46.0% on another model, targeting knowledge-heavy domains like law and finance.
More from AGI Musings
- Ramez Naam: power concentration is the biggest AI risk, plural AI makes us safer — sebkrier · 2026-11-04
- Sequoia summit: the diffusion gap between model progress and enterprise adoption is the opportunity — FinanceYF5 · 2026-11-04
- Sequoia summit: ease of use is underrated; build around customers, not model releases — FinanceYF5 · 2026-11-04
- Sequoia's advice to AI founders: chase customers, not model launches — FinanceYF5 · 2026-11-04
- Sequoia: AI shifts competition from systems of record to selling outcomes — FinanceYF5 · 2026-11-04
- Sequoia: next-gen AI apps must deliver 10-40x gains, not 10-40% — FinanceYF5 · 2026-11-04