Xiaomi's A-RSI lifts 11-benchmark average from 57.77 to 61.05 via adaptive task refresh
mark_k · x · 2026-11-01
Researchers from Xiaomi and several universities propose A-RSI, an adaptive self-improvement loop that monitors training progress and refreshes the task pool when the learning signal fades. Model-written programs generate multimodal questions with independently checkable answers, reducing reliance on the model grading itself. Reported result: average score up from 57.77 to 61.05 across 11 benchmarks using 800 fresh tasks, with no external training examples or human-written labels. Note: the gain applies to this additional training loop only, and results await independent replication.
Related event: Xiaomi and Universities Propose A-RSI for Recursive Self-Improvement(2 posts)→
More from Research
- awesome-ai-constitutions repo collects AI constitutions & model specs with generalization research — dhadfieldmenell · 2026-11-01
- Karpathy's nanochat hits 59K stars: a full LLM training lab in one repo — techNmak · 2026-11-01
- Same 2.4 eV band gap, different routes: SHAP explains per-prediction feature contributions — bravo_abad · 2026-11-01
- Hands-on tutorial trains a decision model on Qwen3-0.6B from scratch — dl_weekly · 2026-11-01
- From-scratch 8M-parameter LLM: dev blogs the whole training journey — the-random-guy-2002 · 2026-11-01
- Fine-grained MoE is the easiest win: DeepSeek-V3-style 128 experts cut loss by up to 6.2% — ivan_bezdomny · 2026-11-01