RecursiveLearner

[Do not include in judging - incomplete result as of deadline, just adding thought process] RecursiveLearner — training-free recursive self-improvement via test-time "lessons memory" The question: Can a small open model improve itself at inference time, with zero weight updates, by writing lessons from its own mistakes and retrieving them when it meets similar problems? And can it be fully self-contained — one model acting as solver, embedder (clustering its own failures using its own hidden-state activations), and lesson-writer? The approach: Run the model on a train split → collect what it gets wrong → cluster those failures by the model's own activations → have the model diagnose each cluster and write a short reusable strategy → at test time, route each new problem to its nearest failure-cluster and inject that lesson into the prompt. No fine-tuning — pure prompt-patching. (Proposer models + a model catalog served via DigitalOcean's inference API; activations + local inference on an RTX 3090; vLLM for throughput.) What we found: - IFEval (instruction-following): a clean positive. Self-derived lessons gave a statistically significant held-out gain (+5.5%, p=0.036), and more lessons helped monotonically (+4.4%, p=0.013). Genuine training-free self-improvement. - BBH (multi-step reasoning): On a fully-consistent, leakage-checked 3-shot protocol (n=675, Qwen2.5-7B), failure-derived self-lessons are neutral-to-marginally-positive on BBH reasoning (strict Δ 0.0; format-agnostic Δ +1.5%, net-positive on ~half the 27 tasks), and a clear win on instruction-following (IFEval +5.5% held-out, p=0.036).

Team

Placement

Participant