Weak-to-Strong Generalization for Reranking

Reference implementation for the DataConLA 2026 talk. Standard practice fine-tunes a reranker with cross-entropy against whatever weak label is available, but that objective's global minimum is the weak label itself, including everywhere it is wrong. This repo implements a confidence-aware loss with a scheduled self-consistency term the weak labeler does not control, and measures recovered capability separately from how confidently the model agrees with itself. Six stages: a candidate pool (BM25 + dense + random negatives, with a recall@100 guard); independent weak scores (BM25, bi-encoder, MiniLM cross-encoder, simulated debiased CTR); rank normalization with RRF fusion; label-free confidence signals (agreement, margin, list entropy) that gate unreliable candidates to exactly zero gradient; the loss itself, (1 - alpha) * reliability-weighted CE + alpha * sharpened self-CE with alpha ramped in after the basic task is learned; and evaluation on nDCG@10, Agreement@10, disagreement precision, and ECE. Five runnable failure-mode diagnostics cover collapse, correlated weak sources, pool recall ceilings, and head-query shift.