New synthetic benchmark exposes two token-level flaws in training attribution
davidbau · x · 2026-10-31
davidbau introduces the first half of the same training-attribution study: the key trick is a new benchmark of novel synthetic data with unambiguous, configurable ground truth. It reveals two token-level problems — most tokens are unrelated to the query so document-level averaging buries the few that matter, and cross-entropy gradients shrink scores from high-confidence tokens, suppressing signal from well-learned facts. These motivated the neg-log-odds loss fix.
More from Research
- Researchers reconstruct Meta Muse's hidden charter from 884 chain-of-thought samples — sethlazar · 2026-10-31
- Working Paper Argues AI Can Break the Coasian Ceiling on Firm Size — zetalyrae · 2026-10-31
- Molence Bio Open-Sources Mol-1, Matching Closed SOTA at Antibody Structure Prediction — AllThingsApx · 2026-10-31
- Revisiting why transformers can't do zero-shot recursion, via Transformer Circuits math — akbirthko · 2026-10-31
- AI-assisted proof: first high-probability regret bounds for Feel-good Thompson sampling — QuanquanGu · 2026-10-31
- Vitória Piai's free psycholinguistics book Speaking joins ChapterPal's open-access library with an AI tutor — burkov · 2026-10-31