Distilling Graded Chunk Relevance Fixes Contextual Embedding Training

Perplexity identified that prior contextual embedding training wrongly treats all supporting context as negatives. Their fix distills graded chunk relevance from a query-aware compressor, scoring tokens and aggregating them at arbitrary chunk boundaries without relabeling.

2026-11-03 ~ 2026-11-03 · 2 related posts

• Prior contextual embedding training treats supporting-context chunks as negatives — antoine_chaffin · 2026-11-03
• Graded chunk relevance distilled from a query-aware context compressor, no re-labeling needed — antoine_chaffin · 2026-11-03