Tencent's GistRank compresses passages to 4 gist tokens for cheap LLM listwise reranking
_reachsumit · x · 2026-10-30
Tencent's GistRank paper studies soft compression for LLM listwise ranking: each candidate passage is compressed into 4 learned gist tokens so the LLM reranks from a short prompt, cutting the linear input growth of verbatim serialization. They find vanilla gist representations fail for ranking (causal encoding weakens absorption; no relative-preference structure), and propose a three-stage framework — semi-bidirectional gist learning, pairwise discriminative training, listwise adaptation — staying close to full-text reranking quality at much lower latency.
More from Research
- UChicago seminar: hippocampal ripples, visual imagery BCIs, minimally invasive neural interfaces — plopesresearch · 2026-10-30
- Oopsie-data: Building the first open dataset of robot failure trajectories — chris_j_paxton · 2026-10-30
- Schmidhuber's 'Miraculous Year 1990-91': claiming roots of GANs, Transformers, pre-training — SchmidhuberAI · 2026-10-30
- NBER opens applications for Stanford Digital Economics and AI tutorial, due Nov 3 — avicgoldfarb · 2026-10-30
- FlexUMM shows unified multimodal models aren't so unified in embedding space — jindong_wang92 · 2026-10-30
- Data leakage hit 648 papers across 30 fields; autoarxiv now auto-reproduces arXiv papers — pandeyparul · 2026-10-30