Generative reranking explained: LLMs that emit rankings as tokens, not scores
dejanseo · x · 2026-09-07
- dejan.ai adds a new concept entry on Generative Reranking: an LLM produces the ranking itself by emitting candidate identifiers as tokens in ranked order, instead of scoring each candidate with a separate model.
- The paradigm was established by RankGPT (Sun et al., 2023), which reranked passages listwise with a sliding window for over-length lists; BlockRank calls this in-context ranking, distinct from generative self-retrieval.
- Cost sits in decoding: N candidates need at least N sequential steps, and free-text output can repeat or omit identifiers, producing invalid permutations.
- Fixing decode: FIRST (Reddy et al., 2024) reads the ranking from the first token's logits; hLLM (Laftchiev et al., 2026) builds an item-by-position score matrix from prefill hidden states and decodes a valid permutation via the Hungarian algorithm.
More from Research
- Stanford's Christopher Potts publishes draft rebutting skeptical views of interpretability research — aryaman2020 · 2026-09-07
- Interp researcher pushes back: probes have advanced well beyond pre-LLM-era techniques — aryaman2020 · 2026-09-07
- AI job impact map: web developers 94% substitutable, teachers only 29%, across 798 US occupations — alvelda · 2026-09-07
- RLSlow team credited with inventing RL at scale for LLMs — morqon · 2026-09-07
- UniSim-SLAM: feed-forward SLAM with unified Sim(3) optimization lands at ECCV 2026 — rsasaki0109 · 2026-09-07
- HCOMP + CI 2026 draft program is live, registration now open — windx0303 · 2026-09-07