Meta's hLLM decodes full rankings in O(1) passes with the Hungarian algorithm, 64x faster at 28ms

_reachsumit · x · 2026-09-03

Meta's hLLM (Hungarian LLM) tackles the decoding bottleneck in generative reranking: instead of autoregressively emitting N ordinals, it reads an N×K item-position score matrix off prefill hidden states with a lightweight attention head and solves optimal bipartite assignment via the Hungarian algorithm, yielding a valid permutation by construction. With LoRA fine-tuning and teacher-ranking distillation, it reaches 28ms end-to-end inference—a 64x speedup—at teacher-level ranking quality, with full ablations.

Original post →

More from Research

Research channel →