Kuaishao's PSG: Generative Reranking Decodes Ordered Item Pairs, Halving Steps Losslessly
_reachsumit · x · 2026-07-30
Kuaishou's team proposed PSG (Pair-Space Generation), a new method for generative reranking to address the high complexity and cumulative errors of Autoregressive (AR) models in list generation.
- Core Innovation: PSG elevates the generation atom from individual items to ordered item pairs. It produces pair token representations on-the-fly via a pretrained module, eliminating the data sparsity of a quadratic vocabulary.
- Theoretical Guarantees: The paper proves PSG is bijective with item-space generation and induces an equivalent family of sequence distributions, incurring no loss of expressiveness.
- Performance: It achieves a theoretical 2x to 4x speedup under moderate settings, effectively halving decoding steps and delivering a 1.83x inference speedup in practice.
More from Research
- Study: Agentic Harnesses Significantly Improve LLM Compositional Generalization — lateinteraction · 2026-07-30
- Stanford's Free CS336 Course: Building LLMs from Scratch to Become an AI Architect — techNmak · 2026-07-30
- Handroid: A Reconfigurable Robot Shifting Between Humanoid and Dexterous Hand — carlosdponx · 2026-07-30
- RATs: Multi-Agent Robot Lifelong Skill Learning Without Gradients or RL — rsasaki0109 · 2026-07-30
- Cisco's Open-Source Security Model Antares-3B Nears GPT-5.5 at Fraction of Cost — joshua_saxe · 2026-07-30
- DeepMind's Andrew Trask: LLMs Are Like Libraries Without an Index — iamtrask · 2026-07-30