Alibaba's PFA progressively focuses attention to speed long-sequence generative recommendation
_reachsumit · x · 2026-10-30
Alibaba researchers observe that attention in generative-recommendation beam search progressively focuses on a small set of critical history tokens whose importance ranking stays stable across decoding steps. Their PFA framework keeps full attention at the first step, builds a shared importance ranking via cross-beam weighted aggregation, reorders KV cache into contiguous memory, then geometrically shrinks the token set — avoiding the semantic loss of static compression and the overhead of repeated top-k selection.
More from Research
- OpenFlyV2: a 1B onboard VLA model for real-time aerial drone navigation — IPEC-COMMUNITY · 2026-10-30
- Prime gap bound shrinks from 70M to 212 since Zhang's breakthrough; twin primes needs 2 — IgorCarron · 2026-10-30
- Researcher: concept neurons alone may not be enough to induce an AI persona — Sauers_ · 2026-10-30
- Sibling model neuron sharing: probing shared representations across related models — Sauers_ · 2026-10-30
- Porting PCRE2 to Verified Rust: Fuzz Aggressively Using the C Version as an Oracle — jd_pressman · 2026-10-30
- UCLouvain Releases Full Video Lectures: Stochastic Optimization and Energy Economics — caglar_ee · 2026-10-30