King's College London's PLatour framework exposes inflated gains in cross-model latent communication
KingsCollegeLondon · hf · 2026-11-06
Researchers from King's College London, Tencent Youtu Lab and NPL study PACT (Prefill-to-Autoregressive Cross-model Transfer), which pairs a large model's prompt representations with a small model's efficient decoding.
- They release PLatour, an open-source framework connecting frozen models via lightweight zero-initialized linear bridges trained directly with task rewards; it beats open-source baselines on generation tasks and surpasses Mostik's reported math/coding results with a much smaller sender.
- Key finding: controlled experiments show some gains can be reproduced by uninformative sender signals, self-communication, or receiver-only training — trainable bridges may act as mere parameter-efficient adaptation without exploiting sender information.
- They propose PACT-Check to quantify how much of a system's gain genuinely comes from communication.
More from Research
- 99% Accuracy Can Be Useless: Why Precision and Recall Beat Accuracy in Fraud Detection — Franc0Fernand0 · 2026-11-06
- Joakim Sundnes' Open Textbook 'Introduction to Scientific Programming with Python' Lands on ChapterPal — burkov · 2026-11-06
- moyix releases slides from Offensive Con Tokyo talk: Fuzzing vs. AI — moyix · 2026-11-06
- Benchmark Error Bars Too Wide: Data Underpowered to Prove Recent AI Speedup — tobyordoxford · 2026-11-06
- TabClustPFN tops 44 real tabular clustering benchmarks, moving clustering into pretraining — jiqizhixin · 2026-11-06
- Attention as spherical interpolation: values combine inside the unit sphere before norm projects back — mgostIH · 2026-11-06