大模型只负责 prefill、小模型自己 decode，35B→4B 的成绩超过了 753B→4B 的闭源方案。但真正值得看的是他们自己泼的那盆冷水。
This is a AI post classified by AGK Intelligence as AI research (research), kept by the AI Radar because it carries real work, not commentary.
大模型只负责 prefill、小模型自己 decode，35B→4B 的成绩超过了 753B→4B 的闭源方案。但真正值得看的是他们自己泼的那盆冷水。 PLatour 的做法很轻：收发两端全部冻结，只训练零初始化的线性桥，把发送方若干层的隐状态按相对深度投影进接收方的残差流，之后 decode 完全不碰大模型；只用一个 GRPO 优化桥参数，不做对齐、不做蒸馏预热。支持混合注意力、16K 生成预算，参数比能拉到 100:1 以上。 数字上，35B→4B 在公开的数学和代码 benchmark 上超过了 Mostik 报告的 753B→4B；也扩到了 235B→1.7B 和 397B→4B 这种极端尺寸差。 但这篇最有价值的不是成绩，是 PACT-Check 这套对照：Dummy Transfer（发送方换成固定模板特征）、Irrelevant Transfer（换成无关问题的表示）、Self-Communication（让接收方自己当发送方）、Trained Receiver（直接对接收方做 LoRA 或全参微调），再加一个 PACT-Score，衡量相对最强对照的准确率余量。 结论不客气。一部分任务上，把发送方信号换掉、甚至直接训接收方，也能复现大部分提升；只有化学推理这类任务，PLatour 的增益超出所有对照。换句话说，光看 benchmark 涨分，证明不了「信息真
Posted by 青稞社区 (3.7k followers) 1 days ago · 36 likes · 2.6k views · view the original post on X. Kept by the AI Radar as AI research. Tools mentioned: platour.
More AI work like this
- alphaXiv 🤝 OpenResearch — @askalphaxiv
- No need for me to manually scan for new model releases anymore, since @paperswithcodee… — @NielsRogge
- We’re introducing ToyBlocks, an open simulated tabletop for controlled experiments in… — @allen_ai
- HotBench - Expert-authored tasks in medicinal chemistry, microbiology and dermatology to… — @SynBio1
- I wonder if you realize how interesting this is. — @SciTechera
- 🔥我靠！终于有大公司把AI模型后训练的“魔法”拆成可复现配方了！ — @NFT_Chen
- BREAKING: the best of 8 frontier models scored 37.5/100 on Science's Last Exam. On true… — @guilleflorvs
- Stanford dropped a 2-hour course on building LLMs from scratch — @RoundtableSpace
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 65.5k posts from 5k X accounts over the last 14 days, 10.4k tools, 19 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-06 16:50 UTC. Full method.