Proprioceptive AI开发的Cygnus技术，通过为冻结的大语言模型添加自感知适配器，使其能读取内部认知几何。该技术将模型的隐藏状态投影到由gl(4,R)李代数定义的数学空间，分离出包含主要精度信号的“暗模式”，从而无需重新训练即可显著提升模型性能。例如，仅用一张RTX 3090显卡，就将Qwen-32B在ARC-Challenge基准上的准确率从82.2%提升至94.97%。其适配器将覆盖从3B到405B的多款模型，服务节点可支持5万用户并发，预计本周末上线。相关设计论文已公开。
用Lie代数揪出模型内部的“黑暗模式”，在ARC上狂涨12个点还只要一张3090，如果真能泛化到其他任务，这或许是今年最巧妙的模型增强方案，但单基准提升仍需更多验证。
Frozen LLMs still carry readable behavior signals deep inside their hidden states.
And Proprioceptive AI has created Cygnus, that lets LLMs sense their own internal thinking patterns and dramatically improve accuracy.
This pushes Qwen-32B from 82.2% to 94.97% on ARC-Challenge using just one RTX 3090.
So Cygnus equips frozen LLMs with self-sensing adapters that read their internal cognitive geometry.
The adapters project hidden states into a mathematical space defined by gl(4,R) Lie algebra to isolate dark modes.
Those dark modes hold the majority of accuracy-relevant signals erased by standard normalization. This design leads to substantial benchmark gains without any model retraining.
Amazing how mathematical insights into activation geometry can improve reliability without full retraining.
They currently host up to 50,000 users concurrently on their droplet.
Final adverserial testing is wrapping up now on small models, GPU’s arriving daily, 1TB ram so far, 128 VRAM and adding. Will have adapters up to Nous Hermes 405 B from 3b Qwen and everything in between. Cluster is up and growing, product is all but complete. End user can download our app, connect to our API , no latency added to end user. Can currently host up to 50,000 users concurrently on our droplet. Should be available by the end of this week if ad testing goes well. Download will be available at https://www.proprioceptiveai.com/cygnus @grok what are your thoughts? Papers on Our Proprioceptive Adapters & Cygnus’ design are available https://proprioceptiveai.com/publications.html https://zenodo.org/records/19223882 - Cygnus Paper - 63 Pg. March - 26 2026’ https://zenodo.org/records/19080172 - 459 Pg. March - 17 2036’ https://zenodo.org/records/18471775 - Unified Behavioral Modification in “LLM’s” 33 Pg. Feb - 3 2026’ https://zenodo.org/records/18247585 - Holonomy Transformer - 36 pg. January 14 2026’ (First concepts - Fiber Theory, Mathematics , Open Code)在 X 查看被引用的帖子
来源：Rohan Paul · x.com