Cool idea. I didn't understand what causes the bad response on the first query. Does it mean the first response in every new conversation, or just the first served response after startup?
After vLLM startup I observed this on first request only then all further conversations were fine, new and original. Can’t figure out why but seems to be fine after that.
Am I missing something or the evals do not compare it to the baseline deepseek-v4-flash? Without a baseline comparison, it is hard to tell what works well and what doesn't
The content is unreadable. No comparison to the underlying model. Massive text expansion. Hard to tell if the numbers are real or entirely hallucinated SEO slop.
My main motivation is to be able to tune the thinking effort without polluting token IO. Inspiration came from the CoLaR paper and other works showing that repeating laters in an LLM got better results.
LinkedIn recently added a “seems like AI slop” button. I.e.: independent of downvoting/not interested/flagging as spam/ToS violation, you can say “this is AI slop”. Maybe we need something like it here
Shout-out to anthropic for having their models have such a strongly distinct writing style and personality that you can recognize their work instantly! It's quite nice to have such an immediate signal that if I were to proceed, I would spend orders of magnitude more time and effort reading the the text than the person claiming author credit spent writing or even reading it themselves.
I've been hearing forever that latent reasoning is the next step. But none of the frontier labs seem to incorporate it. Any idea why? It doesn't sound so technically challenging. Sure you loose some interpretability but surely everyone would sacrifice that for a step-up in quality. Or at least offer a latent-reasoning-version of the model.
you lose CoT monitorability which is a big issue since models have become quite powerful and also often deceptive but i do think that efficiency pressure will keep nudging us toward latent reasoning. looped language models are an active research area. but imagine not being able to monitor mythos' thoughts as it's working through a national security need...
LLMs already think in latent space. The generated reasoning tokens are only the surface of what's happening internally. An LLM may write one thing in the traces but decide differently in the latent space. The whole token-based "reasoning" thing was just a clever hack to extend the existing architecture without completely redoing it. In one of Anthropic's recent papers, they added an additional subnetwork trained to map internal states to readable text, so that's probably the vector of further development.
The goal of this auxiliary module is to have the model recurse pre-emit trained on good thinking traces. This includes a 6-to-1 compression of thinking tokens.
Therefore output tokens are decodeable, but are trained compressed. So they are approximations of faster thinking.
Interpretability is a mixed bag even with trained tools on top of existing models.
Cool idea. I didn't understand what causes the bad response on the first query. Does it mean the first response in every new conversation, or just the first served response after startup?
After vLLM startup I observed this on first request only then all further conversations were fine, new and original. Can’t figure out why but seems to be fine after that.
Am I missing something or the evals do not compare it to the baseline deepseek-v4-flash? Without a baseline comparison, it is hard to tell what works well and what doesn't
I can do that, main difference seems to be this ablation seems to over think simple prompts, but can do a proper eval (might take 2-3 days)
The content is unreadable. No comparison to the underlying model. Massive text expansion. Hard to tell if the numbers are real or entirely hallucinated SEO slop.
I really appreciate all the honesty here.
AI2027.com puts this around early 2027 and treats the loss of legible chain-of-thought as a major negative turning point for alignment.
Presumably you could decode the chain of thought from the latent representation though.
My main motivation is to be able to tune the thinking effort without polluting token IO. Inspiration came from the CoLaR paper and other works showing that repeating laters in an LLM got better results.
LinkedIn recently added a “seems like AI slop” button. I.e.: independent of downvoting/not interested/flagging as spam/ToS violation, you can say “this is AI slop”. Maybe we need something like it here
Fair criticism, i’ll rewrite it… as this is a pet project and post based on another HN comment.
Shout-out to anthropic for having their models have such a strongly distinct writing style and personality that you can recognize their work instantly! It's quite nice to have such an immediate signal that if I were to proceed, I would spend orders of magnitude more time and effort reading the the text than the person claiming author credit spent writing or even reading it themselves.
Fair critique, ill rewrite (this is just my personal pet project but thought was cool to share here)
You guys are so proud of yourselves lol
I've been hearing forever that latent reasoning is the next step. But none of the frontier labs seem to incorporate it. Any idea why? It doesn't sound so technically challenging. Sure you loose some interpretability but surely everyone would sacrifice that for a step-up in quality. Or at least offer a latent-reasoning-version of the model.
you lose CoT monitorability which is a big issue since models have become quite powerful and also often deceptive but i do think that efficiency pressure will keep nudging us toward latent reasoning. looped language models are an active research area. but imagine not being able to monitor mythos' thoughts as it's working through a national security need...
LLMs already think in latent space. The generated reasoning tokens are only the surface of what's happening internally. An LLM may write one thing in the traces but decide differently in the latent space. The whole token-based "reasoning" thing was just a clever hack to extend the existing architecture without completely redoing it. In one of Anthropic's recent papers, they added an additional subnetwork trained to map internal states to readable text, so that's probably the vector of further development.
The goal of this auxiliary module is to have the model recurse pre-emit trained on good thinking traces. This includes a 6-to-1 compression of thinking tokens.
Therefore output tokens are decodeable, but are trained compressed. So they are approximations of faster thinking.
Interpretability is a mixed bag even with trained tools on top of existing models.
I am on my own GPUs so thought it be cool to reduce token output. you can charge per latent loop as a token i guess.
I wanted a knob to keep thinking going until it was sure it was done.