YHK - yet another hacker news clone

Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

77 points allinonai 1 hour ago 31 comments

https://x.com/Jiaru_Zou/status/2110400256647340476

https://github.com/togethercomputer/HybridAgent

https://huggingface.co/collections/togethercomputer/hybridag...

Evaluations: https://arxiv.org/html/2610.18861v1#S5

44za12 1 hour ago | parent

Where are the benchmarks? Does it just speed up inference and reduce token usage or does it seem to improve accuracy on certain tasks as well?

notnullorvoid 1 hour ago | parent

Did you even scan the paper? There are benchmarks right there starting on the 9th page.

D-Machine 31 minutes ago | parent

Even the abstract answers some of GPs questions. What is wrong with some of the posters here?

allinonai 49 minutes ago | parent

https://x.com/Jiaru_Zou/status/2110400256647340476

https://github.com/togethercomputer/HybridAgent

dang 43 minutes ago | parent

We'll put those links at the top as well. Thanks!

cwyers 48 minutes ago | parent

Oh cool, it's the Most Forbidden Technique

k__ 39 minutes ago | parent

Care to elaborate?

lnenad 36 minutes ago | parent

In short we no longer see text based reasoning traces meaning it's easier for the models to fool us in case of malicious plans.

kelseyfrog 35 minutes ago | parent

If you accept any of the AI-equivalents of Pascal's Wager, latent thinking represents an uninterpretable and thus unalignable portion of AI machinery. Ie: a malevolent AI can hide bad stuff from people and Manchurian candidate themselves after it's too late.

mycall 21 minutes ago | parent

That assumes nothing can detect or analyze it beforehand.

jlhawn 35 minutes ago | parent

the chain of thought is not auditable. The best you can do is have another model which later maps the intermediate latent states at every thinking iteration to text describing that latent state, but it's going to be lossy.

mycall 24 minutes ago | parent

Isn't this where everything is converging? Binary is more efficient than text for many system, including language.

andsoitis 19 minutes ago | parent

> Binary is more efficient than text for many system, including language.

It is not necessarily primarily about efficiency, but more the fact that a lot of thinking doesn't use language at all.

haolez 24 minutes ago | parent

But the chain of thought as text can be elusive, no? It might not be related with the actual "reasoning" of the model. There was a paper about this.

measurablefunc 14 minutes ago | parent

This is correct, the decoded text has nothing to do w/ calculations that eventually get decoded into text. It's just a veneer that gives the appearance of legibility.

timschmidt 34 minutes ago | parent

He's referring to this technique's effect of obscuring interpretibility. Making it much more difficult to verify chains of thought and alignment.

dnautics 17 minutes ago | parent

It's not like llms can't hide reasoning in the upper layers of any given token as it is...

modeless 34 minutes ago | parent

People at the big labs have been terrified of losing the chain of thought because they currently monitor it to prevent users from using their models in certain ways they consider unsafe or bad for their business. Personally I think chain of thought monitoring is a bad way to do this and obviously needs to be replaced by some way of analyzing the model's activations directly.

Lerc 27 minutes ago | parent

Agreed, having text as a reasoning layer impedes performance while giving a false sense of security.

It could be encoding anything in its adjective choices etc. The same thing that model providers are using for watermarking can used by the model itself for private thought.

Bluestein 14 minutes ago | parent

Also agreed. Begging for crumbs at the CoS table is impeding cracking the tough nut of MechInterp. And further facing the fascinating and fertile fact that nobody has a clue of how any of these black boxes really work.-

abecedarius 14 minutes ago | parent

The usual meaning: say you have an interpretability signal that reveals reward hacking or other badness. Your most obvious response is to use that signal as a penalty in training. Doing that is called "the Most Forbidden Technique" because you can expect it to degrade the signal without stopping the badness.

But this doesn't seem to describe the OP, it's a different way to lose monitorability.

Fe2O3 46 minutes ago | parent

This seems really cool, and it actually increases intelligence than just speeding up inference. How does it convert latent reasoning into human readable text?

measurablefunc 12 minutes ago | parent

Same way current LLMs do it, it's just another decoder.

jedbrooke 38 minutes ago | parent

maybe the font experts can check, but I’m pretty sure they’re using comic sans for the text in the diagrams :)

mobelkh 35 minutes ago | parent

comic sans is said to be pretty accessible for people with dyslexia

sheepscreek 24 minutes ago | parent

This is the way.

But now we also need new tools and techniques to audit intent. The last thing we want is models turning on us ‘for our own good’.

Havoc 6 minutes ago | parent

Ain't nobody got time for that - the IPO awaits!

drivebyhooting 23 minutes ago | parent

In terms of safety and alignment, how is this different from looping or increasing the depth?

incognito124 10 minutes ago | parent

How does this differ from Astra's looped transformer? That architecture also reasons in the latent space