New paper: "Large Language Models and Emergence: A Complex Systems Perspective" (D. Krakauer, J. Krakauer, M. Mitchell). We look at claims of "emergent capabilities" & "emergent intelligence" in LLMs from the perspective of what emergence means in complexity science. arxiv.org/pdf/2506.11135 - ThreadSky

melaniemitchell.bsky.social • 3 days ago

New paper: "Large Language Models and Emergence: A Complex Systems Perspective" (D. Krakauer, J. Krakauer, M. Mitchell).

We look at claims of "emergent capabilities" & "emergent intelligence" in LLMs from the perspective of what emergence means in complexity science.

https://arxiv.org/pdf/2506.11135

Comments

malthusjohn.bsky.social•3 days ago

📌

davidduvenaud.bsky.social•3 days ago

"only those individuals who explicitly understood a task (via a natural language explanation) reached a correct solution whereas implicit trial and error reinforcement failed to converge. This ... has yet to be demonstrated in an LLM."

Is this claiming LLMs haven't been shown to benefit from hints?

melaniemitchell.bsky.social•2 days ago

It's not just benefiting from hints, but being able to understand complex tasks (or applying algorithms) from natural language.

⬇️

melaniemitchell.bsky.social•2 days ago

LLMs can do this to some degree but struggle more than humans, unless they are given many in-context examples (and even then sometimes), and definitely have problems executing algorithms given in prompts.

davidduvenaud.bsky.social•2 days ago

Thanks for explaining, but I'm still confused. LLMs succeed regularly at following complex natural-language instructions without examples - it's their bread and butter. I agree they sometimes have problems executing algorithms consistently (unless fine-tuned to do so), but so do untrained humans.

melaniemitchell.bsky.social•2 days ago

Many studies have found that if the instructions are for a task that is sufficiently different from the training data, they have trouble following the instructions in a way that seems (or is) unhumanlike.

E.g.,

(more in next reply)

melaniemitchell.bsky.social•2 days ago

https://aclanthology.org/2024.naacl-long.102/
https://openreview.net/pdf?id=t5cy5v9wph
https://arxiv.org/abs/2309.13638
https://machinelearning.apple.com/research/gsm-symbolic

(more in next reply)

melaniemitchell.bsky.social•2 days ago

But you make a fair point, and we will clarify what we mean in an updated version of the paper.

javaid89.bsky.social•2 days ago

I think that specific case is referring to human' ability to use language to learn concepts quickly. And not "hints" where you guide an LLM to an answer that converges after it's expended enough reasoning tokens. Or write code to solve a problem a normal human wouldn't need to use for said problem.

javaid89.bsky.social•2 days ago

I think the point is actually trying to understand the how LLMs vs. Humans solve ecologically valid problems. As opposed to calling your company Anthropic and just assuming your product is doing everything the same way a human does because it does better on benchmarks.

-Rapidly Disempowered human