[Reading Group]: Position: LLMs can't jump

[nicaudinet]

📆 Date & Time

Monday 24th of August, 13:00-15:00

📍 Location or Link

CSE EDIT 8103

📄 Paper Link

https://openreview.net/pdf?id=klU4737opt

📖 Curator

Nikolai Ilinkh

🗒️ Agenda

We will discuss the position taken by the paper: that current AI systems (and LLMs in particular) are great at induction and deduction but are structurally incapable of abduction.

Please feel free to post some questions or comments below before the start of the meeting, and we will use them as a springboard to start the discussion.

See you there!

[nicaudinet]

Some questions:

• LLM coding agents can debug code. Isn't this proof of abductive reasoning?
• What are other good case studies other than Einstein and General Relativity?
• AI systems can be much faster and can scale much better than humans. Even if they are not as good as humans at abduction, they can brute-force a lot more. How good do they actually need to be to be able to compete with us?
• How could we design AI systems that are generally good at abduction? LeCun's JEPA maybe?
• Should we design systems that are good at abduction?

NeuroSymbolic perspective: LLMs as "intuition" systems. According to Kahneman, human intuition works well when

1. The environment is unchanging or slow to change. Complex adaptive systems are poor places to develop intuition.
2. We have a large sample size. That is, we get a lot of practice.
3. We receive immediate and accurate feedback.

[ricj]

I guess there's some similarity to the Bender & Koller argument? https://aclanthology.org/2020.acl-main.463.pdf

[nilinykh]

Some post-meeting comments/notes:

1. What do we mean by abduction?
2. LLMs perform inference to the best explanation: given several explanations/hypotheses, select the one that best accounts for the evidence from a set of explanations. Current LLMs might be good at this.
3. Creative/generative abduction: construct a new hypothesis or conceptualisation that is not simply one of the alternatives suggested within a particular framework. This can involve imagining possible worlds or reconceptualising the problem itself.
4. Perhaps, the relevant distinction is not induction vs deduction vs abduction, but between different kinds of abduction.
5. Can LLMs perform creative abduction? That is, generate truly novel hypotheses rather than selecting or recombining explanations already made salient by their existing knowledge (goes towards extrapolation and interpolation comparison).
6. Human creativity also recombines prior concepts, analogies, experiences, etc. So, what actually counts novel?
7. Scientific discovery is not necessarily "observations -> abduction -> discovery", but could be an iterative cycle with all three forms of reasoning. Abduction -> deduction -> induction/empirical evaluation -> revision -> abduction -> etc.
8. An agent does not necessarily need to spontaneously produce Einstein's insight in a moment. It might generate 1,000 weird hypotheses, deduce their implications, test them in simulations/experiments, discard 997, modify two, generate another hypothesis from the failures, etc.
9. Then scale can become relevant.