New short paper on the limits of one type of inference scaling, by @benediktstroebl.bsky.social @sayash.bsky.social & me. The first page has the main findings and message. (The title is a play on Inference Scaling Laws.) More on the limits of inference scaling coming soon. arxiv.org/abs/2411.17501 - ThreadSky

randomwalker.bsky.social • 118 days ago

New short paper on the limits of one type of inference scaling, by @benediktstroebl.bsky.social @sayash.bsky.social & me. The first page has the main findings and message. (The title is a play on Inference Scaling Laws.) More on the limits of inference scaling coming soon. https://arxiv.org/abs/2411.17501

Comments

mrdrozdov.com•118 days ago

Hi @randomwalker.bsky.social—I love this work! We published a preprint just last week showing a similar phenomenon w/ document reranking: https://arxiv.org/abs/2411.11767

Rerankers are similar architectures to reward models / verifiers, and we suspected the flaw applied generally. Great to see it does!

mrdrozdov.com•118 days ago

aboveearth.bsky.social•117 days ago

This was great. I’ve been putting a hybrid RAG system this into practice and both of these papers validate that our techniques are headed in the right direction. Thanks for sharing!

cschwem2er.bsky.social•118 days ago

In addition to the findings, what is especially interesting to me is the way you structured the first page, including fig. references in the abstract. Was this done based on the assumption that researchers increasingly only read abstracts? Or just to make the main findings as accessible as possible?

randomwalker.bsky.social•118 days ago

Thanks! It started out as a blog post and we expanded it into a paper instead, but we still had the mindset that it should be possible to get the key message across at a glance.

randomwalker.bsky.social•118 days ago

Many types of inference scaling techniques were already known to have limits. But resampling using verifiers seemed promising as a way to increase accuracy across many orders of magnitude of inference compute. The message of our paper is that this only works if the verifier is an oracle.

randomwalker.bsky.social•118 days ago

Note that the above table is about purely inference-time techniques and doesn't include o1-style learning-to-reason techniques. Still, our results suggest that if models are trained using feedback from imperfect verifiers, they might simply learn to exploit weaknesses of those verifiers.

randomwalker.bsky.social•118 days ago

We think inference scaling is promising, but only in a few domains that are amenable to reasoning, transfers poorly even among those domains, and that our efforts are better spent on improving domain-specific verifiers and composition techniques; domain-general inference scaling is fool's gold.

aboveearth.bsky.social•118 days ago

Ooh juicy! I’m going to have to go get my glasses and iPad for this one! Thx

Comments

Posting Rules

Reply