⚡ 5 h early
Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models
2 experts are actively discussing the implications.
Why this matches
Research & technical analysis reaction
1 attributable expert contribution
· Tim Kellogg
“this is one of my favorite papers, similar method. It convinced me a that LLMs weren’t merely autocompleting they showed that EVEN WHEN the answer to a math problem was present in the dataset, the agent didn’t reference it, but rather looked at problems wit…” evidence ↗
2 experts
2 communities
1 sources clustered
“this is one of my favorite papers, similar method. It convinced me a that LLMs weren’t merely autocompleting they showed that EVEN WHEN the answer to a math problem was present in the dataset, the agent didn’t reference it, but rather looked at problems wit…”
2 experts discussed this · 7 posts
Tim Kellogg: this is only stuff you can do on open models not just open weights, but fully open models with open datasets this stuff is extremely important, and we know next to nothing about it
Tim Kellogg: this is one of my favorite papers, similar method. It convinced me a that LLMs weren’t merely autocompleting they showed that EVEN WHEN the answer to a math problem was present in the dataset, the …
Tim Kellogg: that one paper got me off the AI skeptic route and pushed me a lot deeper into how LLMs actually work and again, without being able to cross reference the training dataset, methods like this simply…
Open the full discussion →