Connectionists: new directions in evolution strategy fine-tuning

[Risto Miikkulainen]

Evolution strategies is an alternative to gradient descent in fine-tuning LLMs. Cognizant AI Lab and collaborators recently completed four new papers on it, evaluating it in several new tasks including math and ARC-AGI, extending it to fine-tuning quantized models, demonstrating how LLMs can be trained to know what they know, and showing why high dimensionality of LLMs is a blessing (not a curse) in fine tuning. For more details, see

Overview blog: https://cgnz.at/6005QZNMb
Papers: https://cgnz.at/6007QZN13 (tasks)
        https://cgnz.at/6003QZNHN (quantized)
        https://cgnz.at/6002QZNGG (metacognition)
        https://cgnz.at/6000QZNyI (blessing)
Animated illustration: https://x.com/RobertoDailey1/status/2026791418380087442

-- Risto

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20260225/32bdb58b/attachment.html>