What happens when we patch the KV cache from a finetune into the non-finetuned base LLM?
In our new paper, we find that some behaviors, like answering in German or emergent misalignment, transfer with the KV cache and depend on it, while others don't. ๐งต
What happens when we patch the KV cache from a finetune into the non-finetuned base LLM?
In our new paper, we find that some behaviors, like answering in German or emergent misalignment, transfer with the KV cache and depend on it, while others don't. ๐งต
@ARTartaglini@Mikey_Lepori Nice animation!
Reminds me of the cossim stuff in diffusion models I was working with 2y ago, although the question is completely different.
https://t.co/3rKMsrjOmn
We analyzed image diffusion models like Stable Diffusion and found positional embedding and anomalies in the U-net representation
Will present at #ICLR2025!
@whensdeyi@Sauers_@_lyraaaa_@treetowntree Do you already steer during prefill or only during answer generation?
Prefill sounds like the more "natural" setup, but would probably be much more expensive as it'd break caching.
@Chinese_XU@locknessko@KBlueleaf@SpaceWelder314 Yes, this is true if we don't batch the decoding computations. But if we batch it (either by processing multiple answers at once, or by predicting multiple tokens via speculative decoding), compute can become a bottleneck again
@locknessko@KBlueleaf@SpaceWelder314 Yeah, but doesnt one usually batch decoding via parallel requests or speculative decoding until one hits the compute bounds again?