Kayo Yin
@kayoyin.bsky.social
PhD student at UC Berkeley. NLP for signed languages and LLM interpretability. kayoyin.github.io
🏂🎹🚵♀️🥋
🏂🎹🚵♀️🥋
We speculate that induction heads help models learn the more complex FV mechanism, which ultimately drives in-context learning 🤔
Paper: arxiv.org/abs/2502.14010
Paper: arxiv.org/abs/2502.14010
Which Attention Heads Matter for In-Context Learning?
Large language models (LLMs) exhibit impressive in-context learning (ICL) capability, enabling them to perform new tasks using only a few demonstrations in the prompt. Two different mechanisms have be...
arxiv.org
February 28, 2025 at 4:16 PM
We speculate that induction heads help models learn the more complex FV mechanism, which ultimately drives in-context learning 🤔
Paper: arxiv.org/abs/2502.14010
Paper: arxiv.org/abs/2502.14010
How to reconcile this with previous studies on ICL?
Key difference is that previous works:
- measure ICL using differences between token losses, which we find behaves differently to few-shot ICL accuracy
- don't control for overlap between induction and FV
- focus on small models
Key difference is that previous works:
- measure ICL using differences between token losses, which we find behaves differently to few-shot ICL accuracy
- don't control for overlap between induction and FV
- focus on small models
February 28, 2025 at 4:16 PM
How to reconcile this with previous studies on ICL?
Key difference is that previous works:
- measure ICL using differences between token losses, which we find behaves differently to few-shot ICL accuracy
- don't control for overlap between induction and FV
- focus on small models
Key difference is that previous works:
- measure ICL using differences between token losses, which we find behaves differently to few-shot ICL accuracy
- don't control for overlap between induction and FV
- focus on small models
Other interesting findings:
- FV heads have relatively high induction scores and vice versa compared to other heads
- FV heads emerge later in training than induction heads
- ICL accuracy rises around the same time induction emerges during training, but increases more gradually
- FV heads have relatively high induction scores and vice versa compared to other heads
- FV heads emerge later in training than induction heads
- ICL accuracy rises around the same time induction emerges during training, but increases more gradually
February 28, 2025 at 4:16 PM
Other interesting findings:
- FV heads have relatively high induction scores and vice versa compared to other heads
- FV heads emerge later in training than induction heads
- ICL accuracy rises around the same time induction emerges during training, but increases more gradually
- FV heads have relatively high induction scores and vice versa compared to other heads
- FV heads emerge later in training than induction heads
- ICL accuracy rises around the same time induction emerges during training, but increases more gradually
We also find evidence of induction heads that evolve into FV heads.
Several instances of FV heads have a high induction score earlier in training (around when induction heads first emerge). However, the reverse (induction heads with high FV scores earlier) does not occur.
Several instances of FV heads have a high induction score earlier in training (around when induction heads first emerge). However, the reverse (induction heads with high FV scores earlier) does not occur.
February 28, 2025 at 4:16 PM
We also find evidence of induction heads that evolve into FV heads.
Several instances of FV heads have a high induction score earlier in training (around when induction heads first emerge). However, the reverse (induction heads with high FV scores earlier) does not occur.
Several instances of FV heads have a high induction score earlier in training (around when induction heads first emerge). However, the reverse (induction heads with high FV scores earlier) does not occur.
2 mechanisms have been proposed to explain ICL: induction heads that find and copy relevant tokens, and FV heads that compute a latent encoding of the task from examples.
Our ablations show that FV heads are crucial for few-shot ICL, whereas induction heads are not necessary.
Our ablations show that FV heads are crucial for few-shot ICL, whereas induction heads are not necessary.
February 28, 2025 at 4:16 PM
2 mechanisms have been proposed to explain ICL: induction heads that find and copy relevant tokens, and FV heads that compute a latent encoding of the task from examples.
Our ablations show that FV heads are crucial for few-shot ICL, whereas induction heads are not necessary.
Our ablations show that FV heads are crucial for few-shot ICL, whereas induction heads are not necessary.
Thanks for the kind words, Seth 😊 glad you joined the dinner!
December 17, 2024 at 1:56 AM
Thanks for the kind words, Seth 😊 glad you joined the dinner!
sad I’m not in town for this, looks super exciting!! 🍿
December 4, 2024 at 10:07 PM
sad I’m not in town for this, looks super exciting!! 🍿
oof yeah I was afraid something like that was maybe going on. I hope she gets the help she needs…
December 4, 2024 at 9:26 PM
oof yeah I was afraid something like that was maybe going on. I hope she gets the help she needs…
ahh yes this is it thank you!! I hallucinated the end haha
November 26, 2024 at 11:20 PM
ahh yes this is it thank you!! I hallucinated the end haha
glad to know at least I didn’t just make this up 😭 I think I heard it recently too but can’t remember at alll
November 26, 2024 at 12:05 PM
glad to know at least I didn’t just make this up 😭 I think I heard it recently too but can’t remember at alll
Overall, handshapes in native ASL signs reflect communicative efficiency, but *not in signs borrowed from English*!
Check out our paper+code (w/ Terry Regier & Dan Klein) for more details and why we think that's the case: aclanthology.org/2024.acl-lon...
See you at TISLR in Ethiopia! ☀️ 8/8
Check out our paper+code (w/ Terry Regier & Dan Klein) for more details and why we think that's the case: aclanthology.org/2024.acl-lon...
See you at TISLR in Ethiopia! ☀️ 8/8
November 21, 2024 at 5:40 AM
Overall, handshapes in native ASL signs reflect communicative efficiency, but *not in signs borrowed from English*!
Check out our paper+code (w/ Terry Regier & Dan Klein) for more details and why we think that's the case: aclanthology.org/2024.acl-lon...
See you at TISLR in Ethiopia! ☀️ 8/8
Check out our paper+code (w/ Terry Regier & Dan Klein) for more details and why we think that's the case: aclanthology.org/2024.acl-lon...
See you at TISLR in Ethiopia! ☀️ 8/8
What about perceptual effort - could it be correlated with English usage?
Perceptual effort to distinguish between 2 handshapes is very weakly correlated with how often the 2 letters appear in similar contexts in English, and in the "wrong" direction for efficiency. 7/8
Perceptual effort to distinguish between 2 handshapes is very weakly correlated with how often the 2 letters appear in similar contexts in English, and in the "wrong" direction for efficiency. 7/8
November 21, 2024 at 5:40 AM
What about perceptual effort - could it be correlated with English usage?
Perceptual effort to distinguish between 2 handshapes is very weakly correlated with how often the 2 letters appear in similar contexts in English, and in the "wrong" direction for efficiency. 7/8
Perceptual effort to distinguish between 2 handshapes is very weakly correlated with how often the 2 letters appear in similar contexts in English, and in the "wrong" direction for efficiency. 7/8