Victor Veitch
@vveitch.bsky.social
machine learning and artificial intelligence | University of Chicago / Google
come learn about LLM geometry!
Tomorrow (Fri) at ICLR2025, our paper on the geometry of categorical and hierarchical concepts in LLMs will be presented in both an oral (4:30 PM, Garnet 216–218) and a poster (10 AM, Hall 3 + Hall 2B #525) session. Sadly, I can’t be there in person… but Yo Joong will be presenting!
April 24, 2025 at 7:48 PM
come learn about LLM geometry!
I'll present this poster tonight at East exhibit hall a-c 2510. 5-7:30 pm.
Come chat about alignment!
Come chat about alignment!
LLM Alignment aims at making model outputs preferred by a ranker while changing as little 'off-target' behavior as possible.
Turns out:
-best-of-$n$ is the optimal option!
-you can contrastively train an LLM to mimic its own best-of-$n$ distribution!
BonBon alignment: arxiv.org/abs/2406.00832
Turns out:
-best-of-$n$ is the optimal option!
-you can contrastively train an LLM to mimic its own best-of-$n$ distribution!
BonBon alignment: arxiv.org/abs/2406.00832
On Spurious Associations and LLM Alignment
Large language models are `aligned' to bias them towards outputting responses that are good on various measures---e.g., we may want them to be helpful, factual, and polite. Often, alignment procedures...
simons.berkeley.edu
December 12, 2024 at 6:47 PM
I'll present this poster tonight at East exhibit hall a-c 2510. 5-7:30 pm.
Come chat about alignment!
Come chat about alignment!
I'll be at NeurIPS Thursday-Sunday; send me an email if you'd like to chat :)
December 10, 2024 at 2:11 AM
I'll be at NeurIPS Thursday-Sunday; send me an email if you'd like to chat :)
LLM Alignment aims at making model outputs preferred by a ranker while changing as little 'off-target' behavior as possible.
Turns out:
-best-of-$n$ is the optimal option!
-you can contrastively train an LLM to mimic its own best-of-$n$ distribution!
BonBon alignment: arxiv.org/abs/2406.00832
Turns out:
-best-of-$n$ is the optimal option!
-you can contrastively train an LLM to mimic its own best-of-$n$ distribution!
BonBon alignment: arxiv.org/abs/2406.00832
On Spurious Associations and LLM Alignment
Large language models are `aligned' to bias them towards outputting responses that are good on various measures---e.g., we may want them to be helpful, factual, and polite. Often, alignment procedures...
simons.berkeley.edu
November 23, 2024 at 11:21 PM
LLM Alignment aims at making model outputs preferred by a ranker while changing as little 'off-target' behavior as possible.
Turns out:
-best-of-$n$ is the optimal option!
-you can contrastively train an LLM to mimic its own best-of-$n$ distribution!
BonBon alignment: arxiv.org/abs/2406.00832
Turns out:
-best-of-$n$ is the optimal option!
-you can contrastively train an LLM to mimic its own best-of-$n$ distribution!
BonBon alignment: arxiv.org/abs/2406.00832