[2410.13211] Estimating the Probabilities of Rare Outputs in Language Models
Abstract:We consider the problem of low probability estimation: given a machine learning model and a formally-specified input distribution, how can we estimate the probability of a binary property of the model's output, even when that probability is too small to estimate by random sampling? This problem is motivated by the need to improve worst-case performance, which distribution shift can make much more likely. We study low probability estimation in the context of argmax sampling from small transformer language models. We compare two types of methods: importance sampling, which involves searching for inputs giving rise to the rare output, and activation extrapolation, which involves extrapolating a probability distribution fit to the model's logits. We find that importance sampling outperforms activation extrapolation, but both outperform naive sampling. Finally, we explain how minimizing the probability estimate of an undesirable behavior generalizes adversarial training, and argue that new methods for low probability estimation are needed to provide stronger guarantees about worst-case performance.
Published as a conference paper at ICLR 2025 E STIMATING THE P ROBABILITIES OF R ARE O UTPUTS IN L ANGUAGE M ODELS Gabriel Wu∗ Jacob Hilton Alignment Research Center A BSTRACT We consider the problem of low probability estimation: given a machine learning arXiv:2410.13211v2 [cs.LG] 6…
Explore this link on the map →related reading
- [2410.13211] Estimating the Probabilities of Rare Outputs in Language Modelsarxiv.org
- [2603.02202] Frontier Models Can Take Actions at Low Probabilitiesarxiv.org
- Language Modelinglena-voita.github.io
- Competing with sampling — Alignment Research Centeralignment.org
- State of RL for reasoning LLMs | A. Weersaweers.de
- Large Language Models Must Be Taught to Know What They Don't Knowarxiv.org
- Musings on typicality – Sander Dielemansander.ai
- Large Language Model: world models or surface statistics?thegradient.pub
- ARC progress update: Competing with sampling — LessWronglesswrong.com
- A Gallery of Methods Beyond RL — Part I: Sampling Methodsshengyu-feng.github.io
- 2409.02908arxiv.org
- [1809.09087] Implicit Maximum Likelihood Estimationarxiv.org