Estimating the Probabilities of Rare Outputs in Language Models
This work addresses the challenge of ensuring reliable worst-case performance in language models for safety-critical applications, but it is incremental as it builds on existing estimation techniques.
The paper tackles the problem of estimating extremely low probabilities of specific outputs from language models, which is crucial for improving worst-case performance under distribution shift. They compare importance sampling and activation extrapolation on small transformer models, finding that importance sampling performs best, though both methods surpass naive sampling.
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.