Lucas H. McCabe

Rimon Melamed, Lucas H. McCabe, Tanay Wakhare et al.

We discover that many natural-language prompts can be replaced by corresponding prompts that are unintelligible to humans but that provably elicit similar behavior in language models. We call these prompts "evil twins" because they are obfuscated and uninterpretable (evil), but at the same time mimic the functionality of the original natural-language prompts (twins). Remarkably, evil twins transfer between models. We find these prompts by solving a maximum-likelihood problem which has applications of independent interest.

Lucas H. McCabe, Rimon Melamed, Thomas Hartvigsen et al.

Many black-box techniques for quantifying the uncertainty of large language models (LLMs) rely on repeated LLM sampling, which can be computationally expensive. Therefore, practical applicability demands reliable estimation from few samples. Semantic entropy (SE) is a popular sample-based uncertainty estimator with a discrete formulation attractive for the black-box setting. Recent extensions of semantic entropy exhibit improved LLM hallucination detection, but do so with less interpretable methods that admit additional hyperparameters. For this reason, we revisit the canonical discrete semantic entropy estimator, finding that it underestimates the "true" semantic entropy, as expected from theory. We propose a modified semantic alphabet size estimator, and illustrate that using it to adjust discrete semantic entropy for sample coverage results in more accurate semantic entropy estimation in our setting of interest. Furthermore, our proposed alphabet size estimator flags incorrect LLM responses as well or better than recent top-performing approaches, with the added benefit of remaining highly interpretable.

Rimon Melamed, Lucas H. McCabe, H. Howie Huang

Modern language models (LMs) are not robust to out-of-distribution inputs. Machine generated (``optimized'') prompts can be used to modulate LM outputs and induce specific behaviors while appearing completely uninterpretable. In this work, we investigate the composition of optimized prompts, as well as the mechanisms by which LMs parse and build predictions from optimized prompts. We find that optimized prompts primarily consist of punctuation and noun tokens which are more rare in the training data. Internally, optimized prompts are clearly distinguishable from natural language counterparts based on sparse subsets of the model's activations. Across various families of instruction-tuned models, optimized prompts follow a similar path in how their representations form through the network.