Lucas H. McCabe

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 SE 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 (DSE) 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 DSE for sample coverage results in more accurate SE estimation in our setting of interest. Furthermore, we find that two semantic alphabet size estimators, including our proposed, flag incorrect LLM responses as well or better than many top-performing alternatives, with the added benefit of remaining highly interpretable.

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.

We present a nation-wide network analysis of non-fatal opioid-involved overdose journeys in the United States. Leveraging a unique proprietary dataset of Emergency Medical Services incidents, we construct a journey-to-overdose geospatial network capturing nearly half a million opioid-involved overdose events spanning 2018-2023. We analyze the structure and sociological profile of the nodes, which are counties or their equivalents, characterize the distribution of overdose journey lengths, and investigate changes in the journey network between 2018 and 2023. Our findings include that authority and hub nodes identified by the HITS algorithm tend to be located in urban areas and involved in overdose journeys with particularly long geographical distances.

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.