We know from prior work that LLMs encode social biases, and that this manifests in clinical tasks. In this work we adopt tools from mechanistic interpretability to unveil sociodemographic representations and biases within LLMs in the context of healthcare. Specifically, we ask: Can we identify activations within LLMs that encode sociodemographic information (e.g., gender, race)? We find that, in three open weight LLMs, gender information is highly localized in MLP layers and can be reliably manipulated at inference time via patching. Such interventions can surgically alter generated clinical vignettes for specific conditions, and also influence downstream clinical predictions which correlate with gender, e.g., patient risk of depression. We find that representation of patient race is somewhat more distributed, but can also be intervened upon, to a degree. To our knowledge, this is the first application of mechanistic interpretability methods to LLMs for healthcare.

Model distillation – using outputs from a large teacher model to teach a small student model – is a practical means of creating efficient models for a particular task. We ask: Can we identify a students’ teacher based on its outputs? Such “footprints” left by teacher LLMs would be interesting artifacts. Beyond this, reliable teacher inference may have practical implications as actors seek to distill specific capabilities of massive proprietary LLMs into deployed smaller LMs, potentially violating terms of service. We consider practical task distillation targets including summarization, question answering, and instruction-following. We assume a finite set of candidate teacher models, which we treat as blackboxes. We design discriminative models that operate over lexical features. We find that n-gram similarity alone is unreliable for identifying teachers, but part-of-speech (PoS) templates preferred by student models mimic those of their teachers.

The diversity across outputs generated by LLMs shapes perception of their quality and utility. High lexical diversity is often desirable, but there is no standard method to measure this property. Templated answer structures and “canned” responses across different documents are readily noticeable, but difficult to visualize across large corpora. This work aims to standardize measurement of text diversity. Specifically, we empirically investigate the convergent validity of existing scores across English texts, and release diversity, an open-source Python package (https://pypi.org/project/diversity/, https://github.com/cshaib/diversity) for measuring and extracting repetition in text. We also build a platform (https://ai-templates.app) based on diversity for users to interactively explore repetition in text. We find that fast compression algorithms capture information similar to what is measured by slow-to-compute n-gram overlap homogeneity scores. Further, a combination of measures—compression ratios, self-repetition of long n-grams, and Self-BLEU—are sufficient to report, as they have low mutual correlation with each other.

The diversity of text can be measured beyond word-level features, however existing diversity evaluation focuses primarily on word-level features. Here we propose a method for evaluating diversity over syntactic features to characterize general repetition in models, beyond frequent n-grams. Specifically, we define syntactic templates (e.g., strings comprising parts-of-speech) and show that models tend to produce templated text in downstream tasks at a higher rate than what is found in human-reference textsWe find that most (76%) templates in model-generated text can be found in pre-training data (compared to only 35% of human-authored text), and are not overwritten during fine-tuning or alignment processes such as RLHF. The connection between templates in generated text and the pre-training data allows us to analyze syntactic templates in models where we do not have the pre-training data.We also find that templates as features are able to differentiate between models, tasks, and domains, and are useful for qualitatively evaluating common model constructions.Finally, we demonstrate the use of templates as a useful tool for analyzing style memorization of training data in LLMs.

LLMs process text as sequences of tokens that roughly correspond to words, where less common words are represented by multiple tokens. However, individual tokens are often semantically unrelated to the meanings of the words/concepts they comprise. For example, Llama-2-7b’s tokenizer splits the word “patrolling” into two tokens, “pat” and “rolling”, neither of which correspond to semantically meaningful units like “patrol” or "-ing.” Similarly, the overall meanings of named entities like “Neil Young” and multi-word expressions like “break a leg” cannot be directly inferred from their constituent tokens. Mechanistically, how do LLMs convert such arbitrary groups of tokens into useful higher-level representations? In this work, we find that last token representations of named entities and multi-token words exhibit a pronounced “erasure” effect, where information about previous and current tokens is rapidly forgotten in early layers. Using this observation, we propose a method to “read out” the implicit vocabulary of an autoregressive LLM by examining differences in token representations across layers, and present results of this method for Llama-2-7b and Llama-3-8B. To our knowledge, this is the first attempt to probe the implicit vocabulary of an LLM.

Eliciting chain of thought (CoT) rationales - sequences of token that convey a “reasoning” process has been shown to consistently improve LLM performance on tasks like question answering. More recent efforts have shown that such rationales can also be used for model distillation: Including CoT sequences (elicited from a large “teacher” model) in addition to target labels when fine-tuning a small student model yields (often substantial) improvements. In this work we ask: Why and how does this additional training signal help in model distillation? We perform ablations to interrogate this, and report some potentially surprising results. Specifically: (1) Placing CoT sequences after labels (rather than before) realizes consistently better downstream performance – this means that no student “reasoning” is necessary at test time to realize gains. (2) When rationales are appended in this way, they need not be coherent reasoning sequences to yield improvements; performance increases are robust to permutations of CoT tokens, for example. In fact, (3) a small number of key tokens are sufficient to achieve improvements equivalent to those observed when full rationales are used in model distillation.

Entity matching is the task of linking records from different sources that refer to the same real-world entity. Past work has primarily treated entity linking as a standard supervised learning problem. However, supervised entity matching models often do not generalize well to new data, and collecting exhaustive labeled training data is often cost prohibitive. Further, recent efforts have adopted LLMs for this task in few/zero-shot settings, exploiting their general knowledge. But LLMs are prohibitively expensive for performing inference at scale for real-world entity matching tasks.As an efficient alternative, we re-cast entity matching as a conditional generation task as opposed to binary classification. This enables us to “distill” LLM reasoning into smaller entity matching models via natural language explanations. This approach achieves strong performance, especially on out-of-domain generalization tests (10.85% F-1) where standalone generative methods struggle. We perform ablations that highlight the importance of explanations, both for performance and model robustness.