Byron C. Wallace

Eric Lehman, Vladislav Lialin, Katelyn Y. Legaspi et al.

Existing question answering (QA) datasets derived from electronic health records (EHR) are artificially generated and consequently fail to capture realistic physician information needs. We present Discharge Summary Clinical Questions (DiSCQ), a newly curated question dataset composed of 2,000+ questions paired with the snippets of text (triggers) that prompted each question. The questions are generated by medical experts from 100+ MIMIC-III discharge summaries. We analyze this dataset to characterize the types of information sought by medical experts. We also train baseline models for trigger detection and question generation (QG), paired with unsupervised answer retrieval over EHRs. Our baseline model is able to generate high quality questions in over 62% of cases when prompted with human selected triggers. We release this dataset (and all code to reproduce baseline model results) to facilitate further research into realistic clinical QA and QG: https://github.com/elehman16/discq.

Junru Lu, Jiazheng Li, Byron C. Wallace et al.

Accessing medical literature is difficult for laypeople as the content is written for specialists and contains medical jargon. Automated text simplification methods offer a potential means to address this issue. In this work, we propose a summarize-then-simplify two-stage strategy, which we call NapSS, identifying the relevant content to simplify while ensuring that the original narrative flow is preserved. In this approach, we first generate reference summaries via sentence matching between the original and the simplified abstracts. These summaries are then used to train an extractive summarizer, learning the most relevant content to be simplified. Then, to ensure the narrative consistency of the simplified text, we synthesize auxiliary narrative prompts combining key phrases derived from the syntactical analyses of the original text. Our model achieves results significantly better than the seq2seq baseline on an English medical corpus, yielding 3%~4% absolute improvements in terms of lexical similarity, and providing a further 1.1% improvement of SARI score when combined with the baseline. We also highlight shortcomings of existing evaluation methods, and introduce new metrics that take into account both lexical and high-level semantic similarity. A human evaluation conducted on a random sample of the test set further establishes the effectiveness of the proposed approach. Codes and models are released here: https://github.com/LuJunru/NapSS.

Sarthak Jain, Varun Manjunatha, Byron C. Wallace et al.

Many language tasks (e.g., Named Entity Recognition, Part-of-Speech tagging, and Semantic Role Labeling) are naturally framed as sequence tagging problems. However, there has been comparatively little work on interpretability methods for sequence tagging models. In this paper, we extend influence functions - which aim to trace predictions back to the training points that informed them - to sequence tagging tasks. We define the influence of a training instance segment as the effect that perturbing the labels within this segment has on a test segment level prediction. We provide an efficient approximation to compute this, and show that it tracks with the true segment influence, measured empirically. We show the practical utility of segment influence by using the method to identify systematic annotation errors in two named entity recognition corpora. Code to reproduce our results is available at https://github.com/successar/Segment_Influence_Functions.

Kaijie Mo, Thomas Yang, Chantal Shaib et al.

The morphological form of a word can often give cues to its meaning, but purely relying on these mappings can lead to overgeneralization in high-stakes domains. In the medical domain, for instance, LLMs can confidently reason about fictitious drugs from their affixes alone (e.g., wugcillin) and generate plausible-looking clinical content. We present a behavioral and mechanistic study of LLM "affix heuristics" in pharmacology. Using fictitious drug names built from real affixes, we show that affix signals alone elicit class-level pharmacological responses. We introduce a framework for identifying whether a model's drug semantics are driven mainly by the affix, the stem, or the drug name as a whole. Applied across 653 drugs, our framework reveals that models often induce drug meaning primarily through affix cues, yet rarely explicitly indicate this reliance, and sometimes incorrectly conflate properties among affix-sharing drugs. Activation patching across models further localizes this behavior to early-mid layers. These findings show that morphological shortcuts pose a subtle but measurable risk to safety.

Denis Jered McInerney, Geoffrey Young, Jan-Willem van de Meent et al. · amazon-science, salesforce

We propose CHiLL (Crafting High-Level Latents), an approach for natural-language specification of features for linear models. CHiLL prompts LLMs with expert-crafted queries to generate interpretable features from health records. The resulting noisy labels are then used to train a simple linear classifier. Generating features based on queries to an LLM can empower physicians to use their domain expertise to craft features that are clinically meaningful for a downstream task of interest, without having to manually extract these from raw EHR. We are motivated by a real-world risk prediction task, but as a reproducible proxy, we use MIMIC-III and MIMIC-CXR data and standard predictive tasks (e.g., 30-day readmission) to evaluate this approach. We find that linear models using automatically extracted features are comparably performant to models using reference features, and provide greater interpretability than linear models using "Bag-of-Words" features. We verify that learned feature weights align well with clinical expectations.

Hiba Ahsan, Denis Jered McInerney, Jisoo Kim et al. · amazon-science, salesforce

Unstructured data in Electronic Health Records (EHRs) often contains critical information -- complementary to imaging -- that could inform radiologists' diagnoses. But the large volume of notes often associated with patients together with time constraints renders manually identifying relevant evidence practically infeasible. In this work we propose and evaluate a zero-shot strategy for using LLMs as a mechanism to efficiently retrieve and summarize unstructured evidence in patient EHR relevant to a given query. Our method entails tasking an LLM to infer whether a patient has, or is at risk of, a particular condition on the basis of associated notes; if so, we ask the model to summarize the supporting evidence. Under expert evaluation, we find that this LLM-based approach provides outputs consistently preferred to a pre-LLM information retrieval baseline. Manual evaluation is expensive, so we also propose and validate a method using an LLM to evaluate (other) LLM outputs for this task, allowing us to scale up evaluation. Our findings indicate the promise of LLMs as interfaces to EHR, but also highlight the outstanding challenge posed by "hallucinations". In this setting, however, we show that model confidence in outputs strongly correlates with faithful summaries, offering a practical means to limit confabulations.

Jaden Fiotto-Kaufman, Alexander R. Loftus, Eric Todd et al.

We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. The National Deep Inference Fabric (NDIF) is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the Intervention Graph, an architecture developed to decouple experimental design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code, documentation, and tutorials are available at https://nnsight.net/.

Zhaoyue Sun, Jiazheng Li, Gabriele Pergola et al.

The primary goal of drug safety researchers and regulators is to promptly identify adverse drug reactions. Doing so may in turn prevent or reduce the harm to patients and ultimately improve public health. Evaluating and monitoring drug safety (i.e., pharmacovigilance) involves analyzing an ever growing collection of spontaneous reports from health professionals, physicians, and pharmacists, and information voluntarily submitted by patients. In this scenario, facilitating analysis of such reports via automation has the potential to rapidly identify safety signals. Unfortunately, public resources for developing natural language models for this task are scant. We present PHEE, a novel dataset for pharmacovigilance comprising over 5000 annotated events from medical case reports and biomedical literature, making it the largest such public dataset to date. We describe the hierarchical event schema designed to provide coarse and fine-grained information about patients' demographics, treatments and (side) effects. Along with the discussion of the dataset, we present a thorough experimental evaluation of current state-of-the-art approaches for biomedical event extraction, point out their limitations, and highlight open challenges to foster future research in this area.

Ashwin Devaraj, William Sheffield, Byron C. Wallace et al.

Automated simplification models aim to make input texts more readable. Such methods have the potential to make complex information accessible to a wider audience, e.g., providing access to recent medical literature which might otherwise be impenetrable for a lay reader. However, such models risk introducing errors into automatically simplified texts, for instance by inserting statements unsupported by the corresponding original text, or by omitting key information. Providing more readable but inaccurate versions of texts may in many cases be worse than providing no such access at all. The problem of factual accuracy (and the lack thereof) has received heightened attention in the context of summarization models, but the factuality of automatically simplified texts has not been investigated. We introduce a taxonomy of errors that we use to analyze both references drawn from standard simplification datasets and state-of-the-art model outputs. We find that errors often appear in both that are not captured by existing evaluation metrics, motivating a need for research into ensuring the factual accuracy of automated simplification models.

Nikita Salkar, Thomas Trikalinos, Byron C. Wallace et al.

We provide a quantitative and qualitative analysis of self-repetition in the output of neural summarizers. We measure self-repetition as the number of n-grams of length four or longer that appear in multiple outputs of the same system. We analyze the behavior of three popular architectures (BART, T5, and Pegasus), fine-tuned on five datasets. In a regression analysis, we find that the three architectures have different propensities for repeating content across output summaries for inputs, with BART being particularly prone to self-repetition. Fine-tuning on more abstractive data, and on data featuring formulaic language, is associated with a higher rate of self-repetition. In qualitative analysis we find systems produce artefacts such as ads and disclaimers unrelated to the content being summarized, as well as formulaic phrases common in the fine-tuning domain. Our approach to corpus-level analysis of self-repetition may help practitioners clean up training data for summarizers and ultimately support methods for minimizing the amount of self-repetition.

Denis Jered McInerney, Geoffrey Young, Jan-Willem van de Meent et al. · amazon-science, salesforce

Pretraining multimodal models on Electronic Health Records (EHRs) provides a means of learning representations that can transfer to downstream tasks with minimal supervision. Recent multimodal models induce soft local alignments between image regions and sentences. This is of particular interest in the medical domain, where alignments might highlight regions in an image relevant to specific phenomena described in free-text. While past work has suggested that attention "heatmaps" can be interpreted in this manner, there has been little evaluation of such alignments. We compare alignments from a state-of-the-art multimodal (image and text) model for EHR with human annotations that link image regions to sentences. Our main finding is that the text has an often weak or unintuitive influence on attention; alignments do not consistently reflect basic anatomical information. Moreover, synthetic modifications -- such as substituting "left" for "right" -- do not substantially influence highlights. Simple techniques such as allowing the model to opt out of attending to the image and few-shot finetuning show promise in terms of their ability to improve alignments with very little or no supervision. We make our code and checkpoints open-source.

Jay DeYoung, Stephanie C. Martinez, Iain J. Marshall et al.

Multi-document summarization entails producing concise synopses of collections of inputs. For some applications, the synopsis should accurately synthesize inputs with respect to a key aspect, e.g., a synopsis of film reviews written about a particular movie should reflect the average critic consensus. As a more consequential example, narrative summaries that accompany biomedical systematic reviews of clinical trial results should accurately summarize the potentially conflicting results from individual trials. In this paper we ask: To what extent do modern multi-document summarization models implicitly perform this sort of synthesis? We run experiments over opinion and evidence synthesis datasets using a suite of summarization models, from fine-tuned transformers to GPT-4. We find that existing models partially perform synthesis, but imperfectly: even the best performing models are over-sensitive to changes in input ordering and under-sensitive to changes in input compositions (e.g., ratio of positive to negative reviews). We propose a simple, general, effective method for improving model synthesis capabilities by generating an explicitly diverse set of candidate outputs, and then selecting from these the string best aligned with the expected aggregate measure for the inputs, or abstaining when the model produces no good candidate.

Gongbo Zhang, Qiao Jin, Denis Jered McInerney et al. · amazon-science, salesforce

Evidence-based medicine promises to improve the quality of healthcare by empowering medical decisions and practices with the best available evidence. The rapid growth of medical evidence, which can be obtained from various sources, poses a challenge in collecting, appraising, and synthesizing the evidential information. Recent advancements in generative AI, exemplified by large language models, hold promise in facilitating the arduous task. However, developing accountable, fair, and inclusive models remains a complicated undertaking. In this perspective, we discuss the trustworthiness of generative AI in the context of automated summarization of medical evidence.

Eric Todd, Millicent L. Li, Arnab Sen Sharma et al.

We report the presence of a simple neural mechanism that represents an input-output function as a vector within autoregressive transformer language models (LMs). Using causal mediation analysis on a diverse range of in-context-learning (ICL) tasks, we find that a small number attention heads transport a compact representation of the demonstrated task, which we call a function vector (FV). FVs are robust to changes in context, i.e., they trigger execution of the task on inputs such as zero-shot and natural text settings that do not resemble the ICL contexts from which they are collected. We test FVs across a range of tasks, models, and layers and find strong causal effects across settings in middle layers. We investigate the internal structure of FVs and find while that they often contain information that encodes the output space of the function, this information alone is not sufficient to reconstruct an FV. Finally, we test semantic vector composition in FVs, and find that to some extent they can be summed to create vectors that trigger new complex tasks. Our findings show that compact, causal internal vector representations of function abstractions can be explicitly extracted from LLMs. Our code and data are available at https://functions.baulab.info.

Jinghan Yang, Sarthak Jain, Byron C. Wallace

We consider the problem of identifying a minimal subset of training data $\mathcal{S}_t$ such that if the instances comprising $\mathcal{S}_t$ had been removed prior to training, the categorization of a given test point $x_t$ would have been different. Identifying such a set may be of interest for a few reasons. First, the cardinality of $\mathcal{S}_t$ provides a measure of robustness (if $|\mathcal{S}_t|$ is small for $x_t$, we might be less confident in the corresponding prediction), which we show is correlated with but complementary to predicted probabilities. Second, interrogation of $\mathcal{S}_t$ may provide a novel mechanism for contesting a particular model prediction: If one can make the case that the points in $\mathcal{S}_t$ are wrongly labeled or irrelevant, this may argue for overturning the associated prediction. Identifying $\mathcal{S}_t$ via brute-force is intractable. We propose comparatively fast approximation methods to find $\mathcal{S}_t$ based on influence functions, and find that -- for simple convex text classification models -- these approaches can often successfully identify relatively small sets of training examples which, if removed, would flip the prediction.

Sanjana Ramprasad, Denis Jered McInerney, Iain J. Marshal et al. · amazon-science, salesforce

We present TrialsSummarizer, a system that aims to automatically summarize evidence presented in the set of randomized controlled trials most relevant to a given query. Building on prior work, the system retrieves trial publications matching a query specifying a combination of condition, intervention(s), and outcome(s), and ranks these according to sample size and estimated study quality. The top-k such studies are passed through a neural multi-document summarization system, yielding a synopsis of these trials. We consider two architectures: A standard sequence-to-sequence model based on BART and a multi-headed architecture intended to provide greater transparency to end-users. Both models produce fluent and relevant summaries of evidence retrieved for queries, but their tendency to introduce unsupported statements render them inappropriate for use in this domain at present. The proposed architecture may help users verify outputs allowing users to trace generated tokens back to inputs.

Diego Garcia-Olano, Yasumasa Onoe, Joydeep Ghosh et al.

Interpretable entity representations (IERs) are sparse embeddings that are "human-readable" in that dimensions correspond to fine-grained entity types and values are predicted probabilities that a given entity is of the corresponding type. These methods perform well in zero-shot and low supervision settings. Compared to standard dense neural embeddings, such interpretable representations may permit analysis and debugging. However, while fine-tuning sparse, interpretable representations improves accuracy on downstream tasks, it destroys the semantics of the dimensions which were enforced in pre-training. Can we maintain the interpretable semantics afforded by IERs while improving predictive performance on downstream tasks? Toward this end, we propose Intermediate enTity-based Sparse Interpretable Representation Learning (ItsIRL). ItsIRL realizes improved performance over prior IERs on biomedical tasks, while maintaining "interpretability" generally and their ability to support model debugging specifically. The latter is enabled in part by the ability to perform "counterfactual" fine-grained entity type manipulation, which we explore in this work. Finally, we propose a method to construct entity type based class prototypes for revealing global semantic properties of classes learned by our model.

Alberto Mario Ceballos Arroyo, Monica Munnangi, Jiuding Sun et al. · amazon-science, salesforce

Instruction-tuned Large Language Models (LLMs) can perform a wide range of tasks given natural language instructions to do so, but they are sensitive to how such instructions are phrased. This issue is especially concerning in healthcare, as clinicians are unlikely to be experienced prompt engineers and the potential consequences of inaccurate outputs are heightened in this domain. This raises a practical question: How robust are instruction-tuned LLMs to natural variations in the instructions provided for clinical NLP tasks? We collect prompts from medical doctors across a range of tasks and quantify the sensitivity of seven LLMs -- some general, others specialized -- to natural (i.e., non-adversarial) instruction phrasings. We find that performance varies substantially across all models, and that -- perhaps surprisingly -- domain-specific models explicitly trained on clinical data are especially brittle, compared to their general domain counterparts. Further, arbitrary phrasing differences can affect fairness, e.g., valid but distinct instructions for mortality prediction yield a range both in overall performance, and in terms of differences between demographic groups.

Jiuding Sun, Chantal Shaib, Byron C. Wallace

Instruction fine-tuning has recently emerged as a promising approach for improving the zero-shot capabilities of Large Language Models (LLMs) on new tasks. This technique has shown particular strength in improving the performance of modestly sized LLMs, sometimes inducing performance competitive with much larger model variants. In this paper we ask two questions: (1) How sensitive are instruction-tuned models to the particular phrasings of instructions, and, (2) How can we make them more robust to such natural language variation? To answer the former, we collect a set of 319 instructions manually written by NLP practitioners for over 80 unique tasks included in widely used benchmarks, and we evaluate the variance and average performance of these instructions as compared to instruction phrasings observed during instruction fine-tuning. We find that using novel (unobserved) but appropriate instruction phrasings consistently degrades model performance, sometimes substantially so. Further, such natural instructions yield a wide variance in downstream performance, despite their semantic equivalence. Put another way, instruction-tuned models are not especially robust to instruction re-phrasings. We propose a simple method to mitigate this issue by introducing ``soft prompt'' embedding parameters and optimizing these to maximize the similarity between representations of semantically equivalent instructions. We show that this method consistently improves the robustness of instruction-tuned models.

Koyena Pal, Jiuding Sun, Andrew Yuan et al.

We conjecture that hidden state vectors corresponding to individual input tokens encode information sufficient to accurately predict several tokens ahead. More concretely, in this paper we ask: Given a hidden (internal) representation of a single token at position $t$ in an input, can we reliably anticipate the tokens that will appear at positions $\geq t + 2$? To test this, we measure linear approximation and causal intervention methods in GPT-J-6B to evaluate the degree to which individual hidden states in the network contain signal rich enough to predict future hidden states and, ultimately, token outputs. We find that, at some layers, we can approximate a model's output with more than 48% accuracy with respect to its prediction of subsequent tokens through a single hidden state. Finally we present a "Future Lens" visualization that uses these methods to create a new view of transformer states.

Chantal Shaib, Joe Barrow, Jiuding Sun et al.

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 we release diversity, an open-source Python package for measuring and extracting repetition in text. We also build a platform 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 and BERTScore -- are sufficient to report, as they have low mutual correlation with each other.

Jan Trienes, Sebastian Joseph, Jörg Schlötterer et al.

Text simplification aims to make technical texts more accessible to laypeople but often results in deletion of information and vagueness. This work proposes InfoLossQA, a framework to characterize and recover simplification-induced information loss in form of question-and-answer (QA) pairs. Building on the theory of Question Under Discussion, the QA pairs are designed to help readers deepen their knowledge of a text. We conduct a range of experiments with this framework. First, we collect a dataset of 1,000 linguist-curated QA pairs derived from 104 LLM simplifications of scientific abstracts of medical studies. Our analyses of this data reveal that information loss occurs frequently, and that the QA pairs give a high-level overview of what information was lost. Second, we devise two methods for this task: end-to-end prompting of open-source and commercial language models, and a natural language inference pipeline. With a novel evaluation framework considering the correctness of QA pairs and their linguistic suitability, our expert evaluation reveals that models struggle to reliably identify information loss and applying similar standards as humans at what constitutes information loss.

Zihao Yang, Mosh Levy, Yoav Goldberg et al.

Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and CoT faithfulness. But in this work we argue that observed effects cannot be attributed to the targeted factor without accounting for baseline ``meaning-preserving'' modifications to text that establish general model sensitivity. This is because every counterfactual edit is a compound treatment that bundles the variable of interest with incidental surface-form variation; this violates treatment variation irrelevance. We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender. However, this is statistically indistinguishable from the flip rates induced by simply paraphrasing inputs (14.1%). In this case, it would therefore be unwarranted to conclude that the LLM is especially sensitive to patient gender. To account for this and robustly measure the effects of targeted interventions, we propose a framework in which we compare (via statistical testing) differences observed under target interventions to those induced by paraphrasing inputs. We then use this framework to revisit a analysis done on the MedPerturb dataset, which reported evidence of model sensitivity to patient demographics and stylistic cues. We find that these effects largely dissipate when we account for general model sensitivity, with only 5 of 120 tests reaching statistical significance. Applying the same framework to occupational biography classification, we detect clearly significant directional gender bias, showing that the framework identifies real directional effects even when they are small. We evaluate a range of metrics -- aggregate, per-sample distributional, and regression -- and find that per-sample metrics are dramatically more powerful than aggregate metrics and regression powerfully and uniquely characterizes effect direction and magnitude.

Kaijie Mo, Siddhartha Venkatayogi, Chantal Shaib et al.

In high-stakes domains like medicine, it may be generally desirable for models to faithfully adhere to the context provided. But what happens if the context does not align with model priors or safety protocols? In this paper, we investigate how LLMs behave and reason when presented with counterfactual (or even adversarial) medical evidence. We first construct MedCounterFact, a counterfactual medical QA dataset that requires the models to answer clinical comparison questions (i.e., judge the efficacy of certain treatments, with evidence consisting of randomized controlled trials provided as context). In MedCounterFact, real-world medical interventions within the questions and evidence are systematically replaced with four types of counterfactual stimuli, ranging from unknown words to toxic substances. Our evaluation across multiple frontier LLMs on MedCounterFact reveals that in the presence of counterfactual evidence, existing models overwhelmingly accept such "evidence" at face value even when it is dangerous or implausible, and provide confident and uncaveated answers. While it may be prudent to draw a boundary between faithfulness and safety, our findings suggest that models arguably overemphasize the former.

Hye Sun Yun, Geetika Kapoor, Michael Mackert et al.

Patients are increasingly turning to large language models (LLMs) with medical questions that are complex and difficult to articulate clearly. However, LLMs are sensitive to prompt phrasings and can be influenced by the way questions are worded. Ideally, LLMs should respond consistently regardless of phrasing, particularly when grounded in the same underlying evidence. We investigate this through a systematic evaluation in a controlled retrieval-augmented generation (RAG) setting for medical question answering (QA), where expert-selected documents are used rather than retrieved automatically. We examine two dimensions of patient query variation: question framing (positive vs. negative) and language style (technical vs. plain language). We construct a dataset of 6,614 query pairs grounded in clinical trial abstracts and evaluate response consistency across eight LLMs. Our findings show that positively- and negatively-framed pairs are significantly more likely to produce contradictory conclusions than same-framing pairs. This framing effect is further amplified in multi-turn conversations, where sustained persuasion increases inconsistency. We find no significant interaction between framing and language style. Our results demonstrate that LLM responses in medical QA can be systematically influenced through query phrasing alone, even when grounded in the same evidence, highlighting the importance of phrasing robustness as an evaluation criterion for RAG-based systems in high-stakes settings.

Debjyoti Saha Roy, Byron C. Wallace, Javed A. Aslam

State-of-the-art Extreme Multi-Label Text Classification models rely on multi-label attention to focus on key tokens in input text, but learning good attention weights is challenging. We introduce PLANT - Pretrained and Leveraged Attention - a plug-and-play strategy for initializing attention. PLANT works by planting label-specific attention using a pretrained Learning-to-Rank model guided by mutual information gain. This architecture-agnostic approach integrates seamlessly with large language model backbones such as Mistral-7B, LLaMA3-8B, DeepSeek-V3, and Phi-3. PLANT outperforms state-of-the-art methods across tasks including ICD coding, legal topic classification, and content recommendation. Gains are especially pronounced in few-shot settings, with substantial improvements on rare labels. Ablation studies confirm that attention initialization is a key driver of these gains. For code and trained models, see https://github.com/debjyotiSRoy/xcube/tree/plant

Eric Lehman, Sarthak Jain, Karl Pichotta et al.

Large Transformers pretrained over clinical notes from Electronic Health Records (EHR) have afforded substantial gains in performance on predictive clinical tasks. The cost of training such models (and the necessity of data access to do so) coupled with their utility motivates parameter sharing, i.e., the release of pretrained models such as ClinicalBERT. While most efforts have used deidentified EHR, many researchers have access to large sets of sensitive, non-deidentified EHR with which they might train a BERT model (or similar). Would it be safe to release the weights of such a model if they did? In this work, we design a battery of approaches intended to recover Personal Health Information (PHI) from a trained BERT. Specifically, we attempt to recover patient names and conditions with which they are associated. We find that simple probing methods are not able to meaningfully extract sensitive information from BERT trained over the MIMIC-III corpus of EHR. However, more sophisticated "attacks" may succeed in doing so: To facilitate such research, we make our experimental setup and baseline probing models available at https://github.com/elehman16/exposing_patient_data_release

Pouya Pezeshkpour, Sarthak Jain, Byron C. Wallace et al.

Widespread adoption of deep models has motivated a pressing need for approaches to interpret network outputs and to facilitate model debugging. Instance attribution methods constitute one means of accomplishing these goals by retrieving training instances that (may have) led to a particular prediction. Influence functions (IF; Koh and Liang 2017) provide machinery for doing this by quantifying the effect that perturbing individual train instances would have on a specific test prediction. However, even approximating the IF is computationally expensive, to the degree that may be prohibitive in many cases. Might simpler approaches (e.g., retrieving train examples most similar to a given test point) perform comparably? In this work, we evaluate the degree to which different potential instance attribution agree with respect to the importance of training samples. We find that simple retrieval methods yield training instances that differ from those identified via gradient-based methods (such as IFs), but that nonetheless exhibit desirable characteristics similar to more complex attribution methods. Code for all methods and experiments in this paper is available at: https://github.com/successar/instance_attributions_NLP.

Sarthak Jain, Sarah Wiegreffe, Yuval Pinter et al.

In many settings it is important for one to be able to understand why a model made a particular prediction. In NLP this often entails extracting snippets of an input text `responsible for' corresponding model output; when such a snippet comprises tokens that indeed informed the model's prediction, it is a faithful explanation. In some settings, faithfulness may be critical to ensure transparency. Lei et al. (2016) proposed a model to produce faithful rationales for neural text classification by defining independent snippet extraction and prediction modules. However, the discrete selection over input tokens performed by this method complicates training, leading to high variance and requiring careful hyperparameter tuning. We propose a simpler variant of this approach that provides faithful explanations by construction. In our scheme, named FRESH, arbitrary feature importance scores (e.g., gradients from a trained model) are used to induce binary labels over token inputs, which an extractor can be trained to predict. An independent classifier module is then trained exclusively on snippets provided by the extractor; these snippets thus constitute faithful explanations, even if the classifier is arbitrarily complex. In both automatic and manual evaluations we find that variants of this simple framework yield predictive performance superior to `end-to-end' approaches, while being more general and easier to train. Code is available at https://github.com/successar/FRESH

Sarthak Jain, Byron C. Wallace

Attention mechanisms have seen wide adoption in neural NLP models. In addition to improving predictive performance, these are often touted as affording transparency: models equipped with attention provide a distribution over attended-to input units, and this is often presented (at least implicitly) as communicating the relative importance of inputs. However, it is unclear what relationship exists between attention weights and model outputs. In this work, we perform extensive experiments across a variety of NLP tasks that aim to assess the degree to which attention weights provide meaningful `explanations' for predictions. We find that they largely do not. For example, learned attention weights are frequently uncorrelated with gradient-based measures of feature importance, and one can identify very different attention distributions that nonetheless yield equivalent predictions. Our findings show that standard attention modules do not provide meaningful explanations and should not be treated as though they do. Code for all experiments is available at https://github.com/successar/AttentionExplanation.

Zhaoyue Sun, Gabriele Pergola, Byron C. Wallace et al.

With the advent of large language models (LLMs), there has been growing interest in exploring their potential for medical applications. This research aims to investigate the ability of LLMs, specifically ChatGPT, in the context of pharmacovigilance event extraction, of which the main goal is to identify and extract adverse events or potential therapeutic events from textual medical sources. We conduct extensive experiments to assess the performance of ChatGPT in the pharmacovigilance event extraction task, employing various prompts and demonstration selection strategies. The findings demonstrate that while ChatGPT demonstrates reasonable performance with appropriate demonstration selection strategies, it still falls short compared to fully fine-tuned small models. Additionally, we explore the potential of leveraging ChatGPT for data augmentation. However, our investigation reveals that the inclusion of synthesized data into fine-tuning may lead to a decrease in performance, possibly attributed to noise in the ChatGPT-generated labels. To mitigate this, we explore different filtering strategies and find that, with the proper approach, more stable performance can be achieved, although constant improvement remains elusive.

Hiba Ahsan, Byron C. Wallace

LLMs are increasingly being used in healthcare. This promises to free physicians from drudgery, enabling better care to be delivered at scale. But the use of LLMs in this space also brings risks; for example, such models may worsen existing biases. How can we spot when LLMs are (spuriously) relying on patient race to inform predictions? In this work we assess the degree to which Sparse Autoencoders (SAEs) can reveal (and control) associations the model has made between race and stigmatizing concepts. We first identify SAE latents in Gemma-2 models which appear to correlate with Black individuals. We find that this latent activates on reasonable input sequences (e.g., "African American") but also problematic words like "incarceration". We then show that we can use this latent to steer models to generate outputs about Black patients, and further that this can induce problematic associations in model outputs as a result. For example, activating the Black latent increases the risk assigned to the probability that a patient will become "belligerent". We evaluate the degree to which such steering via latents might be useful for mitigating bias. We find that this offers improvements in simple settings, but is less successful for more realistic and complex clinical tasks. Overall, our results suggest that: SAEs may offer a useful tool in clinical applications of LLMs to identify problematic reliance on demographics but mitigating bias via SAE steering appears to be of marginal utility for realistic tasks.

Kundan Krishna, Sanjana Ramprasad, Prakhar Gupta et al. · cmu

LLMs can generate factually incorrect statements even when provided access to reference documents. Such errors can be dangerous in high-stakes applications (e.g., document-grounded QA for healthcare or finance). We present GenAudit -- a tool intended to assist fact-checking LLM responses for document-grounded tasks. GenAudit suggests edits to the LLM response by revising or removing claims that are not supported by the reference document, and also presents evidence from the reference for facts that do appear to have support. We train models to execute these tasks, and design an interactive interface to present suggested edits and evidence to users. Comprehensive evaluation by human raters shows that GenAudit can detect errors in 8 different LLM outputs when summarizing documents from diverse domains. User studies demonstrate that using GenAudit can substantially improve the performance of humans at finding errors in LLM-generated summaries. We release our tool (GenAudit) and fact-checking model for public use.

Sanjana Ramprasad, Byron C. Wallace

Modern LLMs can now produce highly readable abstractive summaries, to the point that traditional automated metrics for evaluating summary quality, such as ROUGE, have saturated. However, LLMs still sometimes introduce inaccuracies into summaries, i.e., information inconsistent with or unsupported by the corresponding source. Measuring the occurrence of these often subtle factual inconsistencies automatically has proved challenging. This in turn has motivated development of metrics intended to measure the factual consistency of generated summaries against sources. But are these approaches measuring what they purport to? Or are they mostly exploiting artifacts? In this work, we stress test a range of automatic factuality metrics, including specialized models and LLM-based prompting methods, to probe what they actually capture. Using a shallow classifier to separate ``easy'' examples for factual evaluation where surface features suffice from ``hard'' cases requiring deeper reasoning, we find that all metrics show substantial performance drops on the latter. Furthermore, some metrics are more sensitive to benign, fact-preserving edits than to factual corrections. Building on this observation, we demonstrate that most automatic factuality metrics can be gamed, i.e., their scores can be artificially inflated by appending innocuous, content-free sentences to summaries. Among the metrics tested, the prompt based ChatGPT-DA approach is the most robust and reliable. However, this comes with a notable caveat: Prompting LLMs to assess factuality may overly rely on their parametric knowledge rather than the provided reference when making judgments. Taken together, our findings call into question the reliability of current factuality metrics and prompt a broader reflection on what these metrics are truly measuring.

Hye Sun Yun, David Pogrebitskiy, Iain J. Marshall et al.

Meta-analyses statistically aggregate the findings of different randomized controlled trials (RCTs) to assess treatment effectiveness. Because this yields robust estimates of treatment effectiveness, results from meta-analyses are considered the strongest form of evidence. However, rigorous evidence syntheses are time-consuming and labor-intensive, requiring manual extraction of data from individual trials to be synthesized. Ideally, language technologies would permit fully automatic meta-analysis, on demand. This requires accurately extracting numerical results from individual trials, which has been beyond the capabilities of natural language processing (NLP) models to date. In this work, we evaluate whether modern large language models (LLMs) can reliably perform this task. We annotate (and release) a modest but granular evaluation dataset of clinical trial reports with numerical findings attached to interventions, comparators, and outcomes. Using this dataset, we evaluate the performance of seven LLMs applied zero-shot for the task of conditionally extracting numerical findings from trial reports. We find that massive LLMs that can accommodate lengthy inputs are tantalizingly close to realizing fully automatic meta-analysis, especially for dichotomous (binary) outcomes (e.g., mortality). However, LLMs -- including ones trained on biomedical texts -- perform poorly when the outcome measures are complex and tallying the results requires inference. This work charts a path toward fully automatic meta-analysis of RCTs via LLMs, while also highlighting the limitations of existing models for this aim.

Denis Jered McInerney, William Dickinson, Lucy C. Flynn et al. · amazon-science, salesforce

Many diagnostic errors occur because clinicians cannot easily access relevant information in patient Electronic Health Records (EHRs). In this work we propose a method to use LLMs to identify pieces of evidence in patient EHR data that indicate increased or decreased risk of specific diagnoses; our ultimate aim is to increase access to evidence and reduce diagnostic errors. In particular, we propose a Neural Additive Model to make predictions backed by evidence with individualized risk estimates at time-points where clinicians are still uncertain, aiming to specifically mitigate delays in diagnosis and errors stemming from an incomplete differential. To train such a model, it is necessary to infer temporally fine-grained retrospective labels of eventual "true" diagnoses. We do so with LLMs, to ensure that the input text is from before a confident diagnosis can be made. We use an LLM to retrieve an initial pool of evidence, but then refine this set of evidence according to correlations learned by the model. We conduct an in-depth evaluation of the usefulness of our approach by simulating how it might be used by a clinician to decide between a pre-defined list of differential diagnoses.

Somin Wadhwa, Chantal Shaib, Silvio Amir et al.

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.

Hiba Ahsan, Arnab Sen Sharma, Silvio Amir et al.

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 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.

Chantal Shaib, Joe Barrow, Alexa F. Siu et al.

Modern instruction-tuned models have become highly capable in text generation tasks such as summarization, and are expected to be released at a steady pace. In practice one may now wish to choose confidently, but with minimal effort, the best performing summarization model when applied to a new domain or purpose. In this work, we empirically investigate the test sample size necessary to select a preferred model in the context of news summarization. Empirical results reveal that comparative evaluation converges quickly for both automatic and human evaluation, with clear preferences for a system emerging from under 100 examples. The human preference data allows us to quantify how well automatic scores can reproduce preference rankings across a variety of downstream summarization tasks. We find that, while automatic metrics are stable at smaller sample sizes, only some automatic metrics are able to moderately predict model win rates according to human preference.

Hye Sun Yun, Karen Y. C. Zhang, Ramez Kouzy et al.

Medical research faces well-documented challenges in translating novel treatments into clinical practice. Publishing incentives encourage researchers to present "positive" findings, even when empirical results are equivocal. Consequently, it is well-documented that authors often spin study results, especially in article abstracts. Such spin can influence clinician interpretation of evidence and may affect patient care decisions. In this study, we ask whether the interpretation of trial results offered by Large Language Models (LLMs) is similarly affected by spin. This is important since LLMs are increasingly being used to trawl through and synthesize published medical evidence. We evaluated 22 LLMs and found that they are across the board more susceptible to spin than humans. They might also propagate spin into their outputs: We find evidence, e.g., that LLMs implicitly incorporate spin into plain language summaries that they generate. We also find, however, that LLMs are generally capable of recognizing spin, and can be prompted in a way to mitigate spin's impact on LLM outputs.

Chantal Shaib, Tuhin Chakrabarty, Diego Garcia-Olano et al.

AI "slop" is an increasingly popular term used to describe low-quality AI-generated text, but there is currently no agreed upon definition of this term nor a means to measure its occurrence. In this work, we develop a taxonomy of "slop" through interviews with experts in NLP, writing, and philosophy, and propose a set of interpretable dimensions for its assessment in text. Through span-level annotation, we find that binary "slop" judgments are (somewhat) subjective, but such determinations nonetheless correlate with latent dimensions such as coherence and relevance. Our framework can be used to evaluate AI-generated text in both detection and binary preference tasks, potentially offering new insights into the linguistic and stylistic factors that contribute to quality judgments.

Millicent Li, Alberto Mario Ceballos Arroyo, Giordano Rogers et al.

Recent interpretability methods have proposed to translate LLM internal representations into natural language descriptions using a second verbalizer LLM. This is intended to illuminate how the target model represents and operates on inputs. But do such activation verbalization approaches actually provide privileged knowledge about the internal workings of the target model, or do they merely convey information about its inputs? We critically evaluate popular verbalization methods across datasets used in prior work and find that they can succeed at benchmarks without any access to target model internals, suggesting that these datasets may not be ideal for evaluating verbalization methods. We then run controlled experiments which reveal that verbalizations often reflect the parametric knowledge of the verbalizer LLM which generated them, rather than the knowledge of the target LLM whose activations are decoded. Taken together, our results indicate a need for targeted benchmarks and experimental controls to rigorously assess whether verbalization methods provide meaningful insights into the operations of LLMs.

Somin Wadhwa, Silvio Amir, Byron C. Wallace

Model distillation typically focuses on behavioral mimicry, where a student model is trained to replicate a teacher's output while treating its internal computations as a black box. In this work we propose an alternative approach: Distilling the underlying computational mechanisms implemented by a teacher model. Specifically, we propose circuit distillation, which introduces an objective to align internal representations between analogous circuit components in teacher and student models. We propose a method to match ``functionally correspondent'' circuit components and introduce a loss reflecting similarities between the representations that these induce. We evaluate circuit distillation on entity tracking and theory of mind (ToM) tasks using models from the Llama3 family. Our results demonstrate that circuit distillation outperforms standard distillation, successfully transferring algorithmic capabilities by adjusting only a small, targeted subset of student model parameters. This work establishes the feasibility of transferring mechanisms, which may in turn allow for efficient distillation of targeted teacher capabilities via interpretable and controllable internal student mechanisms.

Chantal Shaib, Vinith M. Suriyakumar, Levent Sagun et al.

For an LLM to correctly respond to an instruction it must understand both the semantics and the domain (i.e., subject area) of a given task-instruction pair. However, syntax can also convey implicit information Recent work shows that syntactic templates -- frequent sequences of Part-of-Speech (PoS) tags -- are prevalent in training data and often appear in model outputs. In this work we characterize syntactic templates, domain, and semantics in task-instruction pairs. We identify cases of spurious correlations between syntax and domain, where models learn to associate a domain with syntax during training; this can sometimes override prompt semantics. Using a synthetic training dataset, we find that the syntactic-domain correlation can lower performance (mean 0.51 +/- 0.06) on entity knowledge tasks in OLMo-2 models (1B-13B). We introduce an evaluation framework to detect this phenomenon in trained models, and show that it occurs on a subset of the FlanV2 dataset in open (OLMo-2-7B; Llama-4-Maverick), and closed (GPT-4o) models. Finally, we present a case study on the implications for safety finetuning, showing that unintended syntactic-domain correlations can be used to bypass refusals in OLMo-2-7B Instruct and GPT-4o. Our findings highlight two needs: (1) to explicitly test for syntactic-domain correlations, and (2) to ensure syntactic diversity in training data, specifically within domains, to prevent such spurious correlations.

Sebastian Joseph, Lily Chen, Barry Wei et al.

Technological progress has led to concrete advancements in tasks that were regarded as challenging, such as automatic fact-checking. Interest in adopting these systems for public health and medicine has grown due to the high-stakes nature of medical decisions and challenges in critically appraising a vast and diverse medical literature. Evidence-based medicine connects to every individual, and yet the nature of it is highly technical, rendering the medical literacy of majority users inadequate to sufficiently navigate the domain. Such problems with medical communication ripens the ground for end-to-end fact-checking agents: check a claim against current medical literature and return with an evidence-backed verdict. And yet, such systems remain largely unused. To understand this, we present the first study examining how clinical experts verify real claims from social media by synthesizing medical evidence. In searching for this upper-bound, we reveal fundamental challenges in end-to-end fact-checking when applied to medicine: Difficulties connecting claims in the wild to scientific evidence in the form of clinical trials; ambiguities in underspecified claims mixed with mismatched intentions; and inherently subjective veracity labels. We argue that fact-checking should be approached and evaluated as an interactive communication problem, rather than an end-to-end process.

Chantal Shaib, Yanai Elazar, Junyi Jessy Li et al.

Recent work on evaluating the diversity of text generated by LLMs has focused on word-level features. Here we offer an analysis of syntactic features to characterize general repetition in models, beyond frequent n-grams. Specifically, we define syntactic templates and show that models tend to produce templated text in downstream tasks at a higher rate than what is found in human-reference texts. We 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 processes such as RLHF. This connection to 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.

Somin Wadhwa, Silvio Amir, Byron C. Wallace

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.

Somin Wadhwa, Adit Krishnan, Runhui Wang et al.

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.

Kundan Krishna, Prakhar Gupta, Sanjana Ramprasad et al.

While the NLP community has produced numerous summarization benchmarks, none provide the rich annotations required to simultaneously address many important problems related to control and reliability. We introduce a Wikipedia-derived benchmark, complemented by a rich set of crowd-sourced annotations, that supports $8$ interrelated tasks: (i) extractive summarization; (ii) abstractive summarization; (iii) topic-based summarization; (iv) compressing selected sentences into a one-line summary; (v) surfacing evidence for a summary sentence; (vi) predicting the factual accuracy of a summary sentence; (vii) identifying unsubstantiated spans in a summary sentence; (viii) correcting factual errors in summaries. We compare various methods on this benchmark and discover that on multiple tasks, moderately-sized fine-tuned models consistently outperform much larger few-shot prompted language models. For factuality-related tasks, we also evaluate existing heuristics to create training data and find that training on them results in worse performance than training on $20\times$ less human-labeled data. Our articles draw from $6$ domains, facilitating cross-domain analysis. On some tasks, the amount of training data matters more than the domain where it comes from, while for other tasks training specifically on data from the target domain, even if limited, is more beneficial.

Lucy Lu Wang, Yulia Otmakhova, Jay DeYoung et al.

Evaluating multi-document summarization (MDS) quality is difficult. This is especially true in the case of MDS for biomedical literature reviews, where models must synthesize contradicting evidence reported across different documents. Prior work has shown that rather than performing the task, models may exploit shortcuts that are difficult to detect using standard n-gram similarity metrics such as ROUGE. Better automated evaluation metrics are needed, but few resources exist to assess metrics when they are proposed. Therefore, we introduce a dataset of human-assessed summary quality facets and pairwise preferences to encourage and support the development of better automated evaluation methods for literature review MDS. We take advantage of community submissions to the Multi-document Summarization for Literature Review (MSLR) shared task to compile a diverse and representative sample of generated summaries. We analyze how automated summarization evaluation metrics correlate with lexical features of generated summaries, to other automated metrics including several we propose in this work, and to aspects of human-assessed summary quality. We find that not only do automated metrics fail to capture aspects of quality as assessed by humans, in many cases the system rankings produced by these metrics are anti-correlated with rankings according to human annotators.