Papers by Aaron Mueller
Language model acceptability judgements are not always robust to context (2023.acl-long)
Copied to clipboard
Koustuv Sinha, Jon Gauthier, Aaron Mueller, Kanishka Misra, Keren Fuentes, Roger Levy, Adina Williams
| Challenge: | a recent study found that models prefer acceptable inputs over acceptable ones. |
| Approach: | They find that model judgements are generally robust when placed in randomly sampled linguistic contexts, but unstable when contexts match the test stimuli in syntactic structure. |
| Outcome: | The proposed model performance improves when contexts match syntactic structure, and declines when they are unacceptable. |
CRISP: Persistent Concept Unlearning via Sparse Autoencoders (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work has explored sparse autoencoders (SAEs) to perform precise interventions on monosemantic features, but most SAE-based methods operate at inference time, which does not create persistent changes in the model’s parameters. |
| Approach: | They propose a parameter-efficient method for persistent concept unlearning using SAEs that automatically identifies salient SAE features across multiple layers and suppresses their activations. |
| Outcome: | The proposed method outperforms previous methods on safety-critical unlearning tasks from the WMDP benchmark, successfully removing harmful knowledge while preserving general and in-domain capabilities. |
Coloring the Blank Slate: Pre-training Imparts a Hierarchical Inductive Bias to Sequence-to-sequence Models (2022.findings-acl)
Copied to clipboard
| Challenge: | Sequence-to-sequence models fail to generalize in hierarchy-sensitive manner when performing syntactic transformations. |
| Approach: | They evaluate whether seq2seq models generalize hierarchically on two transformations . they use pre-trained models and their multilingual variants to test their generalization . |
| Outcome: | The proposed models generalize hierarchically on two transformations in English and German. |
SAEs Are Good for Steering – If You Select the Right Features (2025.emnlp-main)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) can learn a decomposition of a model’s latent space by analyzing the input tokens that activate them. |
| Approach: | They propose an unsupervised approach to learn a decomposition of a model’s latent space by analyzing the input tokens that activate them. |
| Outcome: | The proposed approach matches the performance of existing supervised methods by identifying features with low output scores and identifying them with input and output scores. |
Improving the OOD Performance of Closed-Source LLMs on NLI Through Strategic Data Selection (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to improve robustness require changing the fine-tuning process or large-scale data augmentation, which are infeasible or cost prohibitive for closed-source models. |
| Approach: | They propose to prioritize more complex examples or replace existing training examples with LLM-generated data to improve performance on OOD NLI datasets. |
| Outcome: | The proposed methods improve performance on difficult OOD datasets while training with synthetic data leads to substantial improvements on easier OOD data. |
Bernice: A Multilingual Pre-trained Encoder for Twitter (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing language models for Twitter are monolingual, adapted from other domains, or trained on limited amount of in-domain data. |
| Approach: | They propose a multilingual RoBERTa language model that is trained from scratch on 2.5 billion tweets with a custom tweet-focused tokenizer. |
| Outcome: | The proposed model outperforms or matches models trained on monolingual and multilingual tweets on a variety of benchmarks and is more efficient compute- and data-wise to train completely on in-domain data with a specialized domain-specific tokenizer. |
Measuring Mechanistic Independence: Can Bias Be Removed Without Erasing Demographics? (2026.eacl-long)
Copied to clipboard
| Challenge: | Using multi-task evaluation, we examine how independent demographic bias mechanisms are from general demographic recognition in language models. |
| Approach: | They compare attribution-based and correlation-based methods for locating bias features in language models to find out which features are independent from general demographic recognition. |
| Outcome: | The proposed method reduces bias without degrading recognition performance. |
The Johns Hopkins University Bible Corpus: 1600+ Tongues for Typological Exploration (2020.lrec-1)
Copied to clipboard
Arya D. McCarthy, Rachel Wicks, Dylan Lewis, Aaron Mueller, Winston Wu, Oliver Adams, Garrett Nicolai, Matt Post, David Yarowsky
| Challenge: | Our corpus spans 1611 diverse written languages, with constituents of more than 90 language families. |
| Approach: | They propose to scrape and merge online resources and merge them with existing corpora to create a verse-parallel scheme for all translations. |
| Outcome: | The results show that the Bible provides high coverage of core vocabulary. |
Causal Analysis of Syntactic Agreement Mechanisms in Neural Language Models (2021.acl-long)
Copied to clipboard
| Challenge: | Targeted syntactic evaluations have demonstrated the ability of language models to perform subject-verb agreement given difficult contexts. |
| Approach: | They apply causal mediation analysis to pre-trained neural language models to investigate their models' preferences for grammatical inflections and whether neurons process subject-verb agreement similarly across sentences with different syntactic structures. |
| Outcome: | The proposed model can predict correct token from grammatically minimally different continuations with high accuracy even in difficult contexts. |
How to Plant Trees in Language Models: Data and Architectural Effects on the Emergence of Syntactic Inductive Biases (2023.acl-long)
Copied to clipboard
| Challenge: | a recent study found that pre-training can teach language models to rely on hierarchical syntactic features . aaron ramirez: we find that pretraining on simpler language induces a hierarchic bias . |
| Approach: | They find that pre-training can teach language models to rely on hierarchical syntactic features . authors: this suggests that in cognitively plausible language acquisition settings, models may be more data-efficient . |
| Outcome: | a recent study shows that pre-training can teach language models to rely on hierarchical features . the findings suggest that in plausible language acquisition settings, language models may be more data-efficient than previously thought . |
What Do NLP Researchers Believe? Results of the NLP Community Metasurvey (2023.acl-long)
Copied to clipboard
Julian Michael, Ari Holtzman, Alicia Parrish, Aaron Mueller, Alex Wang, Angelica Chen, Divyam Madaan, Nikita Nangia, Richard Yuanzhe Pang, Jason Phang, Samuel R. Bowman
| Challenge: | Getting sociological beliefs wrong can slow research and lead to wasted effort, missed opportunities, and needless fights. |
| Approach: | They present the results of the NLP Community Metasurvey, run from May to June 2022. |
| Outcome: | The NLP community metasurvey elicited opinions on controversial issues from May to June 2022. |
Crosscoding Through Time: Tracking Emergence & Consolidation Of Linguistic Representations Throughout LLM Pretraining (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models learn non-trivial abstractions during pretraining, but it is not well understood when and how these specific linguistic abilities emerge. |
| Approach: | They propose a method to track the evolution of linguistic features during pretraining by using sparse crosscoders to discover and align features across model checkpoints. |
| Outcome: | The proposed approach can detect features emergence, maintenance, and discontinuation during training stages. |
Meta-training with Demonstration Retrieval for Efficient Few-shot Learning (2023.findings-acl)
Copied to clipboard
| Challenge: | Large language models have impressive fewshot performance on many NLP tasks and domains. |
| Approach: | They propose a meta-training approach that uses demonstration retrieval to train parameter-efficient models that generalize well on a larger variety of tasks. |
| Outcome: | The proposed approach outperforms many parameter-efficient methods on QA, NLI, and text classification tasks. |
Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages (2025.naacl-long)
Copied to clipboard
| Challenge: | In the brains of human bilinguals, syntax processing may occur in similar regions for their first and second language, depending on factors like when the second language was learned and language proficiency. |
| Approach: | They propose to use sparse autoencoders to train Llama-3-8B and Aya-23-8B models to train multilingual models that share morphsyntactic representations of grammatical concepts. |
| Outcome: | The proposed model can predict plural verbs in different languages by activating the same plural feature. |
Characterizing the Role of Similarity in the Property Inferences of Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | Property inheritance is a phenomenon where novel properties are projected from higher level categories to lower level ones. |
| Approach: | They investigate how LMs perform property inheritance with behavioral and causal analysis experiments. |
| Outcome: | The results provide insight into the conceptual structure of language models and may suggest new psycholinguistic experiments for human subjects. |
Quantity doesn’t buy quality syntax with neural language models (D19-1)
Copied to clipboard
| Challenge: | Recurrent neural network language models can learn to predict upcoming words with remarkably low perplexity . but in syntactically complex contexts, they often assign unexpectedly high probabilities to ungrammatical words . |
| Approach: | They investigate whether recurrent neural networks can learn to predict upcoming words with remarkably low perplexity. |
| Outcome: | The proposed models perform worse than GPT and BERT in some constructions than LSTMs in other contexts. |
In-context Learning Generalizes, But Not Always Robustly: The Case of Syntax (2024.naacl-long)
Copied to clipboard
| Challenge: | In-context learning is a common method for teaching large language models new tasks . given labeled examples in the input context, the model learns to perform the task without weight updates. |
| Approach: | They examine whether models guided via ICL infer the underlying structure of the task defined by the context or rely on superficial heuristics that only generalize to identically distributed examples. |
| Outcome: | The proposed model generalizes syntactically or linearly on out-of-distribution examples . the proposed model is able to generalize better on pre-trained models . |
Position-aware Automatic Circuit Discovery (2025.acl-long)
Copied to clipboard
| Challenge: | Existing circuit analysis methods treat component positions as equally relevant across input positions, limiting their ability to capture cross-positional interactions or mechanisms that vary across positions. |
| Approach: | They propose to extend edge attribution patching to differentiate between token positions and introduce a dataset schema to enable position-aware circuit discovery in datasets with variable length examples. |
| Outcome: | The proposed approach enables position-aware circuit discovery even on tasks with variable length examples. |
Incremental Sentence Processing Mechanisms in Autoregressive Transformer Language Models (2025.naacl-long)
Copied to clipboard
| Challenge: | a recent study has found LMs focus on processing syntactic relations, but not structural information. |
| Approach: | They examine the mechanisms underlying garden path sentence processing in LMs . they use sparse autoencoders to identify interpretable features that determine which continuation . |
| Outcome: | The proposed model lacks syntactic features and shallow heuristics to perform incremental sentence processing. |
Modeling Color Terminology Across Thousands of Languages (D19-1)
Copied to clipboard
| Challenge: | Existing studies on what constitutes a "basic" color term and its acquisition sequence are flawed . a pan-lingual approach may reveal general color trends more reliably than smaller datasets. |
| Approach: | They propose to operationalize and critique the Berlin and Kay color term hypotheses . they use 14 empirically-grounded computational linguistic metrics to analyze cross-linguistic data . |
| Outcome: | The proposed measures correlate strongly with the Berlin and Kay color term partition and their hypothesized universal acquisition sequence. |
An Analysis of Massively Multilingual Neural Machine Translation for Low-Resource Languages (2020.lrec-1)
Copied to clipboard
| Challenge: | In this study, we explore massively multilingual low-resource neural machine translation. |
| Approach: | They propose to use Bible translations to train models with up to 1,107 source languages and create multilingual corpora varying the number and relatedness of source languages. |
| Outcome: | The proposed approach is highly language-specific and can be tailored to the source language and its typology. |
Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate (2026.acl-long)
Copied to clipboard
| Challenge: | Multi-agent debate is compute-intensive and requires long transcripts before answering questions. |
| Approach: | They propose a framework that distills multi-agent debate into a single LLM by combining debate structure learning with internalization via dynamic reward scheduling and length clipping. |
| Outcome: | The proposed model matches or exceeds explicit multi-agent debate performance using 93% fewer tokens across multiple models and benchmarks. |
Fine-tuning Encoders for Improved Monolingual and Zero-shot Polylingual Neural Topic Modeling (2021.naacl-main)
Copied to clipboard
| Challenge: | Topic models can augment or replace bag-of-words inputs with pre-trained transformer-based word prediction models. |
| Approach: | They propose several methods for fine-tuning encoders to improve both monolingual and zero-shot polylingual topic modeling. |
| Outcome: | The proposed methods improve both monolingual and zero-shot polylingual topic modeling. |
Cross-Linguistic Syntactic Evaluation of Word Prediction Models (2020.acl-main)
Copied to clipboard
| Challenge: | A range of studies have concluded that neural word prediction models can distinguish grammatical from ungrammatically sentences with high accuracy. |
| Approach: | They propose to use CLAMS to evaluate LSTM and multilingual BERT models. |
| Outcome: | The proposed model can learn syntax on English, French, German, Hebrew and Russian, and LSTM language models on multilingual and multilingual models. |
Fine-grained Morphosyntactic Analysis and Generation Tools for More Than One Thousand Languages (2020.lrec-1)
Copied to clipboard
| Challenge: | Using morphosyntactic tools, we train and distribute tools for approximately one thousand languages. |
| Approach: | They train and distribute morphosyntactic tools for approximately one thousand languages. |
| Outcome: | The results show that the tools generalize well across rare and common forms alike. |
Label Semantic Aware Pre-training for Few-shot Text Classification (2022.acl-long)
Copied to clipboard
| Challenge: | Existing models for text classification use label semantics but few studies have attempted to give models access to informative representations of labels. |
| Approach: | They propose to use label semantics to train generative models by performing secondary pre-training on labeled sentences from a variety of domains. |
| Outcome: | The proposed approach improves generalization and data efficiency of text classification systems while maintaining comparable performance to state-of-the-art models. |
From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts? (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to evaluate features disentangle concepts from activations of neural networks are limited by their quality . current methods for concept identification and steering are sparse autoencoders, but they are not reliable. |
| Approach: | They propose to evaluate how well featurization methods disentangle one concept from another . they use sentiment, domain, voice, and tense to steer these features . |
| Outcome: | The proposed evaluations show that featurization methods are insufficient to establish steering selectivity . the results suggest that steering a feature affects many concepts despite a near absence of interaction effects. |