Papers by Percy Liang
Backpack Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | We train a 170Mparameter Backpack language model on OpenWebText, matching the loss of a 6Bparameter Transformer. |
| Approach: | They propose a neural architecture that learns multiple non-contextual sense vectors for each word in a vocabulary and represents a word as a context-dependent, non-negative linear combination of sense vector. |
| Outcome: | The proposed model outperforms a GPT-2's word embeddings on lexical similarity evaluations and can be used to perform controllable text generation and debiasing. |
Task-Oriented Dialogue as Dataflow Synthesis (2020.tacl-1)
Copied to clipboard
Jacob Andreas, John Bufe, David Burkett, Charles Chen, Josh Clausman, Jean Crawford, Kate Crim, Jordan DeLoach, Leah Dorner, Jason Eisner, Hao Fang, Alan Guo, David Hall, Kristin Hayes, Kellie Hill, Diana Ho, Wendy Iwaszuk, Smriti Jha, Dan Klein, Jayant Krishnamurthy, Theo Lanman, Percy Liang, Christopher H. Lin, Ilya Lintsbakh, Andy McGovern, Aleksandr Nisnevich, Adam Pauls, Dmitrij Petters, Brent Read, Dan Roth, Subhro Roy, Jesse Rusak, Beth Short, Div Slomin, Ben Snyder, Stephon Striplin, Yu Su, Zachary Tellman, Sam Thomson, Andrei Vorobev, Izabela Witoszko, Jason Wolfe, Abby Wray, Yuchen Zhang, Alexander Zotov
| Challenge: | Existing approaches to task-oriented dialogue represent dialogue state as a dataflow graph . microsoft's SMCalFlow dataset features complex dialogues about events, weather, places, and people . |
| Approach: | They propose a dataflow graph-based dialogue agent that maps each user utterance to a program that extends this graph. |
| Outcome: | The proposed framework improves representability and predictability in natural dialogues . it uses dataflow graphs and metacomputation to map user intents to a program . |
LinkBERT: Pretraining Language Models with Document Links (2022.acl-long)
Copied to clipboard
| Challenge: | Existing language model pretraining methods do not capture dependencies or knowledge that span across documents. |
| Approach: | They propose a language model pretraining method that leverages links between documents . they use masked language modeling and document relation prediction to model LMs . |
| Outcome: | The proposed method outperforms existing methods on downstream tasks across two domains. |
Unifying Human and Statistical Evaluation for Natural Language Generation (N19-1)
Copied to clipboard
| Challenge: | Human evaluation captures quality but fails to capture diversity . statistical evaluation fails to catch models that plagiarize from training set . |
| Approach: | They propose a framework which evaluates both diversity and quality based on the optimal error rate of predicting whether a sentence is human-generated. |
| Outcome: | The proposed framework evaluates diversity and quality on summarization and chit-chat dialogue. |
On the Importance of Adaptive Data Collection for Extremely Imbalanced Pairwise Tasks (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Recent datasets heuristically choose examples to ensure label balance . state-of-the-art models trained on QQP and WikiQA have only 2.4% average precision . |
| Approach: | They show that recent datasets heuristically choose examples to ensure label balance . they instead use active learning to retrieve uncertain points from a large pool of unlabeled utterance pairs . |
| Outcome: | The proposed model improves on QQP and WikiQA by using more informative negative examples. |
Mapping natural language commands to web elements (D18-1)
Copied to clipboard
| Challenge: | a dataset of over 50,000 natural language commands captures various phenomena, including functional references, relational reasoning, and visual reasoning. |
| Approach: | They propose a task that requires the user to choose the correct element on a web page . they use a dataset of over 50,000 natural language commands to map these to web pages . |
| Outcome: | The proposed task can be viewed as a reference game based on a dataset of over 50,000 natural language commands . |
Prefix-Tuning: Optimizing Continuous Prompts for Generation (2021.acl-long)
Copied to clipboard
| Challenge: | Fine-tuning is the prevalent paradigm for using large pretrained language models for downstream tasks, but it requires updating and storing all the parameters of the LM. |
| Approach: | They propose a lightweight alternative to fine-tuning for natural language generation tasks that optimizes a sequence of continuous vectors, which they call the prefix. |
| Outcome: | The proposed approach outperforms fine-tuning in the full data setting and extrapolates better to examples with topics that are unseen during training. |
LM-Critic: Language Models for Unsupervised Grammatical Error Correction (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent work casts GEC as a translation problem using encoder-decoder models to map bad (ungrammatical) sentences into good (grammatically) sentences. |
| Approach: | They propose to use a pretrained language model to define an LM-Critic that judges a sentence to be grammatical if the LM assigns it a higher probability than its local perturbations. |
| Outcome: | The proposed method outperforms existing methods in both the unsupervised and supervised setting. |
Do Question Answering Modeling Improvements Hold Across Benchmarks? (2023.acl-long)
Copied to clipboard
| Challenge: | a new study finds that human-constructed and downsampled benchmarks hold more concurrence than downsampled benchmarks. |
| Approach: | They propose to measure concurrence between two QA benchmarks on a set of 20 models . they find that human-constructed benchmarks have high concurrence amongst themselves . |
| Outcome: | The proposed models hold broadly across the diverse landscape of question answering (QA) benchmarks. |
Know What You Don’t Know: Unanswerable Questions for SQuAD (P18-2)
Copied to clipboard
| Challenge: | Existing datasets focus on answerable questions or use automatically generated unanswerable questions that are easy to identify. |
| Approach: | They propose a dataset that combines the Stanford Question Answering Dataset with 50,000 unanswerable questions written by crowdworkers to look similar to answerable ones. |
| Outcome: | The proposed dataset looks similar to answerable questions on crowd-written questions . strong neural system that gets 86% F1 on SQuAD achieves only 66% F1. |
Contrastive Decoding: Open-ended Text Generation as Optimization (2023.acl-long)
Copied to clipboard
Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, Mike Lewis
| Challenge: | Using a language model, maximum probability is a poor decoding objective because it produces short and repetitive text. |
| Approach: | They propose a reliable decoding approach that optimizes a contrastive objective subject to a plausibility constraint. |
| Outcome: | The proposed approach outperforms four strong decoding algorithms in automatic and human evaluations across wikipedia, news and story domains. |
The price of debiasing automatic metrics in natural language evalaution (P18-1)
Copied to clipboard
| Challenge: | Existing methods to evaluate natural language systems are expensive and expensive. |
| Approach: | They propose to combine automatic metrics with human judgment to obtain an unbiased estimator at lower cost than human evaluation alone. |
| Outcome: | The proposed estimator reduces the cost of evaluating summarization and open-response questions by 7-13%. |
s1: Simple test-time scaling (2025.emnlp-main)
Copied to clipboard
Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candes, Tatsunori Hashimoto
| Challenge: | OpenAI’s o1 model showed this capability but did not publicly share its methodology, leading to many replication efforts. |
| Approach: | They curate a small dataset s1K with 1,000 reasoning questions based on three criteria we validate through ablations: difficulty, diversity, and quality. |
| Outcome: | The proposed model exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24). |
Truncation Sampling as Language Model Desmoothing (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Long samples of text from neural language models can be of poor quality. |
| Approach: | They propose to think of a neural language model as a mixture of k and a true distribution that avoids infinite perplexity. |
| Outcome: | The proposed methods generate more plausible long documents according to humans and break out of repetition. |
Distributionally Robust Language Modeling (D19-1)
Copied to clipboard
| Challenge: | Language models are generally trained on data spanning a wide range of topics but might be applied to an unknown target distribution. |
| Approach: | They propose a distributionally robust optimization procedure which minimizes the loss of the model over the worst-case mixture of topics with sufficient overlap with the training distribution. |
| Outcome: | The proposed method reduces the loss of the model over the worst-case mixture of topics with sufficient overlap with the training distribution. |
Enabling Language Models to Fill in the Blanks (2020.acl-main)
Copied to clipboard
| Challenge: | Infilling is the task of predicting missing spans of text at any position in a document. |
| Approach: | They propose a framework which can be used to infill entire sentences . they train off-the-shelf LMs on sequences containing concatenation of masked text . |
| Outcome: | The proposed approach can infill entire sentences on short stories, scientific abstracts, and lyrics. |
Beyond Positive Scaling: How Negation Impacts Scaling Trends of Language Models (2023.findings-acl)
Copied to clipboard
Yuhui Zhang, Michihiro Yasunaga, Zhengping Zhou, Jeff Z. HaoChen, James Zou, Percy Liang, Serena Yeung
| Challenge: | Recent studies show that some tasks exhibit inverse scaling, or U-shaped scaling, where the performance degrades as models are scaled up. |
| Approach: | They propose a task that asks questions with negation to show positive scaling . they hypothesize that solving NeQA depends on question answering and negation understanding . |
| Outcome: | The proposed task can exhibit inverse scaling, U-shaped scaling, or positive scaling, and the scaling trends shift as the task is more powerful. |
RNNs can generate bounded hierarchical languages with optimal memory (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have shown that RNNs can efficiently generate bounded hierarchical languages with high syntactic fidelity, but their success is not well-understood theoretically. |
| Approach: | They propose a language of well-nested brackets and m-bounded nesting depth . they prove that an RNN with O(m log k) hidden units suffices, an exponential reduction in memory, by an explicit construction. |
| Outcome: | The proposed language is well-nested brackets and has m-bounded nesting depth . it shows that an RNN with O(m log k) hidden units suffices, an exponential reduction in memory, by an explicit construction. |
QuAC: Question Answering in Context (D18-1)
Copied to clipboard
Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen-tau Yih, Yejin Choi, Percy Liang, Luke Zettlemoyer
| Challenge: | a dataset for Question Answering in Context contains 14K information-seeking QA dialogs . questions are often more open-ended, unanswerable, or only meaningful within the dialog context . |
| Approach: | They propose a dataset for Question Answering in Context that contains 14K dialogs . they use a student to ask questions about a Wikipedia section and a teacher to answer them . |
| Outcome: | The proposed dataset underperforms humans in a number of reference models . the dataset contains 14K information-seeking dialogs over sections from Wikipedia . |
Are Sample-Efficient NLP Models More Robust? (2023.acl-short)
Copied to clipboard
| Challenge: | Recent work in image classification and extractive question answering have observed that pre-trained models trained on less in-distribution data have better out-of-distortion performance. |
| Approach: | They conduct a large empirical study to investigate the relationship between sample efficiency and robustness. |
| Outcome: | The results show that pre-trained models with lower sample efficiency perform better on some tasks but not others. |
ExpBERT: Representation Engineering with Natural Language Explanations (2020.acl-main)
Copied to clipboard
| Challenge: | Existing models with explanations use semantic parsers to interpret explanations, but they can only parse low-level statements like "'wife' appears between o 1 and 2 " |
| Approach: | They propose to use natural language explanations to specify inductive biases . they use BERT fine-tuned on MultiNLI to "interpret" the explanations with respect to the input sentence . |
| Outcome: | The proposed method matches a baseline with less labeled data and improves on the baseline with the same amount of labele. |
Pun Generation with Surprise (N19-1)
Copied to clipboard
| Challenge: | In this paper, we explore creative generation with a focus on puns. |
| Approach: | They propose an unsupervised approach to generating puns using lots of raw text and a surprisal principle. |
| Outcome: | The proposed approach generates puns 30% of the time, doubles the neural generation baseline. |
Beyond a Single Extractor: Re-thinking HTML-to-Text Extraction for LLM Pre-training (2026.findings-eacl)
Copied to clipboard
Jeffrey Li, Joshua P Gardner, Doug Kang, Fangping Shi, Karanjeet Singh, Chun-Liang Li, Herumb Shandilya, David Leo Wright Hall, Oncel Tuzel, Percy Liang, Ludwig Schmidt, Hadi Pouransari, Fartash Faghri
| Challenge: | Existing open-source datasets predominantly apply a single fixed extractor to all webpages. |
| Approach: | They propose to take a Union over different extractors to improve model performance . they show that extractor choice can significantly impact downstream task performance based on content type . |
| Outcome: | The proposed approach can increase the token yield of DCLM-Baseline by 71% while maintaining benchmark performance. |
Training Classifiers with Natural Language Explanations (P18-1)
Copied to clipboard
| Challenge: | a semantic parser converts explanations into programmatic labeling functions . a standard protocol for obtaining a labeled dataset provides only one bit of information per example . |
| Approach: | They propose a framework where an annotator provides an explanation for each labeling decision . they use a semantic parser to convert these explanations into programmatic labeling functions . |
| Outcome: | The proposed framework trains classifiers faster by providing explanations instead of labels . the proposed framework is based on a rule-based semantic parser . |
Robust Encodings: A Framework for Combating Adversarial Typos (2020.acl-main)
Copied to clipboard
| Challenge: | Existing procedures to defend against adversarial perturbations of inputs are heuristic in nature and susceptible to stronger attacks. |
| Approach: | They propose a framework that provides robustness without compromises on model architecture. |
| Outcome: | The proposed framework achieves robustness on six tasks while incompatible with state-of-the-art models like BERT. |
Selective Question Answering under Domain Shift (2020.acl-main)
Copied to clipboard
| Challenge: | Deployed question answering (QA) models need to know when to abstain from answering questions that diverge from their training data. |
| Approach: | They propose a selective question answering under domain shift in which a QA model is tested on a mixture of in-domain and out-of-domain data and must answer (i.e., not abstain on) as many questions as possible. |
| Outcome: | The proposed method answers 56% of questions while maintaining 80% accuracy. |
Evaluating Verifiability in Generative Search Engines (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing generative search engines are rapidly gaining users, according to a new study . existing systems are poorly cited and lack reliability, a study finds . |
| Approach: | They conduct human evaluations of four popular generative search engines . they find that existing generative engines are fluent and appear informative . |
| Outcome: | The results show that existing generative search engines are not reliable and often contain unsupported statements and inaccurate citations. |
QA-GNN: Reasoning with Language Models and Knowledge Graphs for Question Answering (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing question answering systems lack the ability to access relevant knowledge and reason over it. |
| Approach: | They propose a model that uses KGs to identify relevant knowledge in QA contexts and perform joint reasoning over them. |
| Outcome: | The proposed model improves on the CommonsenseQA and OpenBookQA datasets and performs interpretable and structured reasoning. |
Conditional probing: measuring usable information beyond a baseline (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for probing representations are limited to predicting part-of-speech . current methods cannot detect when a representation is predictive of just aspects of part- of-seech not explainable by the word identity. |
| Approach: | They propose to condition on the information in a baseline representation to test whether it is predictive of part-of-speech. |
| Outcome: | The proposed method is based on a theory of usable information called V-information and conditions on the information in the baseline. |
Designing and Interpreting Probes with Control Tasks (D19-1)
Copied to clipboard
| Challenge: | Existing studies on supervised models to predict properties from representations have shown high accuracy on a range of linguistic tasks. |
| Approach: | They propose control tasks which associate word types with random outputs to complement linguistic tasks by construction . they find that popular probes on ELMo representations are not selective . |
| Outcome: | The proposed tasks associate word types with random outputs to complement linguistic tasks. |
Textual Analogy Parsing: What’s Shared and What’s Compared among Analogous Facts (D18-1)
Copied to clipboard
| Challenge: | Existing methods to extract information from text do not capture disparity between demographic groups. |
| Approach: | They propose a task of Textual Analogy Parsing to model higher-order meanings by comparing poverty rates between different demographic groups. |
| Outcome: | The proposed model can be used to generate graphs from quantitative text. |
Delete, Retrieve, Generate: a Simple Approach to Sentiment and Style Transfer (N18-1)
Copied to clipboard
| Challenge: | Previous work using adversarial methods has struggled to produce high-quality outputs. |
| Approach: | They propose a method that transforms a sentence to alter a specific attribute while preserving its attribute-independent content. |
| Outcome: | The proposed method generates grammatical and appropriate responses on 22% more inputs than the best previous system, averaged over three attribute transfer datasets. |
Swords: A Benchmark for Lexical Substitution with Improved Data Coverage and Quality (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing benchmarks for lexical substitution depend on human recall as the only source of data, authors say . existing benchmarks lack coverage of the appropriate substitutes that would be most helpful to humans . |
| Approach: | They propose a benchmark for lexical substitution to find appropriate substitutes for a target word in context . existing benchmarks depend on human recall as the only source of data, they argue . |
| Outcome: | The new benchmark for lexical substitution uses a context-free thesaurus . it has 3x as many substitutes per target word for the same quality, and substitutes are 1.4x more appropriate . |
Decoupling Strategy and Generation in Negotiation Dialogues (D18-1)
Copied to clipboard
| Challenge: | Recent work on negotiation trains neural models, but their end-to-end nature makes it hard to control their strategy. |
| Approach: | They propose a modular approach that decouples strategy and generation by coarse dialogue acts . they test their approach on a recently proposed DEALORNODEAL game . |
| Outcome: | The proposed approach can decouple strategy and generation without degeneracy. |
Shaping Visual Representations with Language for Few-Shot Classification (2020.acl-main)
Copied to clipboard
| Challenge: | Existing models use natural language descriptions to classify images, but no model uses it for new tasks. |
| Approach: | They propose a model that regularizes visual representations to predict language in a few-shot setting . they propose to use language to improve few- shot visual classification . |
| Outcome: | The proposed model outperforms baseline models in two challenging few-shot domains. |
Certified Robustness to Adversarial Word Substitutions (D19-1)
Copied to clipboard
| Challenge: | State-of-the-art NLP models can be fooled by adversaries that apply seemingly innocuous label-preserving transformations to input text. |
| Approach: | They propose to train models that are provably robust to all word substitutions in a family of label-preserving transformations that can be replaced with a similar word without changing the original sentiment. |
| Outcome: | The proposed models achieve 75% adversarial accuracy on both sentiment analysis and natural language inference on IMDB and SNLI compared to models trained normally and ones trained with data augmentation. |
LawInstruct: A Resource for Studying Language Model Adaptation to the Legal Domain (2025.findings-naacl)
Copied to clipboard
Joel Niklaus, Lucia Zheng, Arya D. McCarthy, Christopher Hahn, Brian M Rosen, Peter Henderson, Daniel E. Ho, Garrett Honke, Percy Liang, Christopher D Manning
| Challenge: | In general, instruction tuning is important for direct user interaction, but the legal domain is underrepresented in typical instruction datasets. |
| Approach: | They aggregate 58 annotated legal datasets and write instructions for each to create LawInstruct. |
| Outcome: | The proposed model improves on LegalBench across all model sizes, but no drop in MMLU. |