Papers by Mike Lewis
Compute Optimal Scaling of Skills: Knowledge vs Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Scaling laws are a critical component of the LLM development pipeline, but little is known about whether the COs of individual skills such as mathematical reasoning, question answering (QA) or coding, align with these APEs. |
| Approach: | They examine knowledge-based QA and code generation to find out whether skill-dependent scaling is an artefact of the pretraining datamix. |
| Outcome: | The proposed scaling laws are skill-dependent, and knowledge and code exhibit fundamental differences in scaling behaviour when corrected for datamix differences. |
Conversational Semantic Parsing (2020.emnlp-main)
Copied to clipboard
Armen Aghajanyan, Jean Maillard, Akshat Shrivastava, Keith Diedrick, Michael Haeger, Haoran Li, Yashar Mehdad, Veselin Stoyanov, Anuj Kumar, Mike Lewis, Sonal Gupta
| Challenge: | Structured representations for task-oriented assistant systems are limited due to the limitations of the representation. |
| Approach: | They propose a semantic representation for task-oriented conversational systems that can represent co-reference and context carryover. |
| Outcome: | The proposed model improves the best results on ATIS, SNIPS, TOP and DSTC2 by up to 5 points for slot-carryover. |
A Dataset for Telling the Stories of Social Media Videos (D18-1)
Copied to clipboard
| Challenge: | Existing datasets focused on pre-selected human activities, whereas social media videos contain a great diversity of topics. |
| Approach: | They propose a large-scale dataset for video description as a new challenge for multi-sentence video description. |
| Outcome: | The proposed dataset contains 20k videos with 123k sentences, temporally aligned to the video. |
Hierarchical Neural Story Generation (P18-1)
Copied to clipboard
| Challenge: | a hierarchical model that generates a premise and then conditions on it creates fluent text . a novel form of model fusion improves the relevance of the story to the prompt . |
| Approach: | They use a hierarchical model that first generates a premise, then transforms it into a text . they use fusion to improve relevance of the story to the prompt and add a gated mechanism to model context . |
| Outcome: | The proposed model improves on strong baselines on automated and human evaluations. |
Measuring and Narrowing the Compositionality Gap in Language Models (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a language model can correctly answer all sub-problems but not generate the overall solution. |
| Approach: | They propose a method that asks itself and then answers follow-up questions to narrow the compositionality gap by reasoning explicitly instead of implicitly. |
| Outcome: | The proposed method improves on chain of thought by asking itself and answering follow-up questions. |
Shortformer: Better Language Modeling using Shorter Inputs (2021.acl-long)
Copied to clipboard
| Challenge: | Existing methods require computationally expensive relative position embeddings. |
| Approach: | They propose two methods that decrease input length to improve perplexity and perplexability. |
| Outcome: | The proposed methods speed up training by a factor of 1.65 and reduce memory usage. |
Joint Verification and Reranking for Open Fact Checking Over Tables (2021.acl-long)
Copied to clipboard
Michael Sejr Schlichtkrull, Vladimir Karpukhin, Barlas Oguz, Mike Lewis, Wen-tau Yih, Sebastian Riedel
| Challenge: | Existing research into structured data has focused on textual data and the closed-domain setting is not reflective of real-world fact checking tasks. |
| Approach: | They propose a joint reranking-and-verification model which fuses evidence documents in the verification component and a heuristic retrieval baseline. |
| Outcome: | The proposed model achieves comparable performance to the closed-domain state-of-the-art on the TabFact dataset and significantly improves over a heuristic retrieval baseline. |
Questions Are All You Need to Train a Dense Passage Retriever (2023.tacl-1)
Copied to clipboard
| Challenge: | Existing methods for dense retrieval require large supervised datasets with custom hard-negative mining and denoising of positive examples. |
| Approach: | They propose a new corpus-level autoencoding approach for training dense retrieval models that does not require labeled training data. |
| Outcome: | The proposed method matches or surpasses strong supervised performance levels on multiple QA benchmarks with no labeled training data or task-specific losses. |
Interpretation of Natural Language Rules in Conversational Machine Reading (D18-1)
Copied to clipboard
Marzieh Saeidi, Max Bartolo, Patrick Lewis, Sameer Singh, Tim Rocktäschel, Mike Sheldon, Guillaume Bouchard, Sebastian Riedel
| Challenge: | Existing work on question answering problems requires the reading of text because it contains a recipe to derive an answer together with the reader’s background knowledge. |
| Approach: | They formalise a task and develop a crowd-sourcing strategy to collect 37k task instances based on real-world rules and crowd-generated questions and scenarios. |
| Outcome: | The proposed task is based on 37k task instances based in real-world rules and crowd-generated questions and scenarios. |
Nonparametric Masked Language Modeling (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing language models (LMs) predict tokens with a softmax over a finite vocabulary, which can make it difficult to predict rare tokens or phrases. |
| Approach: | They introduce a nonparametric masked language model that replaces a softmax with a distribution over every phrase in a reference corpus and uses an in-batch approximation to train it. |
| Outcome: | The proposed model outperforms larger parametric models on 16 tasks including classification, fact probing and question answering. |
Residual Prompt Tuning: improving prompt tuning with residual reparameterization (2023.findings-acl)
Copied to clipboard
| Challenge: | Prompt tuning is one of the most parameter-efficient approaches for parameter-effective tuning of pre-trained language models. |
| Approach: | They propose to reparameterize soft prompt embeddings using a shallow network with a residual connection and use it to tune prompt embeds P. |
| Outcome: | The proposed method outperforms prompt tuning on SuperGLUE, T5-Base and BERT-Bass models and can reduce the prompt length by 10 times without hurting performance. |
Multilingual Denoising Pre-training for Neural Machine Translation (2020.tacl-1)
Copied to clipboard
Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer
| Challenge: | Existing approaches to pre-train models focus on only English corpora, but this is not common in machine translation. |
| Approach: | They propose a sequence-to-sequence denoising auto-encoder pre-trained on monolingual corpora . they show that it produces significant performance gains across MT tasks . |
| Outcome: | The proposed model can achieve significant performance gains across a wide variety of MT tasks. |
MetaICL: Learning to Learn In Context (2022.naacl-main)
Copied to clipboard
| Challenge: | Large language models can do in-context learning by conditioning on a few training examples with no parameter updates or task-specific templates. |
| Approach: | They propose a meta-training framework where a pretrained language model is tuned to do in-context learning on a large set of training tasks. |
| Outcome: | The proposed framework outperforms baseline models on 142 NLP datasets and a range of target tasks with domain shifts. |
DEMix Layers: Disentangling Domains for Modular Language Modeling (2022.naacl-main)
Copied to clipboard
| Challenge: | Extensive experiments with autoregressive transformer LMs show that DEMix layers reduce test-time perplexity and increase training efficiency. |
| Approach: | They introduce a new domain expert mixture layer that enables conditioning a language model on the domain of the input text. |
| Outcome: | Experiments with 1.3B LMs show that DEMix layers reduce test-time perplexity, increase training efficiency, and enable rapid adaptation. |
Trusting Your Evidence: Hallucinate Less with Context-aware Decoding (2024.naacl-short)
Copied to clipboard
| Challenge: | Language models (LMs) often struggle to pay enough attention to the input context, and generate texts that are unfaithful or contain hallucinations. |
| Approach: | They propose a context-aware decoding technique that amplifies the difference between the output probabilities when a model is used with and without context. |
| Outcome: | The proposed model significantly improves faithfulness of different LM families including OPT, GPT, LLaMA, and FLAN-T5 for summarization tasks. |
Span-based Hierarchical Semantic Parsing for Task-Oriented Dialog (D19-1)
Copied to clipboard
| Challenge: | Existing semantic parsers score intents and slots as labels of nesting nodes, but decode a valid tree globally. |
| Approach: | They propose a span-based semantic parser for parsing compositional utterances into Task Oriented Parse (TOP) the parsers score labels of the tree nodes covering each token span independently, but decode a valid tree globally. |
| Outcome: | The proposed parser outperforms previous methods on the TOP dataset in accuracy and training speed. |
Sparse Distillation: Speeding Up Text Classification by Using Bigger Student Models (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing methods to reduce inference cost by distilling transformer models into lightweight student models are limited for high-volume use cases. |
| Approach: | They propose to distill state-of-the-art transformer models into lightweight student models to reduce computation cost at inference time. |
| Outcome: | The proposed pipeline achieves up to 600x speed-up on GPUs and CPUs on six single-sentence text classification tasks and in domain generalization settings. |
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. |
Improving Passage Retrieval with Zero-Shot Question Generation (2022.emnlp-main)
Copied to clipboard
Devendra Sachan, Mike Lewis, Mandar Joshi, Armen Aghajanyan, Wen-tau Yih, Joelle Pineau, Luke Zettlemoyer
| Challenge: | Existing re-ranking methods for open-domain question answering are not domain- or task-specific. |
| Approach: | They propose a simple and effective re-ranking method for improving passage retrieval in open-domain question answering. |
| Outcome: | The proposed method outperforms strong supervised models on open-domain questions and triviaQA datasets on top-1000 passages. |
Strategies for Structuring Story Generation (P19-1)
Copied to clipboard
| Challenge: | Existing language models generate word by word, but fail to capture high-level interactions . a novel decomposition approach allows more abstract representations to be generated first . |
| Approach: | They propose models which abstract over actions and entities to create stories . they generate predicate-argument structure, then replace placeholders with context-sensitive names . |
| Outcome: | The proposed models improve diversity and coherence of events and entities in generated stories. |
REPLUG: Retrieval-Augmented Black-Box Language Models (2024.naacl-long)
Copied to clipboard
Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, Wen-tau Yih
| Challenge: | Existing retrieval-augmented language models require access to internal representations to enhance performance. |
| Approach: | They introduce a retrieval-augmented language modeling framework that treats the language model as a black box and augments it with a tuneable retrieval model. |
| Outcome: | The proposed framework improves performance on language modeling tasks by 6.3% and 5.1%. |
FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation (2023.emnlp-main)
Copied to clipboard
Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, Hannaneh Hajishirzi
| Challenge: | Evaluating the factuality of long-form text generated by large language models (LMs) is non-trivial because (1) generations often contain a mixture of supported and unsupported pieces of information, making binary judgments of quality inadequate and (2) human evaluation is time-consuming and costly. |
| Approach: | They introduce a new evaluation that breaks a generation into a series of atomic facts and computes the percentage of atom facts supported by a reliable knowledge source. |
| Outcome: | The proposed model breaks a generation into atomic facts and computes the percentage of atomic fact supported by a reliable knowledge source. |
BTS: Harmonizing Specialized Experts into a Generalist LLM (2025.emnlp-main)
Copied to clipboard
Qizhen Zhang, Prajjwal Bhargava, Chloe Bi, Chris X. Cai, Jakob Nicolaus Foerster, Jeremy Fu, Punit Singh Koura, Ruan Silva, Sheng Shen, Emily Dinan, Suchin Gururangan, Mike Lewis
| Challenge: | Branch-Train-Stitch (BTS) is an efficient and flexible training algorithm for combining independently trained large language model (LLM) experts into a single, capable generalist model. |
| Approach: | They propose an efficient and flexible training algorithm for combining large language model (LLM) experts into a single, capable generalist model using lightweight stitch layers. |
| Outcome: | The proposed model can generalize to new domains despite being frozen . it yields the best generalist performance on a variety of downstream tasks, retaining the specialized capabilities of each of the experts. |
AutoReply: Detecting Nonsense in Dialogue with Discriminative Replies (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models for dialogue detection make many errors in their own messages . a dataset of long dialogues richly grounded in the game state contains many errors . |
| Approach: | They propose to use an annotated dialogue dataset to generate automatic responses for dialogue models. |
| Outcome: | The proposed model outperforms handcrafted replies and performs on par with supervised learning approaches. |
Neural Compositional Denotational Semantics for Question Answering (D18-1)
Copied to clipboard
| Challenge: | a new model for compositional questions is needed to answer multi-step reasoning . the model is inspired by formal approaches to compositional semantics . |
| Approach: | They propose an end-to-end differentiable model for interpreting compositional questions . they build a latent tree of interpretable expressions over a sentence . |
| Outcome: | The proposed model outperforms RNN encoders when test questions are longer than training questions. |
Semantic Parsing for Task Oriented Dialog using Hierarchical Representations (D18-1)
Copied to clipboard
| Challenge: | Existing work on task oriented dialog systems has limited expressive power to one intent per query and one slot label per token. |
| Approach: | They propose a hierarchical annotation scheme for semantic parsing that allows representation of compositional queries. |
| Outcome: | The proposed representation outperforms sequence-to-sequence approaches on a 44k annotated query dataset. |
Byte Latent Transformer: Patches Scale Better Than Tokens (2025.acl-long)
Copied to clipboard
Artidoro Pagnoni, Ramakanth Pasunuru, Pedro Rodriguez, John Nguyen, Benjamin Muller, Margaret Li, Chunting Zhou, Lili Yu, Jason E Weston, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Ari Holtzman, Srini Iyer
| Challenge: | Existing large language models (LLMs) are trained on bytes, except for tokenization, which groups bytes into a static set of tokens. |
| Approach: | They propose a new byte-level LLM architecture that encodes bytes into dynamically sized patches, which serve as the primary units of computation. |
| Outcome: | The proposed architecture matches tokenization-based models with improvements in inference efficiency and robustness. |
Effective Long-Context Scaling of Foundation Models (2024.naacl-long)
Copied to clipboard
Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, Madian Khabsa, Han Fang, Yashar Mehdad, Sharan Narang, Kshitiz Malik, Angela Fan, Shruti Bhosale, Sergey Edunov, Mike Lewis, Sinong Wang, Hao Ma
| Challenge: | Large language models (LLMs) are rapidly deployed and continue to evolve through scaling. |
| Approach: | They propose a method to train strong long-context LLMs that are capable of utilizing massive context windows of up to 32,000 tokens. |
| Outcome: | The proposed model can surpass gpt-3.5-turbo-16k's overall performance on long-context benchmarks with a cost-effective instruction tuning procedure that is free of expensive annotations. |
BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension (2020.acl-main)
Copied to clipboard
Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, Luke Zettlemoyer
| Challenge: | Recent work has shown gains by improving the distribution of masked tokens and the order in which mucked tokens are predicted. |
| Approach: | They propose a denoising autoencoder for pretraining sequence-to-sequence models that corrupts text with an arbitrary noising function and learns a model to reconstruct the original text. |
| Outcome: | The proposed model outperforms RoBERTa on GLUE and SQUAD and provides a 1.1 BLEU increase over a back-translation system for machine translation. |
In-context Examples Selection for Machine Translation (2023.findings-acl)
Copied to clipboard
| Challenge: | Large-scale generative models can perform a wide range of NLP tasks using in-context learning. |
| Approach: | They aim to understand the properties of good in-context examples for machine translation in both in-domain and out-of-domain settings. |
| Outcome: | The proposed model outperforms a strong kNN-MT baseline in 2 out of 4 out-of-domain datasets. |
Asking and Answering Questions to Evaluate the Factual Consistency of Summaries (2020.acl-main)
Copied to clipboard
| Challenge: | Existing automatic evaluation metrics for summarization are insensitive to factual inconsistencies. |
| Approach: | They propose an automatic evaluation protocol that detects factual inconsistencies in a model-generated summary. |
| Outcome: | QAGS has higher correlations with human judgments of factual consistency than other evaluation metrics. |
Cross-lingual Transfer Learning for Multilingual Task Oriented Dialog (N19-1)
Copied to clipboard
| Challenge: | a lack of multilingual training data has hindered development of conversational AI models for task-oriented tasks . a new data set of 57k annotated utterances in english, spanish, and Thai is used to evaluate cross-lingual methods . |
| Approach: | They present a data set of 57k annotated utterances in English, Spanish and Thai . they evaluate three different cross-lingual transfer methods to identify user intents and slots . |
| Outcome: | The proposed model outperforms existing methods in English, Spanish and Thai . the proposed model is based on training data from three languages . |
Noisy Channel Language Model Prompting for Few-Shot Text Classification (2022.acl-long)
Copied to clipboard
| Challenge: | Prior work has suggested methods for finding better prompt or scoring of the output from the model. |
| Approach: | They propose a noisy channel approach for language model prompting in few-shot text classification by in-context demonstration or prompt tuning. |
| Outcome: | The proposed model outperforms direct models in both demonstration and prompt tuning. |
Multitasking Inhibits Semantic Drift (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing studies have found that LLP training is prone to semantic drift (use of messages inconsistent with their natural language meanings) |
| Approach: | They propose to use latent language policies to train neural LLPs to eliminate semantic drift in a well-studied family of signaling games to reduce drift and improve sample efficiency. |
| Outcome: | The proposed model eliminates semantic drift in a well-studied family of signaling games while improving sample efficiency. |
Grounded Adaptation for Zero-shot Executable Semantic Parsing (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing semantic parsers are usually engineered for each application environment, but they struggle when deployed to a new database. |
| Approach: | They propose a method to adapt existing semantic parsers to new environments . they propose combining a forward semantic parsed with a backward utterance generator to synthesize data in the new environment and select cycle-consistent examples to adapt the parser. |
| Outcome: | The proposed procedure outperforms data-augmentation and improves execution accuracy on the Spider, Sparc, and CoSQL zero-shot semantic parsing tasks. |
Tricks for Training Sparse Translation Models (2022.naacl-main)
Copied to clipboard
| Challenge: | Multitask learning with an unbalanced data distribution skews model learning towards high resource tasks. |
| Approach: | They propose to use a temperature heating mechanism and dense pre-training to mitigate this by training models with a fixed model capacity. |
| Outcome: | The proposed techniques improve performance on two multilingual translation benchmarks compared to BASELayers and Dense scaling baselines and in combination, more than 2x model convergence speed. |
Question Answering Infused Pre-training of General-Purpose Contextualized Representations (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing pretraining objectives for question answering (QA) are not optimized for being immediately useful without fine-tuning. |
| Approach: | They propose a pre-training objective based on question answering (QA) that is based more directly on context. |
| Outcome: | The proposed model matches predictions of a more accurate cross-encoder model on 80 million synthesized QA pairs and achieves large improvements over previous state-of-the-art models on paraphrase detection and fewshot named entity recognition. |
Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? (2022.emnlp-main)
Copied to clipboard
Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, Luke Zettlemoyer
| Challenge: | Large language models can in-context learn by conditioning on a few input-label pairs and making predictions for new inputs. |
| Approach: | They propose to use ground truth demonstrations to replace labels in demonstrations . they also show that other aspects of the demonstrations are key drivers of endtask performance . |
| Outcome: | The proposed model outperforms zeroshot inference on a wide range of tasks using ground truth demonstrations. |