We propose an efficient framework to compress massive video-frame features before feeding them into large multimodal models, thereby mitigating the severe token explosion arising from hour-long videos. Our design leverages a bidirectional state-space model equipped with a gated skip connection and a learnable weighted-average pooling mechanism applied to periodically inserted learned queries. This structure enables hierarchical downsampling across both spatial and temporal dimensions, preserving performance in a cost-effective manner. Across challenging hour-long video understanding tasks, our approach demonstrates competitive results against state-of-the-art models, while significantly reducing overall token budget. Notably, replacing our state-space model with conventional modules results in substantial performance degradation, highlighting the advantages of the proposed state-space modeling for effectively compressing multi-frame video information. Our framework emphasizes resource-conscious efficiency, making it practical for real-world deployments. We validate its scalability and generality across multiple benchmarks, achieving the dual objectives of efficient resource usage and comprehensive video understanding.

Jiyeon Kim, Hyunji Lee, Dylan Zhou, Sue Hyun Park, Seunghyun Yoon, Trung Bui, Franck Dernoncourt, Sungmin Cha, Minjoon Seo. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Information retrieval methods often rely on a single embedding model trained on large, general-domain datasets like MSMARCO. While this approach can produce a retriever with reasonable overall performance, they often underperform models trained on domain-specific data when testing on their respective domains. Prior work in information retrieval has tackled this through multi-task training, but the idea of routing over a mixture of domain-specific expert retrievers remains unexplored despite the popularity of such ideas in language model generation research. In this work, we introduce RouterRetriever, a retrieval model that leverages a mixture of domain-specific experts by using a routing mechanism to select the most appropriate expert for each query. RouterRetriever is lightweight and allows easy addition or removal of experts without additional training. Evaluation on the BEIR benchmark demonstrates that RouterRetriever outperforms both models trained on MSMARCO (+2.1 absolute nDCG@10) and multi-task models (+3.2). This is achieved by employing our routing mechanism, which surpasses other routing techniques (+1.8 on average) commonly used in language modeling. Furthermore, the benefit generalizes well to other datasets, even in the absence of a specific expert on the dataset. RouterRetriever is the first work to demonstrate the advantages of routing over a mixture of domain-specific expert embedding models as an alternative to a single, general-purpose embedding model, especially when retrieving from diverse, specialized domains.

In this work, we investigate how a model's tendency to broadly integrate its parametric knowledge evolves throughout pretraining, and how this behavior affects overall performance, particularly in terms of knowledge acquisition and forgetting. We introduce the concept of knowledge entropy, which quantifies the range of memory sources the model engages with; high knowledge entropy indicates that the model utilizes a wide range of memory sources, while low knowledge entropy suggests reliance on specific sources with greater certainty. Our analysis reveals a consistent decline in knowledge entropy as pretraining advances. We also find that the decline is closely associated with a reduction in the model's ability to acquire and retain knowledge, leading us to conclude that diminishing knowledge entropy (smaller number of active memory sources) impairs the model's knowledge acquisition and retention capabilities. We find further support for this by demonstrating that increasing the activity of inactive memory sources enhances the model's capacity for knowledge acquisition and retention.

Vision-Language adaptation (VL adaptation) transforms Large Language Models (LLMs) into Large Vision-Language Models (LVLMs) for multimodal tasks, but this process often compromises the inherent safety capabilities embedded in the original LLMs. Despite potential harmfulness due to weakened safety measures, in-depth analysis on the effects of VL adaptation on safety remains under-explored. This study examines how VL adaptation influences safety and evaluates the impact of safety fine-tuning methods. Our analysis reveals that safety degradation occurs during VL adaptation, even when the training data is safe. While safety tuning techniques like supervised fine-tuning with safety datasets or reinforcement learning from human feedback mitigate some risks, they still lead to safety degradation and a reduction in helpfulness due to over-rejection issues. Further analysis of internal model weights suggests that VL adaptation may impact certain safety-related layers, potentially lowering overall safety levels. Additionally, our findings demonstrate that the objectives of VL adaptation and safety tuning are divergent, which often results in their simultaneous application being suboptimal. To address this, we suggest the weight merging approach as an optimal solution effectively reducing safety degradation while maintaining helpfulness. These insights help guide the development of more reliable and secure LVLMs for real-world applications.

We introduce Latent Action Pretraining for general Action models (LAPA), the first unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation models.

Seungone Kim, Juyoung Suk, Ji Yong Cho, Shayne Longpre, Chaeeun Kim, Dongkeun Yoon, Guijin Son, Yejin Cho, Sheikh Shafayat, Jinheon Baek, Sue Hyun Park, Hyeonbin Hwang, Jinkyung Jo, Hyowon Cho, Haebin Shin, Seongyun Lee, Hanseok Oh, Noah Lee, Namgyu Ho, Se June Joo, Miyoung Ko, Yoonjoo Lee, Hyungjoo Chae, Jamin Shin, Joel Jang, Seonghyeon Ye, Bill Yuchen Lin, Sean Welleck, Graham Neubig, Moontae Lee, Kyungjae Lee, Minjoon Seo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Haebin Shin, Lei Ji, Yeyun Gong, Sungdong Kim, Eunbi Choi, Minjoon Seo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Despite their strengths, large language models (LLMs) often fail to communicate their confidence accurately, making it difficult to assess when they might be wrong and limiting their reliability. In this work, we demonstrate that reasoning models that engage in extended chain-of-thought (CoT) reasoning exhibit superior performance not only in problem-solving but also in accurately expressing their confidence. Specifically, we benchmark six reasoning models across six datasets and find that they achieve strictly better confidence calibration than their non-reasoning counterparts in 33 out of the 36 settings. Our detailed analysis reveals that these gains in calibration stem from the slow thinking behaviors of reasoning models (e.g., exploring alternative approaches and backtracking) which enable them to adjust their confidence dynamically throughout their CoT, making it progressively more accurate. In particular, we find that reasoning models become increasingly better calibrated as their CoT unfolds, a trend not observed in non-reasoning models. Moreover, removing slow thinking behaviors from the CoT leads to a significant drop in calibration. Lastly, we show that non-reasoning models also demonstrate enhanced calibration when simply guided to slow think via in-context learning, fully isolating slow thinking as the source of the calibration gains.

In this paper, we present our finding that prepending a Task-Agnostic Prefix Prompt (TAPP) to the input improves the instruction-following ability of various Large Language Models (LLMs) during inference. TAPP is different from canonical prompts for LLMs in that it is a fixed prompt prepended to the beginning of every input regardless of the target task for zero-shot generalization. We observe that both base LLMs (i.e. not fine-tuned to follow instructions) and instruction-tuned models benefit from TAPP, resulting in 34.58% and 12.26% improvement on average, respectively. This implies that the instruction-following ability of LLMs can be improved during inference time with a fixed prompt constructed with simple heuristics. We hypothesize that TAPP assists language models to better estimate the output distribution by focusing more on the instruction of the target task during inference. In other words, such ability does not seem to be sufficiently activated in not only base LLMs but also many instruction-fine-tuned LLMs.

Hyunji Lee, Doyoung Kim, Jihoon Jun, Se June Joo, Joel Jang, Kyoung-Woon On, Minjoon Seo. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Sangkyu Lee, Sungdong Kim, Ashkan Yousefpour, Minjoon Seo, Kang Min Yoo, Youngjae Yu. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Dongkeun Yoon, Joel Jang, Sungdong Kim, Seungone Kim, Sheikh Shafayat, Minjoon Seo. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Recent advancements in language and vision assistants have showcased impressive capabilities but suffer from a lack of transparency, limiting broader research and reproducibility.While open-source models handle general image tasks effectively, they face challenges with the high computational demands of complex visuallysituated text understanding.Such tasks often require increased token inputs and large vision modules to harness high-resolution information.Striking a balance between model size and data importance remains an open question.This study aims to redefine the design of visionlanguage models by identifying key components and creating efficient models with constrained inference costs.By strategically formulating datasets, optimizing vision modules, and enhancing supervision techniques, we achieve significant improvements in inference throughput while maintaining high performance.Extensive experiments across models ranging from 160M to 13B parameters offer insights into model optimization.We will fully opensource our codebase, models, and datasets at https://github.com/naver-ai/elva.

Learning from human feedback via proxy reward modeling has been studied to align Large Language Models (LLMs) with human values.However, achieving reliable training through that proxy reward model (RM) is not a trivial problem, and its behavior remained as a blackbox.In this paper, we study the role of proxy rewards in the LLM alignment via 'reverse reward engineering' by composing interpretable features as a white-box reward function.We aim to replicate the ground truth (gold) reward signal by achieving a monotonic relationship between the proxy and gold reward signals after training the model using the proxy reward in reinforcement learning (RL).Our findings indicate that successfully emulating the gold reward requires generating responses that are relevant with enough length to open-ended questions, while also ensuring response consistency in closed-ended questions.Furthermore, resulting models optimizing our devised white-box reward show competitive performances with strong open-source RMs in alignment benchmarks.We highlight its potential usage as a simple but strong reward baseline for the LLM alignment, not requiring explicit pairwise human feedback dataset and RM training 1 .

Seungone Kim, Juyoung Suk, Shayne Longpre, Bill Yuchen Lin, Jamin Shin, Sean Welleck, Graham Neubig, Moontae Lee, Kyungjae Lee, Minjoon Seo. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Benchmarking the performance of information retrieval (IR) is mostly conducted with a fixed set of documents (static corpora).However, in realistic scenarios, this is rarely the case and the documents to be retrieved are constantly updated and added.In this paper, we focus on Generative Retrievals (GR), which apply autoregressive language models to IR problems, and explore their adaptability and robustness in dynamic scenarios.We also conduct an extensive evaluation of computational and memory efficiency, crucial factors for real-world deployment of IR systems handling vast and ever-changing document collections.Our results on the StreamingQA benchmark demonstrate that GR is more adaptable to evolving knowledge (4 -11%), robust in learning knowledge with temporal information, and efficient in terms of inference FLOPs (ˆ2), indexing time (ˆ6), and storage footprint (ˆ4) compared to Dual Encoders (DE), which are commonly used in retrieval systems.Our paper highlights the potential of GR for future use in practical IR systems within dynamic environments.

Despite the advances in large language models (LLMs), how they use their knowledge for reasoning is not yet well understood.In this study, we propose a method that deconstructs complex real-world questions into a graph, representing each question as a node with predecessors of background knowledge needed to solve the question.We develop the DEPTHQA dataset, deconstructing questions into three depths: (i) recalling conceptual knowledge, (ii) applying procedural knowledge, and (iii) analyzing strategic knowledge.Based on a hierarchical graph, we quantify forward discrepancy, discrepancies in LLMs' performance on simpler sub-problems versus complex questions.We also measure backward discrepancy, where LLMs answer complex questions but struggle with simpler ones.Our analysis shows that smaller models exhibit more discrepancies than larger models.Distinct patterns of discrepancies are observed across model capacity and possibility of training data memorization.Additionally, guiding models from simpler to complex questions through multiturn interactions improves performance across model sizes, highlighting the importance of structured intermediate steps in knowledge reasoning.This work enhances our understanding of LLM reasoning and suggests ways to improve their problem-solving abilities.

Large language models (LLMs) have made significant advancements in various natural language processing tasks, including question answering (QA) tasks. While incorporating new information with the retrieval of relevant passages is a promising way to improve QA with LLMs, the existing methods often require additional fine-tuning which becomes infeasible with recent LLMs. Augmenting retrieved passages via prompting has the potential to address this limitation, but this direction has been limitedly explored. To this end, we design a simple yet effective framework to enhance open-domain QA (ODQA) with LLMs, based on the summarized retrieval (SuRe). SuRe helps LLMs predict more accurate answers for a given question, which are well-supported by the summarized retrieval that could be viewed as an explicit rationale extracted from the retrieved passages. Specifically, SuRe first constructs summaries of the retrieved passages for each of the multiple answer candidates. Then, SuRe confirms the most plausible answer from the candidate set by evaluating the validity and ranking of the generated summaries. Experimental results on diverse ODQA benchmarks demonstrate the superiority of SuRe, with improvements of up to 4.6\% in exact match (EM) and 4.0\% in F1 score over standard prompting approaches. SuRe also can be integrated with a broad range of retrieval methods and LLMs. Finally, the generated summaries from SuRe show additional advantages to measure the importance of retrieved passages and serve as more preferred rationales by models and humans.

Recently, GPT-4 has become the de facto evaluator for long-form text generated by large language models (LLMs). However, for practitioners and researchers with large and custom evaluation tasks, GPT-4 is unreliable due to its closed-source nature, uncontrolled versioning, and prohibitive costs. In this work, we propose PROMETHEUS a fully open-source LLM that is on par with GPT-4’s evaluation capabilities when the appropriate reference materials (reference answer, score rubric) are accompanied. For this purpose, we construct a new dataset – FEEDBACK COLLECTION – that consists of 1K fine-grained score rubrics, 20K instructions, and 100K natural language feedback generated by GPT-4. Using the FEEDBACK COLLECTION, we train PROMETHEUS, a 13B evaluation-specific LLM that can assess any given response based on novel and unseen score rubrics and reference materials provided by the user. Our dataset’s versatility and diversity make our model generalize to challenging real-world criteria, such as prioritizing conciseness, child-readability, or varying levels of formality. We show that PROMETHEUS shows a stronger correlation with GPT-4 evaluation compared to ChatGPT on seven evaluation benchmarks (Two Feedback Collection testsets, MT Bench, Vicuna Bench, Flask Eval, MT Bench Human Judgment, and HHH Alignment), showing the efficacy of our model and dataset design. During human evaluation with hand-crafted score rubrics, PROMETHEUS shows a Pearson correlation of 0.897 with human evaluators, which is on par with GPT-4-0613 (0.882), and greatly outperforms ChatGPT (0.392). Remarkably, when assessing the quality of the generated feedback, PROMETHEUS demonstrates a win rate of 58.62% when compared to GPT-4 evaluation and a win rate of 79.57% when compared to ChatGPT evaluation. Our findings suggests that by adding reference materials and training on GPT-4 feedback, we can obtain effective open-source evaluator LMs.

Evaluation of Large Language Models (LLMs) is challenging because instruction-following necessitates alignment with human values and the required set of skills varies depending on the instruction. However, previous studies have mainly focused on coarse-grained evaluation (i.e. overall preference-based evaluation), which limits interpretability since it does not consider the nature of user instructions that require instance-wise skill composition. In this paper, we introduce FLASK (Fine-grained Language Model Evaluation based on Alignment Skill Sets), a fine-grained evaluation protocol for both human-based and model-based evaluation which decomposes coarse-level scoring to a skill set-level scoring for each instruction. We experimentally observe that the fine-graininess of evaluation is crucial for attaining a holistic view of model performance and increasing the reliability of the evaluation. Using FLASK, we compare multiple open-source and proprietary LLMs and observe a high correlation between model-based and human-based evaluations.

Hyunji Lee, Se June Joo, Chaeeun Kim, Joel Jang, Doyoung Kim, Kyoung-Woon On, Minjoon Seo. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Seongyun Lee, Sue Hyun Park, Yongrae Jo, Minjoon Seo. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Hanseok Oh, Haebin Shin, Miyoung Ko, Hyunji Lee, Minjoon Seo. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, Wen-tau Yih. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Despite the recent observation that large language models (LLMs) can store substantial factual knowledge, there is a limited understanding of the mechanisms of how they acquire factual knowledge through pretraining. This work addresses this gap by studying how LLMs acquire factual knowledge during pretraining. The findings reveal several important insights into the dynamics of factual knowledge acquisition during pretraining. First, counterintuitively, we observe that pretraining on more data shows no significant improvement in the model's capability to acquire and maintain factual knowledge. Next, LLMs undergo forgetting of memorization and generalization of factual knowledge, and LLMs trained with duplicated training data exhibit faster forgetting. Third, training LLMs with larger batch sizes can enhance the models' robustness to forgetting. Overall, our observations suggest that factual knowledge acquisition in LLM pretraining occurs by progressively increasing the probability of factual knowledge presented in the pretraining data at each step. However, this increase is diluted by subsequent forgetting. Based on this interpretation, we demonstrate that we can provide plausible explanations on recently observed behaviors of LLMs, such as the poor performance of LLMs on long-tail knowledge and the benefits of deduplicating the pretraining corpus.

Although humans inherently have diverse values, current large language model (LLM) alignment methods often assume that aligning LLMs with the general public’s preferences is optimal. A major challenge in adopting a more individualized approach to LLM alignment is its lack of scalability, as it involves repeatedly acquiring preference data and training new reward models and LLMs for each individual’s preferences. To address these challenges, we propose a new paradigm where users specify what they value most within the system message, steering the LLM’s generation behavior to better align with the user’s intentions. However, a naive application of such an approach is non-trivial since LLMs are typically trained on a uniform system message (e.g., “You are a helpful assistant”), which limits their ability to generalize to diverse, unseen system messages. To improve this generalization, we create Multifaceted Collection, augmenting 66k user instructions into 197k system messages through hierarchical user value combinations. Using this dataset, we train a 7B LLM called Janus and test it on 921 prompts from 5 benchmarks (AlpacaEval 2.0, FLASK, Koala, MT-Bench, and Self-Instruct) by adding system messages that reflect unseen user values. JANUS achieves tie+win rate of 75.2%, 72.4%, and 66.4% against Mistral 7B Instruct v0.2, GPT-3.5 Turbo, and GPT-4, respectively. Unexpectedly, on three benchmarks focused on response helpfulness (AlpacaEval 2.0, MT-Bench, Arena Hard Auto v0.1), JANUS also outperforms LLaMA 3 8B Instruct by a +4.0%p, +0.1%p, +3.0%p margin, underscoring that training with a vast array of system messages could also enhance alignment to the general public’s preference as well. Our code, dataset, benchmark, and models are available at https://lklab.kaist.ac.kr/Janus/.

Abstract Previous work in prompt engineering for large language models has introduced different gradient-free probability-based prompt selection methods that aim to choose the optimal prompt among the candidates for a given task but have failed to provide a comprehensive and fair comparison between each other. In this paper, we propose a unified framework to interpret and evaluate the existing probability-based prompt selection methods by performing extensive experiments on 13 common and diverse NLP tasks. We find that each of the existing methods can be interpreted as some variant of the method that maximizes mutual information between the input and the predicted output (MI). Utilizing this finding, we develop several other combinatorial variants of MI and increase the effectiveness of the oracle prompt selection method from 87.79% to 94.98%, measured as the ratio of the performance of the selected prompt to that of the optimal oracle prompt. Furthermore, considering that all the methods rely on the output probability distribution of the model that might be biased, we propose a novel calibration method called Calibration by Marginalization (CBM) that is orthogonal to the existing methods and helps increase the prompt selection effectiveness of the best method to 96.85%, achieving 99.44% of the oracle prompt F1 without calibration.1

Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, Minjoon Seo. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Soyoung Yoon, Sungjoon Park, Gyuwan Kim, Junhee Cho, Kihyo Park, Gyu Tae Kim, Minjoon Seo, Alice Oh. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Aligning large language models (LLMs) to human values has become increasingly important as it enables sophisticated steering of LLMs.However, it requires significant human demonstrations and feedback or distillation from proprietary LLMs such as ChatGPT.In this work, we propose a novel alignment learning framework with synthetic feedback not dependent on extensive human annotations and proprietary LLMs.First, we perform reward modeling (RM) with synthetic feedback by contrasting responses from vanilla LLMs with various sizes and prompts.Then, we use the RM to simulate high-quality demonstrations to train a supervised policy and further optimize the model with reinforcement learning.Our resulting model, Aligned Language Model with Synthetic Training dataset (ALMoST), outperforms recent open-sourced models, which are trained on the outputs of InstructGPT or humanannotated demonstrations, in alignment benchmarks.In human evaluation, our model is preferred to Alpaca and Dolly-v2, 55.0% and 58.5% of the time, respectively.Further analyses demonstrate the efficacy and importance of synthetic feedback in our framework 1 .

Language models (LMs) with less than 100B parameters are known to perform poorly on chain-of-thought (CoT) reasoning in contrast to large LMs when solving unseen tasks.In this work, we aim to equip smaller LMs with the step-by-step reasoning capability by instruction tuning with CoT rationales.In order to achieve this goal, we first introduce a new instructiontuning dataset called the COT COLLECTION, which augments the existing Flan Collection (including only 9 CoT tasks) with additional 1.84 million rationales across 1,060 tasks.We show that CoT fine-tuning Flan-T5 (3B & 11B) with COT COLLECTION enables smaller LMs to have better CoT capabilities on unseen tasks.On the BIG-Bench-Hard (BBH) benchmark, we report an average improvement of +4.34% (Flan-T5 3B) and +2.60% (Flan-T5 11B), in terms of zero-shot task accuracy.Furthermore, we show that instruction tuning with COT COL-LECTION allows LMs to possess stronger fewshot learning capabilities on 4 domain-specific tasks, resulting in an improvement of +2.24% (Flan-T5 3B) and +2.37% (Flan-T5 11B), even outperforming ChatGPT utilizing demonstrations until the max length by a +13.98% margin.Our code, the COT COLLECTION data, and model checkpoints are publicly available 1 .

Recently, Language Models (LMs) instruction-tuned on multiple tasks, also known as multitask-prompted fine-tuning (MT), have shown capabilities to generalize to unseen tasks. Previous work has shown that scaling the number of finetuning datasets and instructions is the key component in making stronger MT LMs. In this work, we report surprising findings that show an expert LM trained on just a single task can outperform an MT LM trained with 300+ different tasks on 11 different unseen datasets and on 13 datasets of the BIG-bench benchmark by an average of 3.20% and 1.29%, respectively. This finding casts doubt on the previously held belief that simply scaling the number of tasks makes stronger MT LMs. Leveraging this finding, we further show that this distributed approach of training multiple expert LMs instead of a single MT LM for zero-shot inference possesses many benefits including (1) avoiding negative task transfer that often occurs during instruction tuning, (2) being able to continually learn new tasks without having to re-train on previous tasks to avoid catastrophic forgetting, and (3) showing compositional capabilities when merging individual experts together.

Training data attribution (TDA) techniques find influential training data for the model's prediction on the test data of interest. They approximate the impact of down- or up-weighting a particular training sample. While conceptually useful, they are hardly applicable to deep models in practice, particularly because of their sensitivity to different model initialisation. In this paper, we introduce a Bayesian perspective on the TDA task, where the learned model is treated as a Bayesian posterior and the TDA estimates as random variables. From this novel viewpoint, we observe that the influence of an individual training sample is often overshadowed by the noise stemming from model initialisation and SGD batch composition. Based on this observation, we argue that TDA can only be reliably used for explaining deep model predictions that are consistently influenced by certain training data, independent of other noise factors. Our experiments demonstrate the rarity of such noise-independent training-test data pairs but confirm their existence. We recommend that future researchers and practitioners trust TDA estimates only in such cases. Further, we find a disagreement between ground truth and estimated TDA distributions and encourage future work to study this gap. Code is provided at https://github.com/ElisaNguyen/bayesian-tda.

Language Models (LMs) become outdated as the world changes; they often fail to perform tasks requiring recent factual information which was absent or different during training, a phenomenon called temporal misalignment.This is especially a challenging problem because the research community still lacks a coherent dataset for assessing the adaptability of LMs to frequently-updated knowledge corpus such as Wikipedia.To this end, we introduce TEMPORALWIKI, a lifelong benchmark for ever-evolving LMs that utilizes the difference between consecutive snapshots of English Wikipedia and English Wikidata for training and evaluation, respectively.The benchmark hence allows researchers to periodically track an LM's ability to retain previous knowledge and acquire updated/new knowledge at each point in time.We also find that training an LM on the diff data through continual learning methods achieves similar or better perplexity than on the entire snapshot in our benchmark with 12 times less computational cost, which verifies that factual knowledge in LMs can be safely updated with minimal training data via continual learning.The dataset and the code is made available at this link. What is the most dominant COVID-19 variant?What is the most dominant COVID-19 variant?Difference

A common practice for text retrieval is to use an encoder to map the documents and the query to a common vector space and perform a nearest neighbor search (NNS); multi-hop retrieval also often adopts the same paradigm, usually with a modification of iteratively reformulating the query vector so that it can retrieve different documents at each hop.However, such a biencoder approach has limitations in multi-hop settings; (1) the reformulated query gets longer as the number of hops increases, which further tightens the embedding bottleneck of the query vector, and (2) it is prone to error propagation.In this paper, we focus on alleviating these limitations in multi-hop settings by formulating the problem in a fully generative way.We propose an encoder-decoder model that performs multi-hop retrieval by simply generating the entire text sequences of the retrieval targets, which means the query and the documents interact in the language model's parametric space rather than L2 or inner product space as in the bi-encoder approach.Our approach, Generative Multi-hop Retrieval (GMR), consistently achieves comparable or higher performance than bi-encoder models in five datasets while demonstrating superior GPU memory and storage footprint.1

The recent advances of deep learning have dramatically changed how machine learning, especially in the domain of natural language processing, can be applied to legal domain. However, this shift to the data-driven approaches calls for larger and more diverse datasets, which are nevertheless still small in number, especially in non-English languages. Here we present the first large-scale benchmark of Korean legal AI datasets, LBOX OPEN, that consists of one legal corpus, two classification tasks, two legal judgement prediction (LJP) tasks, and one summarization task. The legal corpus consists of 147k Korean precedents (259M tokens), of which 63k are sentenced in last 4 years and 96k are from the first and the second level courts in which factual issues are reviewed. The two classification tasks are case names (11.3k) and statutes (2.8k) prediction from the factual description of individual cases. The LJP tasks consist of (1) 10.5k criminal examples where the model is asked to predict fine amount, imprisonment with labor, and imprisonment without labor ranges for the given facts, and (2) 4.7k civil examples where the inputs are facts and claim for relief and outputs are the degrees of claim acceptance. The summarization task consists of the Supreme Court precedents and the corresponding summaries (20k). We also release realistic variants of the datasets by extending the domain (1) to infrequent case categories in case name (31k examples) and statute (17.7k) classification tasks, and (2) to long input sequences in the summarization task (51k). Finally, we release LCUBE, the first Korean legal language model trained on the legal corpus from this study. Given the uniqueness of the Law of South Korea and the diversity of the legal tasks covered in this work, we believe that LBOX OPEN contributes to the multilinguality of global legal research. LBOX OPEN and LCUBE will be publicly available.

We present a new text-to-SQL dataset for electronic health records (EHRs). The utterances were collected from 222 hospital staff, including physicians, nurses, insurance review and health records teams, and more. To construct the QA dataset on structured EHR data, we conducted a poll at a university hospital and templatized the responses to create seed questions. Then, we manually linked them to two open-source EHR databases—MIMIC-III and eICU—and included them with various time expressions and held-out unanswerable questions in the dataset, which were all collected from the poll. Our dataset poses a unique set of challenges: the model needs to 1) generate SQL queries that reflect a wide range of needs in the hospital, including simple retrieval and complex operations such as calculating survival rate, 2) understand various time expressions to answer time-sensitive questions in healthcare, and 3) distinguish whether a given question is answerable or unanswerable based on the prediction confidence. We believe our dataset, EHRSQL, could serve as a practical benchmark to develop and assess QA models on structured EHR data and take one step further towards bridging the gap between text-to-SQL research and its real-life deployment in healthcare. EHRSQL is available at https://github.com/glee4810/EHRSQL.

A real-world information extraction (IE) system for semi-structured document images often involves a long pipeline of multiple modules, whose complexity dramatically increases its development and maintenance cost.One can instead consider an endto-end model that directly maps the input to the target output and simplify the entire process.However, such generation approach is known to lead to unstable performance if not designed carefully.Here we present our recent effort on transitioning from our existing pipeline-based IE system to an end-to-end system focusing on practical challenges that are associated with replacing and deploying the system in real, large-scale production.By carefully formulating document IE as a sequence generation task, we show that a single end-to-end IE system can be built and still achieve competent performance.* Most work done while these authors were at NAVER. 1 In October 2020, the system receives approximately 350k name cards and 650k receipts queries per day.

In open-domain question answering (QA), retrieve-and-read mechanism has the inherent benefit of interpretability and the easiness of adding, removing, or editing knowledge compared to the parametric approaches of closedbook QA models.However, it is also known to suffer from its large storage footprint due to its document corpus and index.Here, we discuss several orthogonal strategies to drastically reduce the footprint of a retrieve-andread open-domain QA system by up to 160x.Our results indicate that retrieve-and-read can be a viable option even in a highly constrained serving environment such as edge devices, as we show that it can achieve better accuracy than a purely parametric model with comparable docker-level system size. 1

Open-domain question answering can be formulated as a phrase retrieval problem, in which we can expect huge scalability and speed benefit but often suffer from low accuracy due to the limitation of existing phrase representation models.In this paper, we aim to improve the quality of each phrase embedding by augmenting it with a contextualized sparse representation (SPARC).Unlike previous sparse vectors that are term-frequencybased (e.g., tf-idf) or directly learned (only few thousand dimensions), we leverage rectified self-attention to indirectly learn sparse vectors in n-gram vocabulary space.By augmenting the previous phrase retrieval model (Seo et al., 2019) with SPARC, we show 4%+ improvement in CuratedTREC and SQuAD-Open.Our CuratedTREC score is even better than the best known retrieve & read model with at least 45x faster inference speed. 1

Existing open-domain question answering (QA) models are not suitable for real-time usage because they need to process several long documents on-demand for every input query, which is computationally prohibitive. In this paper, we introduce query-agnostic indexable representations of document phrases that can drastically speed up open-domain QA. In particular, our dense-sparse phrase encoding effectively captures syntactic, semantic, and lexical information of the phrases and eliminates the pipeline filtering of context documents. Leveraging strategies for optimizing training and inference time, our model can be trained and deployed even in a single 4-GPU server. Moreover, by representing phrases as pointers to their start and end tokens, our model indexes phrases in the entire English Wikipedia (up to 60 billion phrases) using under 2TB. Our experiments on SQuAD-Open show that our model is on par with or more accurate than previous models with 6000x reduced computational cost, which translates into at least 68x faster end-to-end inference benchmark on CPUs. Code and demo are available at nlp.cs.washington.edu/denspi

Jaemin Cho, Minjoon Seo, Hannaneh Hajishirzi. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

We formalize a new modular variant of current question answering tasks by enforcing complete independence of the document encoder from the question encoder.This formulation addresses a key challenge in machine comprehension by requiring a standalone representation of the document discourse.It additionally leads to a significant scalability advantage since the encoding of the answer candidate phrases in the document can be pre-computed and indexed offline for efficient retrieval.We experiment with baseline models for the new task, which achieve a reasonable accuracy but significantly underperform unconstrained QA models.We invite the QA research community to engage in Phrase-Indexed Question Answering (PIQA, pika) for closing the gap.The leaderboard is at: nlp.

We show that relation extraction can be reduced to answering simple reading comprehension questions, by associating one or more natural-language questions with each relation slot.This reduction has several advantages: we can (1) learn relationextraction models by extending recent neural reading-comprehension techniques, (2) build very large training sets for those models by combining relation-specific crowd-sourced questions with distant supervision, and even (3) do zero-shot learning by extracting new relation types that are only specified at test-time, for which we have no labeled training examples.Experiments on a Wikipedia slot-filling task demonstrate that the approach can generalize to new questions for known relation types with high accuracy, and that zero-shot generalization to unseen relation types is possible, at lower accuracy levels, setting the bar for future work on this task.

We introduce the task of Multi-Modal Machine Comprehension (M3C), which aims at answering multimodal questions given a context of text, diagrams and images. We present the Textbook Question Answering (TQA) dataset that includes 1,076 lessons and 26,260 multi-modal questions, taken from middle school science curricula. Our analysis shows that a significant portion of questions require complex parsing of the text and the diagrams and reasoning, indicating that our dataset is more complex compared to previous machine comprehension and visual question answering datasets. We extend state-of-the-art methods for textual machine comprehension and visual question answering to the TQA dataset. Our experiments show that these models do not perform well on TQA. The presented dataset opens new challenges for research in question answering and reasoning across multiple modalities.

This paper introduces GEOS, the first automated system to solve unaltered SAT geometry questions by combining text understanding and diagram interpretation.We model the problem of understanding geometry questions as submodular optimization, and identify a formal problem description likely to be compatible with both the question text and diagram.GEOS then feeds the description to a geometric solver that attempts to determine the correct answer.In our experiments, GEOS achieves a 49% score on official SAT questions, and a score of 61% on practice questions. 1 Finally, we show that by integrating textual and visual information, GEOS boosts the accuracy of dependency and semantic parsing of the question text.

Automatically solving geometry questions is a long-standing AI problem. A geometry question typically includes a textual description accompanied by a diagram. The first step in solving geometry questions is diagram understanding, which consists of identifying visual elements in the diagram, their locations, their geometric properties, and aligning them to corresponding textual descriptions. In this paper, we present a method for diagram understanding that identifies visual elements in a diagram while maximizing agreement between textual and visual data. We show that the method's objective function is submodular; thus we are able to introduce an efficient method for diagram understanding that is close to optimal. To empirically evaluate our method, we compile a new dataset of geometry questions (textual descriptions and diagrams) and compare with baselines that utilize standard vision techniques. Our experimental evaluation shows an F1 boost of more than 17% in identifying visual elements and 25% in aligning visual elements with their textual descriptions.

Health sensing through smartphones has received considerable attention in recent years because of the devices' ubiquity and promise to lower the barrier for tracking medical conditions. In this paper, we focus on using smartphones to monitor newborn jaundice, which manifests as a yellow discoloration of the skin. Although a degree of jaundice is common in healthy newborns, early detection of extreme jaundice is essential to prevent permanent brain damage or death. Current detection techniques, however, require clinical tests with blood samples or other specialized equipment. Consequently, newborns often depend on visual assessments of their skin color at home, which is known to be unreliable. To this end, we present BiliCam, a low-cost system that uses smartphone cameras to assess newborn jaundice. We evaluated BiliCam on 100 newborns, yielding a 0.85 rank order correlation with the gold standard blood test. We also discuss usability challenges and design solutions to make the system practical.