Kyungjae Lee

Joel Jang, Seungone Kim, Seonghyeon Ye et al. · cmu, uw

Recently, Language Models (LMs) instruction-tuned on multiple tasks, also known as multitask-prompted fine-tuning (MT), have shown the capability to generalize to unseen tasks. Previous work has shown that scaling the number of training tasks is the key component in making stronger MT LMs. In this work, we report an unexpected finding that an expert LM fine-tuned 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 a mean accuracy 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 a separate expert LM per training task 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. The code is available at https://github.com/joeljang/ELM.

Kyungjae Lee, Wookje Han, Seung-won Hwang et al.

Language models (LMs) have shown great potential as implicit knowledge bases (KBs). And for their practical use, knowledge in LMs need to be updated periodically. However, existing tasks to assess LMs' efficacy as KBs do not adequately consider multiple large-scale updates. To this end, we first propose a novel task--Continuously-updated QA (CuQA)--in which multiple large-scale updates are made to LMs, and the performance is measured with respect to the success in adding and updating knowledge while retaining existing knowledge. We then present LMs with plug-in modules that effectively handle the updates. Experiments conducted on zsRE QA and NQ datasets show that our method outperforms existing approaches. We find that our method is 4x more effective in terms of updates/forgets ratio, compared to a fine-tuning baseline.

Kyungjae Lee, Sang-eun Han, Seung-won Hwang et al.

This paper studies the problem of open-domain question answering, with the aim of answering a diverse range of questions leveraging knowledge resources. Two types of sources, QA-pair and document corpora, have been actively leveraged with the following complementary strength. The former is highly precise when the paraphrase of given question $q$ was seen and answered during training, often posed as a retrieval problem, while the latter generalizes better for unseen questions. A natural follow-up is thus leveraging both models, while a naive pipelining or integration approaches have failed to bring additional gains over either model alone. Our distinction is interpreting the problem as calibration, which estimates the confidence of predicted answers as an indicator to decide when to use a document or QA-pair corpus. The effectiveness of our method was validated on widely adopted benchmarks such as Natural Questions and TriviaQA.

Seungone Kim, Juyoung Suk, Shayne Longpre et al. · cmu

Proprietary LMs such as GPT-4 are often employed to assess the quality of responses from various LMs. However, concerns including transparency, controllability, and affordability strongly motivate the development of open-source LMs specialized in evaluations. On the other hand, existing open evaluator LMs exhibit critical shortcomings: 1) they issue scores that significantly diverge from those assigned by humans, and 2) they lack the flexibility to perform both direct assessment and pairwise ranking, the two most prevalent forms of assessment. Additionally, they do not possess the ability to evaluate based on custom evaluation criteria, focusing instead on general attributes like helpfulness and harmlessness. To address these issues, we introduce Prometheus 2, a more powerful evaluator LM than its predecessor that closely mirrors human and GPT-4 judgements. Moreover, it is capable of processing both direct assessment and pair-wise ranking formats grouped with a user-defined evaluation criteria. On four direct assessment benchmarks and four pairwise ranking benchmarks, Prometheus 2 scores the highest correlation and agreement with humans and proprietary LM judges among all tested open evaluator LMs. Our models, code, and data are all publicly available at https://github.com/prometheus-eval/prometheus-eval.

Sang-eun Han, Yeonseok Jeong, Seung-won Hwang et al.

Question answering (QA) is a critical task for speech-based retrieval from knowledge sources, by sifting only the answers without requiring to read supporting documents. Specifically, open-domain QA aims to answer user questions on unrestricted knowledge sources. Ideally, adding a source should not decrease the accuracy, but we find this property (denoted as "monotonicity") does not hold for current state-of-the-art methods. We identify the cause, and based on that we propose Judge-Specialist framework. Our framework consists of (1) specialist retrievers/readers to cover individual sources, and (2) judge, a dedicated language model to select the final answer. Our experiments show that our framework not only ensures monotonicity, but also outperforms state-of-the-art multi-source QA methods on Natural Questions. Additionally, we show that our models robustly preserve the monotonicity against noise from speech recognition. We publicly release our code and setting.

Takyoung Kim, Kyungjae Lee, Young Rok Jang et al.

Interactions with large language models (LLMs) often yield long and detailed responses, leveraging both parametric knowledge and retrieval-augmented generation (RAG). While these responses can provide rich insights, they often include redundant or less engaging content not aligned with user interests. This issue becomes apparent when users specify particular subtopics to include or exclude -- termed coverage-conditioned ($C^2$) queries -- as LLMs often struggle to provide tailored responses. To address this challenge, we investigate the role of query outlines, sequences of subqueries designed to guide LLMs in generating responses that meet specific user requirements. To systematically create and evaluate these outlines, we introduce QTree, a dataset of 10K hierarchical sets of information-seeking subqueries that define structured boundaries for outline creation and evaluation in $C^2$ scenarios. Additionally, we develop QPlanner, a 7B language model trained to generate customized outlines within boundaries of QTree. We evaluate the effectiveness of the generated outlines through automatic and human judgements, focusing on their impact within retrieval-augmented generation (RAG) systems. Experimental results demonstrate that QPlanner, especially when trained with alignment techniques like DPO, generates higher-quality outlines that better fulfill diverse user needs.

Wookje Han, Jinsol Park, Kyungjae Lee

Information-seeking questions in long-form question answering (LFQA) often prove misleading due to ambiguity or false presupposition in the question. While many existing approaches handle misleading questions, they are tailored to limited questions, which are insufficient in a real-world setting with unpredictable input characteristics. In this work, we propose PreWoMe, a unified approach capable of handling any type of information-seeking question. The key idea of PreWoMe involves extracting presuppositions in the question and exploiting them as working memory to generate feedback and action about the question. Our experiment shows that PreWoMe is effective not only in tackling misleading questions but also in handling normal ones, thereby demonstrating the effectiveness of leveraging presuppositions, feedback, and action for real-world QA settings.

Hyeon Hwang, Yewon Cho, Chanwoong Yoon et al.

Step-by-step reasoning has become a standard approach for large language models (LLMs) to tackle complex tasks. While this paradigm has proven effective, it raises a fundamental question: How can we verify that an LLM's reasoning is accurately grounded in knowledge? To address this question, we introduce a novel evaluation suite that systematically assesses the knowledge grounding of intermediate reasoning. Our framework comprises three key components. (1) Principal Knowledge Collection, a large-scale repository of atomic knowledge essential for reasoning. Based on the collection, we propose (2) knowledge-grounded evaluation metrics designed to measure how well models recall and apply prerequisite knowledge in reasoning. These metrics are computed by our (3) evaluator LLM, a lightweight model optimized for cost-effective and reliable metric computation. Our evaluation suite demonstrates remarkable effectiveness in identifying missing or misapplied knowledge elements, providing crucial insights for uncovering fundamental reasoning deficiencies in LLMs. Beyond evaluation, we demonstrate how these metrics can be integrated into preference optimization, showcasing further applications of knowledge-grounded evaluation.

Seungone Kim, Juyoung Suk, Ji Yong Cho et al.

As language models (LMs) become capable of handling a wide range of tasks, their evaluation is becoming as challenging as their development. Most generation benchmarks currently assess LMs using abstract evaluation criteria like helpfulness and harmlessness, which often lack the flexibility and granularity of human assessment. Additionally, these benchmarks tend to focus disproportionately on specific capabilities such as instruction following, leading to coverage bias. To overcome these limitations, we introduce the BiGGen Bench, a principled generation benchmark designed to thoroughly evaluate nine distinct capabilities of LMs across 77 diverse tasks. A key feature of the BiGGen Bench is its use of instance-specific evaluation criteria, closely mirroring the nuanced discernment of human evaluation. We apply this benchmark to assess 103 frontier LMs using five evaluator LMs. Our code, data, and evaluation results are all publicly available at https://github.com/prometheus-eval/prometheus-eval/tree/main/BiGGen-Bench.

Dohyeok Lee, Jung Min Lee, Munkyung Kim et al.

Behavior cloning methods for robot learning suffer from poor generalization due to limited data support beyond expert demonstrations. Recent approaches leveraging video prediction models have shown promising results by learning rich spatiotemporal representations from large-scale datasets. However, these models learn action-agnostic dynamics that cannot distinguish between different control inputs, limiting their utility for precise manipulation tasks and requiring large pretraining datasets. We propose a Dynamics-Aligned Flow Matching Policy (DAP) that integrates dynamics prediction into policy learning. Our method introduces a novel architecture where policy and dynamics models provide mutual corrective feedback during action generation, enabling self-correction and improved generalization. Empirical validation demonstrates generalization performance superior to baseline methods on real-world robotic manipulation tasks, showing particular robustness in OOD scenarios including visual distractions and lighting variations.

Jeongeun Park, Jihwan Yoon, Byungwoo Jeon et al.

Prior Vision-Language-Action (VLA) models are typically trained on teleoperated successful demonstrations, while discarding numerous failed attempts that occur naturally during data collection. However, these failures encode where and how policies can be fragile, information that can be exploited to improve robustness. We address this problem by leveraging mixed-quality datasets to learn failure-aware reasoning at planning time. We introduce VINE, a hierarchical vision-language-action model that separates high-level reasoning (System 2) from low-level control (System 1) under a hierarchical reinforcement learning formalism, making failures usable as a structured learning signal rather than noisy supervision. System 2 performs feasibility-guided tree search over a 2D scene-graph abstraction: it proposes subgoal transitions, predicts success probabilities from both successes and failures, and prunes brittle branches before execution, effectively casting plan evaluation as feasibility scoring. The selected subgoal sequence is then passed to System 1, which executes low-level actions without modifying the agent's core skills. Trained entirely from offline teleoperation data, VINE integrates negative experience directly into the decision loop. Across challenging manipulation tasks, this approach consistently improves success rates and robustness, demonstrating that failure data is an essential resource for converting the broad competence of VLAs into robust execution.

Hajung Kim, Chanhwi Kim, Hoonick Lee et al.

Transforming natural language questions into SQL queries is crucial for precise data retrieval from electronic health record (EHR) databases. A significant challenge in this process is detecting and rejecting unanswerable questions that request information beyond the database's scope or exceed the system's capabilities. In this paper, we introduce a novel text-to-SQL framework that robustly handles out-of-domain questions and verifies the generated queries with query execution.Our framework begins by standardizing the structure of questions into a templated format. We use a powerful large language model (LLM), fine-tuned GPT-3.5 with detailed prompts involving the table schemas of the EHR database system. Our experimental results demonstrate the effectiveness of our framework on the EHRSQL-2024 benchmark benchmark, a shared task in the ClinicalNLP workshop. Although a straightforward fine-tuning of GPT shows promising results on the development set, it struggled with the out-of-domain questions in the test set. With our framework, we improve our system's adaptability and achieve competitive performances in the official leaderboard of the EHRSQL-2024 challenge.

Hyungjoo Chae, Dongjin Kang, Jihyuk Kim et al. · gatech

With the release of R1, a publicly available large reasoning model (LRM), researchers commonly train new LRMs by training language models on R1's long chain-of-thought (CoT) inferences. While prior works show that LRMs' capabilities can be reproduced through direct distillation, the continued reliance on the existing models (e.g., R1) remains a critical limitation in advancing the field. As a first step toward independent LRM development, this paper explores the possibility of constructing a long CoT dataset with LLMs that are not trained for inference-time scaling. To this end, we present the Long CoT Collection, a dataset of 100K CoT rationales annotated using existing short CoT LLMs. We develop a pipeline that induces o1's novel reasoning strategies into short CoT LLMs, enabling them to think longer and introducing controllability over the thought budget to better manage the overthinking problem. Our extensive analyses validate that our dataset achieves quality comparable to--or slightly below--R1. Furthermore, our experiments demonstrate that training on our dataset not only strengthens general reasoning skills, but also provides a strong foundation for reinforcement learning--models initialized on our data achieve 2-3x larger gains with RLVR.

Kyungjae Lee, Seung-won Hwang, Sang-eun Han et al.

This paper studies the bias problem of multi-hop question answering models, of answering correctly without correct reasoning. One way to robustify these models is by supervising to not only answer right, but also with right reasoning chains. An existing direction is to annotate reasoning chains to train models, requiring expensive additional annotations. In contrast, we propose a new approach to learn evidentiality, deciding whether the answer prediction is supported by correct evidences, without such annotations. Instead, we compare counterfactual changes in answer confidence with and without evidence sentences, to generate "pseudo-evidentiality" annotations. We validate our proposed model on an original set and challenge set in HotpotQA, showing that our method is accurate and robust in multi-hop reasoning.

Jihyeok Kim, Reinald Kim Amplayo, Kyungjae Lee et al.

The performance of text classification has improved tremendously using intelligently engineered neural-based models, especially those injecting categorical metadata as additional information, e.g., using user/product information for sentiment classification. These information have been used to modify parts of the model (e.g., word embeddings, attention mechanisms) such that results can be customized according to the metadata. We observe that current representation methods for categorical metadata, which are devised for human consumption, are not as effective as claimed in popular classification methods, outperformed even by simple concatenation of categorical features in the final layer of the sentence encoder. We conjecture that categorical features are harder to represent for machine use, as available context only indirectly describes the category, and even such context is often scarce (for tail category). To this end, we propose to use basis vectors to effectively incorporate categorical metadata on various parts of a neural-based model. This additionally decreases the number of parameters dramatically, especially when the number of categorical features is large. Extensive experiments on various datasets with different properties are performed and show that through our method, we can represent categorical metadata more effectively to customize parts of the model, including unexplored ones, and increase the performance of the model greatly.

Reinald Kim Amplayo, Kyungjae Lee, Jinyeong Yeo et al.

In sentence classification tasks, additional contexts, such as the neighboring sentences, may improve the accuracy of the classifier. However, such contexts are domain-dependent and thus cannot be used for another classification task with an inappropriate domain. In contrast, we propose the use of translated sentences as context that is always available regardless of the domain. We find that naive feature expansion of translations gains only marginal improvements and may decrease the performance of the classifier, due to possible inaccurate translations thus producing noisy sentence vectors. To this end, we present multiple context fixing attachment (MCFA), a series of modules attached to multiple sentence vectors to fix the noise in the vectors using the other sentence vectors as context. We show that our method performs competitively compared to previous models, achieving best classification performance on multiple data sets. We are the first to use translations as domain-free contexts for sentence classification.