Evaluating the quality of LLM-generated reasoning traces in expert domains (e.g., law) is essential for ensuring credibility and explainability, yet remains challenging due to the inherent complexity of such reasoning tasks. We introduce LEGIT (LEGal Issue Trees), a novel large-scale (24K instances) expert-level legal reasoning dataset with an emphasis on reasoning trace evaluation. We convert court judgments into hierarchical trees of opposing parties' arguments and the court's conclusions, which serve as rubrics for evaluating the issue coverage and correctness of the reasoning traces. We verify the reliability of these rubrics via human expert annotations and comparison with coarse, less informative rubrics. Using the LEGIT dataset, we show that (1) LLMs' legal reasoning ability is seriously affected by both legal issue coverage and correctness, and that (2) retrieval-augmented generation (RAG) and RL with rubrics bring complementary benefits for legal reasoning abilities, where RAG improves overall reasoning capability, whereas RL improves correctness albeit with reduced coverage.

In real-world services such as ChatGPT, aligning models based on user feedback is crucial for improving model performance.However, due to the simplicity and convenience of providing feedback, users typically offer only basic binary signals, such as 'thumbs-up' or 'thumbsdown'.Most existing alignment research, on the other hand, relies on preference-based approaches that require both positive and negative responses as a pair.We propose Binary Classifier Optimization (BCO), a technique that effectively aligns LLMs using only binary feedback.BCO trains a binary classifier, where the logit serves as an implicit reward, effectively minimizing the Direct Preference Optimization (DPO) loss.We demonstrate that the binary cross-entropy loss employed in classifier training acts as an upper bound for the DPO loss.Additionally, a novel reward shift technique further minimizes the gap between the losses.We validate our methodology in two settings: first, on a paired preference dataset, where our method performs on par with DPO; and second, on a Likert-5 scale annotation dataset which stems from real users' queries.Our model consistently demonstrates effective and robust alignment across four base LLMs and three different datasets, showcasing the strength of our approach to learning from binary signals.

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.

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.

Foundation models have shown outstanding performance and generalization capabilities across domains. Since most studies on foundation models mainly focus on the pretraining phase, a naive strategy to minimize a single task-specific loss is adopted for fine-tuning. However, such fine-tuning methods do not fully leverage other losses that are potentially beneficial for the target task. Therefore, we propose MEta Loss TRansformer (MELTR), a plug-in module that automatically and non-linearly combines various loss functions to aid learning the target task via auxiliary learning. We formulate the auxiliary learning as a bi-level optimization problem and present an efficient optimization algorithm based on Approximate Implicit Differentiation (AID). For evaluation, we apply our framework to various video foundation models (UniVL, Violet and All-in-one), and show significant performance gain on all four downstream tasks: text-to-video retrieval, video question answering, video captioning, and multimodal sentiment analysis. Our qualitative analyses demonstrate that MELTR adequately 'transforms' individual loss functions and 'melts' them into an effective unified loss. Code is available at https://github.com/mlvlab/MELTR.

Human-Object Interaction (HOI) detection is the task of identifying a set of (human, object, interaction) triplets from an image. Recent work proposed transformer encoder-decoder architectures that successfully eliminated the need for many hand-designed components in HOI detection through end-to-end training. However, they are limited to single-scale feature resolution, providing suboptimal performance in scenes containing humans, objects, and their interactions with vastly different scales and distances. To tackle this problem, we propose a Multi-Scale TRansformer (MSTR) for HOI detection powered by two novel HOI-aware deformable attention modules called Dual-Entity attention and Entity-conditioned Context attention. While existing deformable attention comes at a huge cost in HOI detection performance, our proposed attention modules of MSTR learn to effectively attend to sampling points that are essential to identify interactions. In experiments, we achieve the new state-of-the-art performance on two HOI detection benchmarks.

Learning generic joint representations for video and text by a supervised method requires a prohibitively substantial amount of manually annotated video datasets. As a practical alternative, a large-scale but uncurated and narrated video dataset, HowTo100M, has recently been introduced. But it is still challenging to learn joint embeddings of video and text in a self-supervised manner, due to its ambiguity and non-sequential alignment. In this paper, we propose a novel multi-modal self-supervised framework Video-Text Temporally Weak Alignment-based Contrastive Learning (VT-TWINS) to capture significant information from noisy and weakly correlated data using a variant of Dynamic Time Warping (DTW). We observe that the standard DTW inherently cannot handle weakly correlated data and only considers the globally optimal alignment path. To address these problems, we develop a differentiable DTW which also reflects local information with weak temporal alignment. Moreover, our proposed model applies a contrastive learning scheme to learn feature representations on weakly correlated data. Our extensive experiments demonstrate that VT-TWINS attains significant improvements in multi-modal representation learning and outperforms various challenging downstream tasks. Code is available at https://github.com/mlvlab/VT-Twins.

Despite recent progress on computer vision and natural language processing, developing a machine that can understand video story is still hard to achieve due to the intrinsic difficulty of video story. Moreover, researches on how to evaluate the degree of video understanding based on human cognitive process have not progressed as yet. In this paper, we propose a novel video question answering (Video QA) task, DramaQA, for a comprehensive understanding of the video story. The DramaQA focuses on two perspectives: 1) Hierarchical QAs as an evaluation metric based on the cognitive developmental stages of human intelligence. 2) Character-centered video annotations to model local coherence of the story. Our dataset is built upon the TV drama "Another Miss Oh" and it contains 17,983 QA pairs from 23,928 various length video clips, with each QA pair belonging to one of four difficulty levels. We provide 217,308 annotated images with rich character-centered annotations, including visual bounding boxes, behaviors and emotions of main characters, and coreference resolved scripts. Additionally, we suggest Multi-level Context Matching model which hierarchically understands character-centered representations of video to answer questions. We release our dataset and model publicly for research purposes, and we expect our work to provide a new perspective on video story understanding research.

Conventional sequential learning methods such as Recurrent Neural Networks (RNNs) focus on interactions between consecutive inputs, i.e. first-order Markovian dependency. However, most of sequential data, as seen with videos, have complex dependency structures that imply variable-length semantic flows and their compositions, and those are hard to be captured by conventional methods. Here, we propose Cut-Based Graph Learning Networks (CB-GLNs) for learning video data by discovering these complex structures of the video. The CB-GLNs represent video data as a graph, with nodes and edges corresponding to frames of the video and their dependencies respectively. The CB-GLNs find compositional dependencies of the data in multilevel graph forms via a parameterized kernel with graph-cut and a message passing framework. We evaluate the proposed method on the two different tasks for video understanding: Video theme classification (Youtube-8M dataset (Abu-El-Haija et al. 2016)) and Video Question and Answering (TVQA dataset(Lei et al. 2018)). The experimental results show that our model efficiently learns the semantic compositional structure of video data. Furthermore, our model achieves the highest performance in comparison to other baseline methods.

One of the fundamental problems that arise in multimodal learning tasks is the disparity of information levels between different modalities. To resolve this problem, we propose Hypergraph Attention Networks (HANs), which define a common semantic space among the modalities with symbolic graphs and extract a joint representation of the modalities based on a co-attention map constructed in the semantic space. HANs follow the process: constructing the common semantic space with symbolic graphs of each modality, matching the semantics between sub-structures of the symbolic graphs, constructing co-attention maps between the graphs in the semantic space, and integrating the multimodal inputs using the co-attention maps to get the final joint representation. From the qualitative analysis with two Visual Question and Answering datasets, we discover that 1) the alignment of the information levels between the modalities is important, and 2) the symbolic graphs are very powerful ways to represent the information of the low-level signals in alignment. Moreover, HANs dramatically improve the state-of-the-art accuracy on the GQA dataset from 54.6\% to 61.88\% only using the symbolic information in quantitatively.

In this paper, we consider a problem of analyzing human behavioral data to predict the human cognitive states and generate corresponding actions of sever-agent. Specifically, we aim at predicting human cognitive states during meal time and generating relevant dining services for the human. For this study, we collect behavioral data using 2 kinds of wearable devices, which are an eye tracker and a watch type EDA device, during meal time. We focus on the characteristics of the behavioral data, which are heterogeneous, noisy and temporal, and suggest a novel machine learning algorithm which can analyze the data integrally. Suggested model has hierarchical structure: the bottom layer combines the multi-modal behavioral data based on causal structure of the data and extracts the feature vector. Using the extracted feature vectors, the upper layer predicts the cognitive states based on temporal correlation between feature vectors. Experimental results show that the suggested model can analyze the behavioral data efficiently and predict the human cognitive states correctly.