Reward models (RMs) play a central role in aligning large language models (LLMs) with human preferences.However, RMs are often sensitive to spurious features such as response length.Existing inference-time approaches to mitigating these biases typically focus exclusively on response length, resulting in performance trade-offs.In this paper, we propose causally motivated intervention for mitigating multiple types of biases in RMs at inference time.Our method first identifies neurons whose activations are strongly correlated with predefined bias attributes, and applies neuron-level intervention that suppresses these signals.We evaluate our method on RM benchmarks and observe reductions in sensitivity to spurious features across diverse bias types, without inducing performance trade-offs.Moreover, when used for preference annotation, small RMs (2B and 7B) with our method, which edits less than 2% of all the neurons in RMs, enable LLMs to improve alignment, achieving performance comparable to that of a state-of-the-art 70B RM on AlpacaEval and MT-Bench.Further analysis reveals that bias signals are primarily encoded by neurons in early layers, shedding light on the internal mechanisms of bias exploitation in RMs.

Recent concept-based interpretable models have succeeded in providing meaningful explanations by pre-defined concept sets. However, the dependency on the pre-defined concepts restricts the application because of the limited number of concepts for explanations. This paper proposes a novel interpretable deep neural network called explanation bottleneck models (XBMs). XBMs generate a text explanation from the input without pre-defined concepts and then predict a final task prediction based on the generated explanation by leveraging pre-trained vision-language encoder-decoder models. To achieve both the target task performance and the explanation quality, we train XBMs through the target task loss with the regularization penalizing the explanation decoder via the distillation from the frozen pre-trained decoder. Our experiments, including a comparison to state-of-the-art concept bottleneck models, confirm that XBMs provide accurate and fluent natural language explanations without pre-defined concept sets.

Loss spikes, a phenomenon in which the loss value diverges suddenly, is a fundamental issue in the pre-training of large language models.This paper supposes that the non-uniformity of the norm of the parameters is one of the causes of loss spikes.Here, in training of neural networks, the scale of the gradients is required to be kept constant throughout the layers to avoid the vanishing and exploding gradients problem.However, to meet these requirements in the Transformer model, the norm of the model parameters must be non-uniform, and thus, parameters whose norm is smaller are more sensitive to the parameter update.To address this issue, we propose a novel technique, weight scaling as reparameterization (WeSaR).WeSaR introduces a gate parameter per parameter matrix and adjusts it to the value satisfying the requirements.Because of the gate parameter, WeSaR sets the norm of the original parameters uniformly, which results in stable training.Experimental results with the Transformer decoders consisting of 130 million, 1.3 billion, and 13 billion parameters showed that WeSaR stabilizes and accelerates training and that it outperformed compared methods including popular initialization methods.

Visual question answering on document images that contain textual, visual, and layout information, called document VQA, has received much attention recently. Although many datasets have been proposed for developing document VQA systems, most of the existing datasets focus on understanding the content relationships within a single image and not across multiple images. In this study, we propose a new multi-image document VQA dataset, SlideVQA, containing 2.6k+ slide decks composed of 52k+ slide images and 14.5k questions about a slide deck. SlideVQA requires complex reasoning, including single-hop, multi-hop, and numerical reasoning, and also provides annotated arithmetic expressions of numerical answers for enhancing the ability of numerical reasoning. Moreover, we developed a new end-to-end document VQA model that treats evidence selection and question answering as a unified sequence-to-sequence format. Experiments on SlideVQA show that our model outperformed existing state-of-the-art QA models, but that it still has a large gap behind human performance. We believe that our dataset will facilitate research on document VQA.

Although named entity recognition (NER) helps us to extract domain-specific entities from text (e.g., artists in the music domain), it is costly to create a large amount of training data or a structured knowledge base to perform accurate NER in the target domain.Here, we propose selfadaptive NER, which retrieves external knowledge from unstructured text to learn the usages of entities that have not been learned well.To retrieve useful knowledge for NER, we design an effective two-stage model that retrieves unstructured knowledge using uncertain entities as queries.Our model predicts the entities in the input and then finds those of which the prediction is not confident.Then, it retrieves knowledge by using these uncertain entities as queries and concatenates the retrieved text to the original input to revise the prediction.Experiments on CrossNER datasets demonstrated that our model outperforms strong baselines by 2.35 points in F 1 metric.

Multi-hop QA with annotated supporting facts, which is the task of reading comprehension (RC) considering the interpretability of the answer, has been extensively studied. In this study, we define an interpretable reading comprehension (IRC) model as a pipeline model with the capability of predicting unanswerable queries. The IRC model justifies the answer prediction by establishing consistency between the predicted supporting facts and the actual rationale for interpretability. The IRC model detects unanswerable questions, instead of outputting the answer forcibly based on the insufficient information, to ensure the reliability of the answer. We also propose an end-to-end training method for the pipeline RC model. To evaluate the interpretability and the reliability, we conducted the experiments considering unanswerability in a multi-hop question for a given passage. We show that our end-to-end trainable pipeline model outperformed a non-interpretable model on our modified HotpotQA dataset. Experimental results also show that the IRC model achieves comparable results to the previous non-interpretable models in spite of the trade-off between prediction performance and interpretability.

This study tackles unsupervised domain adaptation of reading comprehension (UDARC). Reading comprehension (RC) is a task to learn the capability for question answering with textual sources. State-of-the-art models on RC still do not have general linguistic intelligence; i.e., their accuracy worsens for out-domain datasets that are not used in the training. We hypothesize that this discrepancy is caused by a lack of the language modeling (LM) capability for the out-domain. The UDARC task allows models to use supervised RC training data in the source domain and only unlabeled passages in the target domain. To solve the UDARC problem, we provide two domain adaptation models. The first one learns the out-domain LM and in-domain RC task sequentially. The second one is the proposed model that uses a multi-task learning approach of LM and RC. The models can retain both the RC capability acquired from the supervised data in the source domain and the LM capability from the unlabeled data in the target domain. We evaluated the models on UDARC with five datasets in different domains. The models outperformed the model without domain adaptation. In particular, the proposed model yielded an improvement of 4.3/4.2 points in EM/F1 in an unseen biomedical domain.

Kosuke Nishida, Kyosuke Nishida, Masaaki Nagata, Atsushi Otsuka, Itsumi Saito, Hisako Asano, Junji Tomita. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

Kyosuke Nishida, Itsumi Saito, Kosuke Nishida, Kazutoshi Shinoda, Atsushi Otsuka, Hisako Asano, Junji Tomita. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.