Publications

LLMs sometimes repeat, repeat, repeat, repeat themselves. To uncover the internal mechanism, this paper introduces "repetition neurons", regarded as skill neurons responsible for the repetition problem in text generation tasks.  These neurons become progressively more active as the repetition continues.

Recent work has analyzed how knowledge is represented in activation space. Building on this line of inquiry, this paper examines how LLMs leverage the linear subspace of entity-numerical attributes when answering questions involving numeric comparisons—for example, "Was Cristiano born before Messi?"

In LLMs, early layers transform subword tokens into more meaningful representations that form the model's inner vocabulary. This paper demonstrates that several important aspects of this detokenization stage can be understood purely by analyzing model weights.

This paper introduces MQM-Chat, a new multidimensional quality metric designed specifically for chat translations. It encompasses seven error types, including three that are unique to chat translations, allowing for the evaluation of both lexical and semantic accuracy in chat translation tasks.

NLP datasets may still contain annotation errors, even when they are manually annotated. This paper introduces SubRegWeigh, a time-saving method that leverages subword regularization to simulate multiple error detection models for identifying annotation errors in NLP datasets.

Assessing the quality of free-text explanations remains challenging for LLMs due to their multifaceted, subjective, and labor-intensive nature. This paper introduces ACORN, a dataset of free-text explanations paired with aspect-wise quality ratings.

How is factual knowledge involving numeric properties such as "Karl Popper was born in 1902" encoded in the model’s internal representations? This paper introduces a method for finding and editing representations of numeric properties by identifying interpretable, monotonically encoded directions.

During interactions with a model, contradictory responses can undermine user trust and disrupt dialogue coherence. This paper tackles that challenge head-on by constructing the first large-scale dataset of model-generated contradictions.

Interpreting the internals of Transformer models remains a pivotal challenge. This paper examines the role of feed-forward (FF) blocks in Transformers, which was unexplored, by visualizing their impact on input contextualization through attention maps.

This paper delves into the mechanism of in-context entity tracking, exploring how language models seamlessly bind entities to their corresponding attributes from a given context. It newly introduces the Ordering ID captured by entity activations, which directly determines binding behavior.

This paper reports on the systematic strategy that language models employ in multi-step reasoning processes, revealing that they rely on heuristics such as lexical overlap in the early stages, with this reliance diminishing as they advance toward the final answer.

In education, counterarguments are used to boost critical thinking skills, but giving personalized feedback to students can be challenging for teachers. This paper introduces Counter-Argument Logical Structure Analysis (CALSA), a novel approach that breaks down counterarguments in debates into ten clear logical patterns, offering a promising pathway toward automating effective feedback.

Language models encode world knowledge in their internal parameters through training. This paper investigates the deletion of such encoded knowledge from the model and analyzes the relationship between deletion side effects and the associated entities using a synthetic knowledge graph.

Vision models such as CLIP have been used in research on diagram understanding and generation. This study examines an unexplored capability of these models, specifically, whether they can accurately identify diagram attributes including node colors, shapes, edge colors, and connection patterns.

While current models can assess argument quality, they often fail to provide constructive feedback explaining the basis of their evaluations. This survey explores current NLP feedback systems by categorizing them into four key dimensions—Richness, Visualization, Interactivity, and Personalization.

The performance of sentence encoders can be significantly improved by fine-tuning with contrastive loss. However, what characteristics do models acquire during contrastive learning? This paper reveals these characteristics by shedding light on the inner workings of word weighting.

To create a translation system that excels across diverse domains, this study employs a Multiple Expert Models Collaboration strategy that aggregates the specialized knowledge of individual domain-specific experts and validates the effectiveness.

Factual probing, a method that uses prompts to assess a model's world knowledge, faces the challenge that slight variations in the prompt can drastically change the results. To tackle this issue, the study introduces test-time augmentation, which augments and ensembles prompts to reduce sensitivity.

While answers to questions can evolve over time, traditional QA datasets have been built on static assumptions. This paper presents REALTIME QA, a dynamic platform that announces questions and evaluates systems on a regular basis.

This paper introduces a novel strategy to enhance the trustworthiness of "Short Answer Scoring" systems in educational settings by aligning response features with rubric criteria to mitigate shortcut learning based on superficial cues in the training data.

Reducing the memory required by word embeddings while still maintaining performance is crucial given their large vocabulary and high dimensionality. This paper presents a new compression method that uses a deterministic convex optimization process to produce stable and reproducible representations.

Why do LLMs sometimes generate incorrect answers, even as their capabilities continue to improve? This study hypothesizes that the model's flawed reasoning stems from its reliance on incorrect beliefs, and it demonstrates that correcting the belief space leads to improved performance.

This paper introduces a new method called the self-referencing causal cycle (RECALL). It helps large language models avoid the reversal curse, which is a problem that comes from one-way reasoning. RECALL uses special "cycle tokens" to connect different parts of the training data.

This paper asks whether the long-range dependencies that define human language also appear in animal communication. It employs Transformer models to explore this idea in Bengalese finch songs, which are marked by highly variable and complex syllable sequences.

This paper challenges recent claims from cognitive modeling studies that LLMs poorly align with human reading behavior, showing that focusing solely on their final layers can be misleading, as revealed by a mechanistic analysis of their internal layers.