arXiv subjects

Kentaro Inui

Publications and source records attributed to Kentaro Inui.

At least 37 records · Page 2Linked to original sources

arXiv2026

Sycophancy Hides Linearly in the Attention Heads

We find that correct-to-incorrect sycophancy signals are most linearly separable within multi-head attention activations. Motivated by the linear representation hypothesis, we train linear probes across the residual stream, multilayer perceptron (MLP), and attention layers to analyze where these signals emerge. Although separability appears in the residual stream and MLPs, steering using these probes is most effective in a sparse subset of middle-layer attention heads. Using TruthfulQA as the base dataset, we find that probes trained on it transfer effectively to other factual QA benchmarks. Furthermore, comparing our discovered direction to previously identified "truthful" directions reveals limited overlap, suggesting that factual accuracy, and deference resistance, arise from related but distinct mechanisms. Attention-pattern analysis further indicates that the influential heads attend disproportionately to expressions of user doubt, contributing to sycophantic shifts. Overall, these findings suggest that sycophancy can be mitigated through simple, targeted linear interventions that exploit the internal geometry of attention activations.

cs.CL↗

arXiv2025

Can Language Models Handle a Non-Gregorian Calendar? The Case of the Japanese wareki

Temporal reasoning and knowledge are essential capabilities for language models (LMs). While much prior work has analyzed and improved temporal reasoning in LMs, most studies have focused solely on the Gregorian calendar. However, many non-Gregorian systems, such as the Japanese, Hijri, and Hebrew calendars, are in active use and reflect culturally grounded conceptions of time. If and how well current LMs can accurately handle such non-Gregorian calendars has not been evaluated so far. Here, we present a systematic evaluation of how well language models handle one such non-Gregorian system: the Japanese wareki. We create datasets that require temporal knowledge and reasoning in using wareki dates. Evaluating open and closed LMs, we find that some models can perform calendar conversions, but GPT-4o, Deepseek V3, and even Japanese-centric models struggle with Japanese calendar arithmetic and knowledge involving wareki dates. Error analysis suggests corpus frequency of Japanese calendar expressions and a Gregorian bias in the model's knowledge as possible explanations. Our results show the importance of developing LMs that are better equipped for culture-specific tasks such as calendar understanding.

cs.CL↗

arXiv2025

Understanding and Controlling Repetition Neurons and Induction Heads in In-Context Learning

This paper investigates the relationship between large language models' (LLMs) ability to recognize repetitive input patterns and their performance on in-context learning (ICL). In contrast to prior work that has primarily focused on attention heads, we examine this relationship from the perspective of skill neurons, specifically repetition neurons. Our experiments reveal that the impact of these neurons on ICL performance varies depending on the depth of the layer in which they reside. By comparing the effects of repetition neurons and induction heads, we further identify strategies for reducing repetitive outputs while maintaining strong ICL capabilities.

cs.CL↗

arXiv2025

SPIRIT: Patching Speech Language Models against Jailbreak Attacks

Speech Language Models (SLMs) enable natural interactions via spoken instructions, which more effectively capture user intent by detecting nuances in speech. The richer speech signal introduces new security risks compared to text-based models, as adversaries can better bypass safety mechanisms by injecting imperceptible noise to speech. We analyze adversarial attacks and find that SLMs are substantially more vulnerable to jailbreak attacks, which can achieve a perfect 100% attack success rate in some instances. To improve security, we propose post-hoc patching defenses used to intervene during inference by modifying the SLM's activations that improve robustness up to 99% with (i) negligible impact on utility and (ii) without any re-training. We conduct ablation studies to maximize the efficacy of our defenses and improve the utility/security trade-off, validated with large-scale benchmarks unique to SLMs.

eess.AS↗

arXiv2025

How a Bilingual LM Becomes Bilingual: Tracing Internal Representations with Sparse Autoencoders

This study explores how bilingual language models develop complex internal representations. We employ sparse autoencoders to analyze internal representations of bilingual language models with a focus on the effects of training steps, layers, and model sizes. Our analysis shows that language models first learn languages separately, and then gradually form bilingual alignments, particularly in the mid layers. We also found that this bilingual tendency is stronger in larger models. Building on these findings, we demonstrate the critical role of bilingual representations in model performance by employing a novel method that integrates decomposed representations from a fully trained model into a mid-training model. Our results provide insights into how language models acquire bilingual capabilities.

cs.CL↗

arXiv2025

SmallPlan: Leverage Small Language Models for Sequential Path Planning with Simulation-Powered, LLM-Guided Distillation

Efficient path planning in robotics, particularly within large-scale, complex environments, remains a significant hurdle. While Large Language Models (LLMs) offer strong reasoning capabilities, their high computational cost and limited adaptability hinder real-time deployment on edge devices. We present SmallPlan - a novel framework leveraging LLMs as teacher models to train lightweight Small Language Models (SLMs) for high-level path planning tasks. In SmallPlan, the SLMs provide optimal action sequences to navigate across scene graphs that compactly represent full-scaled 3D scenes. The SLMs are trained in a simulation-powered, interleaved manner with LLM-guided supervised fine-tuning (SFT) and reinforcement learning (RL). This strategy not only enables SLMs to successfully complete navigation tasks but also makes them aware of important factors like distance travel, providing more efficient path planning. Through experiments, we demonstrate that the fine-tuned SLMs perform competitively with larger models like GPT-4o on sequential path planning, without suffering from hallucination and overfitting. SmallPlan is resource-efficient, making it well-suited for edge-device deployment and advancing practical autonomous robotics. Our source code is available here: https://github.com/quangpham2006/SmallPlan

cs.RO↗

arXiv2025

LLMs Can Compensate for Deficiencies in Visual Representations

Many vision-language models (VLMs) that prove very effective at a range of multimodal task, build on CLIP-based vision encoders, which are known to have various limitations. We investigate the hypothesis that the strong language backbone in VLMs compensates for possibly weak visual features by contextualizing or enriching them. Using three CLIP-based VLMs, we perform controlled self-attention ablations on a carefully designed probing task. Our findings show that despite known limitations, CLIP visual representations offer ready-to-read semantic information to the language decoder. However, in scenarios of reduced contextualization in the visual representations, the language decoder can largely compensate for the deficiency and recover performance. This suggests a dynamic division of labor in VLMs and motivates future architectures that offload more visual processing to the language decoder.

cs.CV↗

arXiv2025

Uncovering the Spectral Bias in Diagonal State Space Models

Current methods for initializing state space models (SSMs) parameters mainly rely on the \textit{HiPPO framework}, which is based on an online approximation of orthogonal polynomials. Recently, diagonal alternatives have shown to reach a similar level of performance while being significantly more efficient due to the simplification in the kernel computation. However, the \textit{HiPPO framework} does not explicitly study the role of its diagonal variants. In this paper, we take a further step to investigate the role of diagonal SSM initialization schemes from the frequency perspective. Our work seeks to systematically understand how to parameterize these models and uncover the learning biases inherent in such diagonal state-space models. Based on our observations, we propose a diagonal initialization on the discrete Fourier domain \textit{S4D-DFouT}. The insights in the role of pole placing in the initialization enable us to further scale them and achieve state-of-the-art results on the Long Range Arena benchmark, allowing us to train from scratch on very large datasets as PathX-256.

cs.LG↗

arXiv2025

Annotating Errors in English Learners' Written Language Production: Advancing Automated Written Feedback Systems

Recent advances in natural language processing (NLP) have contributed to the development of automated writing evaluation (AWE) systems that can correct grammatical errors. However, while these systems are effective at improving text, they are not optimally designed for language learning. They favor direct revisions, often with a click-to-fix functionality that can be applied without considering the reason for the correction. Meanwhile, depending on the error type, learners may benefit most from simple explanations and strategically indirect hints, especially on generalizable grammatical rules. To support the generation of such feedback, we introduce an annotation framework that models each error's error type and generalizability. For error type classification, we introduce a typology focused on inferring learners' knowledge gaps by connecting their errors to specific grammatical patterns. Following this framework, we collect a dataset of annotated learner errors and corresponding human-written feedback comments, each labeled as a direct correction or hint. With this data, we evaluate keyword-guided, keyword-free, and template-guided methods of generating feedback using large language models (LLMs). Human teachers examined each system's outputs, assessing them on grounds including relevance, factuality, and comprehensibility. We report on the development of the dataset and the comparative performance of the systems investigated.

cs.CL↗

arXiv2025

Tell Me Who Your Students Are: GPT Can Generate Valid Multiple-Choice Questions When Students' (Mis)Understanding Is Hinted

The primary goal of this study is to develop and evaluate an innovative prompting technique, AnaQuest, for generating multiple-choice questions (MCQs) using a pre-trained large language model. In AnaQuest, the choice items are sentence-level assertions about complex concepts. The technique integrates formative and summative assessments. In the formative phase, students answer open-ended questions for target concepts in free text. For summative assessment, AnaQuest analyzes these responses to generate both correct and incorrect assertions. To evaluate the validity of the generated MCQs, Item Response Theory (IRT) was applied to compare item characteristics between MCQs generated by AnaQuest, a baseline ChatGPT prompt, and human-crafted items. An empirical study found that expert instructors rated MCQs generated by both AI models to be as valid as those created by human instructors. However, IRT-based analysis revealed that AnaQuest-generated questions - particularly those with incorrect assertions (foils) - more closely resembled human-crafted items in terms of difficulty and discrimination than those produced by ChatGPT.

cs.CL↗

arXiv2025

Large Language Models Are Human-Like Internally

Recent cognitive modeling studies have reported that larger language models (LMs) exhibit a poorer fit to human reading behavior (Oh and Schuler, 2023b; Shain et al., 2024; Kuribayashi et al., 2024), leading to claims of their cognitive implausibility. In this paper, we revisit this argument through the lens of mechanistic interpretability and argue that prior conclusions were skewed by an exclusive focus on the final layers of LMs. Our analysis reveals that next-word probabilities derived from internal layers of larger LMs align with human sentence processing data as well as, or better than, those from smaller LMs. This alignment holds consistently across behavioral (self-paced reading times, gaze durations, MAZE task processing times) and neurophysiological (N400 brain potentials) measures, challenging earlier mixed results and suggesting that the cognitive plausibility of larger LMs has been underestimated. Furthermore, we first identify an intriguing relationship between LM layers and human measures: earlier layers correspond more closely with fast gaze durations, while later layers better align with relatively slower signals such as N400 potentials and MAZE processing times. Our work opens new avenues for interdisciplinary research at the intersection of mechanistic interpretability and cognitive modeling.

cs.CL↗

arXiv2025

On Entity Identification in Language Models

We analyze the extent to which internal representations of language models (LMs) identify and distinguish mentions of named entities, focusing on the many-to-many correspondence between entities and their mentions. We first formulate two problems of entity mentions -- ambiguity and variability -- and propose a framework analogous to clustering quality metrics. Specifically, we quantify through cluster analysis of LM internal representations the extent to which mentions of the same entity cluster together and mentions of different entities remain separated. Our experiments examine five Transformer-based autoregressive models, showing that they effectively identify and distinguish entities with metrics analogous to precision and recall ranging from 0.66 to 0.9. Further analysis reveals that entity-related information is compactly represented in a low-dimensional linear subspace at early LM layers. Additionally, we clarify how the characteristics of entity representations influence word prediction performance. These findings are interpreted through the lens of isomorphism between LM representations and entity-centric knowledge structures in the real world, providing insights into how LMs internally organize and use entity information.

cs.CL↗

arXiv2025

TopK Language Models

Sparse autoencoders (SAEs) have become an important tool for analyzing and interpreting the activation space of transformer-based language models (LMs). However, SAEs suffer several shortcomings that diminish their utility and internal validity. Since SAEs are trained post-hoc, it is unclear if the failure to discover a particular concept is a failure on the SAE's side or due to the underlying LM not representing this concept. This problem is exacerbated by training conditions and architecture choices affecting which features an SAE learns. When tracing how LMs learn concepts during training, the lack of feature stability also makes it difficult to compare SAEs features across different checkpoints. To address these limitations, we introduce a modification to the transformer architecture that incorporates a TopK activation function at chosen layers, making the model's hidden states equivalent to the latent features of a TopK SAE. This approach eliminates the need for post-hoc training while providing interpretability comparable to SAEs. The resulting TopK LMs offer a favorable trade-off between model size, computational efficiency, and interpretability. Despite this simple architectural change, TopK LMs maintain their original capabilities while providing robust interpretability benefits. Our experiments demonstrate that the sparse representations learned by TopK LMs enable successful steering through targeted neuron interventions and facilitate detailed analysis of neuron formation processes across checkpoints and layers. These features make TopK LMs stable and reliable tools for understanding how language models learn and represent concepts, which we believe will significantly advance future research on model interpretability and controllability.

cs.CL↗

arXiv2025

Emergence of Primacy and Recency Effect in Mamba: A Mechanistic Point of View

We study memory in state-space language models using primacy and recency effects as behavioral tools to uncover how information is retained and forgotten over time. Applying structured recall tasks to the Mamba architecture, we observe a consistent U-shaped accuracy profile, indicating strong performance at the beginning and end of input sequences. We identify three mechanisms that give rise to this pattern. First, long-term memory is supported by a sparse subset of channels within the model's selective state space block, which persistently encode early input tokens and are causally linked to primacy effects. Second, short-term memory is governed by delta-modulated recurrence: recent inputs receive more weight due to exponential decay, but this recency advantage collapses when distractor items are introduced, revealing a clear limit to memory depth. Third, we find that memory allocation is dynamically modulated by semantic regularity: repeated relations in the input sequence shift the delta gating behavior, increasing the tendency to forget intermediate items. We validate these findings via targeted ablations and input perturbations on two large-scale Mamba-based language models: one with 1.4B and another with 7B parameters.

cs.CL↗

arXiv2025

Rectifying Belief Space via Unlearning to Harness LLMs' Reasoning

Large language models (LLMs) can exhibit advanced reasoning yet still generate incorrect answers. We hypothesize that such errors frequently stem from spurious beliefs, propositions the model internally considers true but are incorrect. To address this, we propose a method to rectify the belief space by suppressing these spurious beliefs while simultaneously enhancing true ones, thereby enabling more reliable inferences. Our approach first identifies the beliefs that lead to incorrect or correct answers by prompting the model to generate textual explanations, using our Forward-Backward Beam Search (FBBS). We then apply unlearning to suppress the identified spurious beliefs and enhance the true ones, effectively rectifying the model's belief space. Empirical results on multiple QA datasets and LLMs show that our method corrects previously misanswered questions without harming overall model performance. Furthermore, our approach yields improved generalization on unseen data, suggesting that rectifying a model's belief space is a promising direction for mitigating errors and enhancing overall reliability.

cs.CL↗

arXiv2025

Spelling-out is not Straightforward: LLMs' Capability of Tokenization from Token to Characters

Large language models (LLMs) can spell out tokens character by character with high accuracy, yet they struggle with more complex character-level tasks, such as identifying compositional subcomponents within tokens. In this work, we investigate how LLMs internally represent and utilize character-level information during the spelling-out process. Our analysis reveals that, although spelling out is a simple task for humans, it is not handled in a straightforward manner by LLMs. Specifically, we show that the embedding layer does not fully encode character-level information, particularly beyond the first character. As a result, LLMs rely on intermediate and higher Transformer layers to reconstruct character-level knowledge, where we observe a distinct "breakthrough" in their spelling behavior. We validate this mechanism through three complementary analyses: probing classifiers, identification of knowledge neurons, and inspection of attention weights.

cs.CL↗

arXiv2025

Do LLMs Need to Think in One Language? Correlation between Latent Language and Task Performance

Large Language Models (LLMs) are known to process information using a proficient internal language consistently, referred to as latent language, which may differ from the input or output languages. However, how the discrepancy between the latent language and the input and output language affects downstream task performance remains largely unexplored. While many studies research the latent language of LLMs, few address its importance in influencing task performance. In our study, we hypothesize that thinking in latent language consistently enhances downstream task performance. To validate this, our work varies the input prompt languages across multiple downstream tasks and analyzes the correlation between consistency in latent language and task performance. We create datasets consisting of questions from diverse domains such as translation and geo-culture, which are influenced by the choice of latent language. Experimental results across multiple LLMs on translation and geo-culture tasks, which are sensitive to the choice of language, indicate that maintaining consistency in latent language is not always necessary for optimal downstream task performance. This is because these models adapt their internal representations near the final layers to match the target language, reducing the impact of consistency on overall performance.

cs.CL↗

arXiv2025

Mechanistic Insights into Grokking from the Embedding Layer

Grokking, a delayed generalization in neural networks after perfect training performance, has been observed in Transformers and MLPs, but the components driving it remain underexplored. We show that embeddings are central to grokking: introducing them into MLPs induces delayed generalization in modular arithmetic tasks, whereas MLPs without embeddings can generalize immediately. Our analysis identifies two key mechanisms: (1) Embedding update dynamics, where rare tokens stagnate due to sparse gradient updates and weight decay, and (2) Bilinear coupling, where the interaction between embeddings and downstream weights introduces saddle points and increases sensitivity to initialization. To confirm these mechanisms, we investigate frequency-aware sampling, which balances token updates by minimizing gradient variance, and embedding-specific learning rates, derived from the asymmetric curvature of the bilinear loss landscape. We prove that an adaptive learning rate ratio, \(\frac{η_E}{η_W} \propto \frac{σ_{\max}(E)}{σ_{\max}(W)} \cdot \frac{f_W}{f_E}\), mitigates bilinear coupling effects, accelerating convergence. Our methods not only improve grokking dynamics but also extend to broader challenges in Transformer optimization, where bilinear interactions hinder efficient training.

cs.LG↗