arXiv subjects

Naoaki Okazaki

Publications and source records attributed to Naoaki Okazaki.

At least 19 recordsLinked to original sources

arXiv2026

DyMT-ESB: Dynamic Multi-Turn Evaluation of Social Bias in User-LLM Interactions

Warning: This paper contains examples of stereotypes and social bias. LLMs are increasingly used in interactive settings by the general public, making the evaluation of model behavior in multi-turn conversational scenarios important for safety, including stereotyping-related harms. However, existing multi-turn social bias evaluations often rely on pre-specified or template-based user inputs that do not adapt to model responses and typically assume a fixed dialogue length in advance. In this paper, we study social bias dynamics in response-conditioned multi-turn interactions using a controlled evaluation protocol that generates follow-up user queries from the evolving dialogue history and allows evaluation over variable numbers of turns. Experimental results show that LLMs exhibit social bias even in coherent, response-conditioned multi-turn interactions, revealing late-emerging bias, non-monotonic bias patterns, and bias re-emergence. These results motivate evaluations that extend beyond fixed-turn, pre-scripted protocols. Our findings highlight the importance of analyzing social bias as a turn-level dynamic phenomenon.

cs.CL↗

arXiv2026

Synthesizing Instruction-Tuning Datasets with Contrastive Decoding

Using responses generated by high-performing large language models (LLMs) for instruction tuning has become a widely adopted approach. However, the existing literature overlooks a property of LLM-generated responses: they conflate world knowledge acquired during pre-training with instruction-following capabilities acquired during post-training. We hypothesize that disentangling the instruction-following capabilities from pre-trained knowledge improves the effectiveness of instruction tuning. To this end, we propose CoDIT, a method that applies contrastive decoding between a post-trained model and its pre-trained counterpart during response generation. The method suppresses pre-trained knowledge shared between the two models while amplifying the instruction-following behavior acquired via post-training, resulting in responses that more purely reflect instruction-following capabilities. Experiment results demonstrate that models trained on datasets constructed via CoDIT consistently outperform those trained on directly generated responses. Training on our datasets also yields better performance than on existing publicly available instruction-tuning datasets across multiple benchmarks. Furthermore, we theoretically and empirically show that CoDIT can be interpreted as distilling the chat vector from parameter space to text space, enabling the transfer of instruction-tuning capabilities across models of different architectures.

cs.CL↗

arXiv2026

Jagle: Building a Large-Scale Japanese Multimodal Post-Training Dataset for Vision-Language Models

Developing vision-language models (VLMs) that generalize across diverse tasks requires large-scale training datasets with diverse content. In English, such datasets are typically constructed by aggregating and curating numerous existing visual question answering (VQA) resources. However, this strategy does not readily extend to other languages, where VQA datasets remain limited in both scale and domain coverage. In this work, we introduce Jagle, the largest Japanese multimodal post-training dataset to date, comprising approximately 9.2 million instances across diverse tasks. Rather than relying on existing VQA datasets, we collect heterogeneous source data, including images, image-text pairs, and PDF documents, and generate VQA pairs through multiple strategies such as VLM-based QA generation, translation, and text rendering. Experiments demonstrate that a 2.2B model trained with Jagle surpasses InternVL3.5-2B in average score across ten Japanese evaluation tasks and comes within five points of Qwen3-VL-2B-Instruct. Moreover, combining Jagle with FineVision improves English performance over training with FineVision alone. These benefits generalize across model backbones and scales: LLM-jp-4-VL 9B beta, a 9B-scale model built on the Japanese-centric LLM-jp-4-8B-instruct, achieves Japanese performance comparable to Qwen3-VL-8B-Instruct. To facilitate reproducibility and future research, we release the dataset, models, and code.

cs.CV↗

arXiv2026

LLM Output Detectability and Task Performance Can be Jointly Optimized

Detecting machine-generated text is essential for transparency and accountability when deploying LLMs. Watermarking enables statistically reliable detection by biasing token distributions to embed detectable signals into LLM outputs. However, it has been reported that watermarked LLMs often perform worse on downstream tasks. We propose PUPPET, a framework that fine-tunes an LLM via DPO to generate text that is both more detectable by a target detector and better performing on downstream tasks. We use two rewards: a detector that outputs a machine-class likelihood and an evaluator that measures a task-specific metric. Just as a watermark is verified with its secret key, this detector-specific design lets an LLM provider track how its published model is used. Experiments on long-form QA, summarization, and essay writing show that LLMs trained with PUPPET achieve detectability competitive with watermarking methods---even at strict low FPRs---while outperforming them on downstream tasks. Moreover, this optimization requires only a few thousand samples and 1--2 GPU hours, and its gains hold across out-of-domain tasks, six detectors of diverse architectures, and different LLM families and sizes, and are even robust to paraphrasing attacks.

cs.CL↗

arXiv2026

Beyond Perplexity: UTF-8 Validity in Byte-aware Language Models

Byte-level tokenization enables language models to handle any Unicode input, but models can generate invalid UTF-8 sequences when encountering rare or unseen characters. We investigate the relationship between training scale and UTF-8 generation reliability with a 355M parameter model trained on 80B tokens from a balanced multilingual corpus of English, Japanese, Korean, and Chinese. We introduce multiple evaluation protocols that isolate UTF-8 structural validity from language modeling. UTF-8 validity convergence lags perplexity by a roughly a factor of two: perplexity stabilizes after 2.1B tokens, but UTF-8 validity requires 4.2B tokens. In context-free generation, rare characters achieve higher structural validity than common characters, suggesting over-specialization of frequent character representations. Through experiments, we observed that reliable UTF-8 generation is a distinct capability requiring evaluation beyond perplexity.

cs.CL↗

arXiv2026

Neuron Level Analysis of Large Language Model in Legal Domain Reasoning

We presented a neuron-level analysis of legal-domain reasoning in LLMs, comparing it with other applied domain tasks across seven open-weight models. Using neuron attribution scores to rank and suppress influential neurons, we confirmed that suppressing the identified neurons collapses accuracy on the target task, whereas suppressing the same number of random neurons does not. We further found a small subset of neurons influential across all seven tasks; once these are removed, suppressing the remaining neurons degrades only the task they were identified from, revealing genuinely task-specific neurons in every model studied. Within the legal domain, the three benchmarks exhibit relatively high neuron overlap and tend to be affected jointly, suggesting of legal components neurons that span jurisdictions. The distribution of identified neurons in our experiments suggests that the hypothesis that influential neurons are concentrated in middle MLP layers may depend on the input format and content, rather than being a universal phenomenon.

cs.CL↗

arXiv2026

From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models

While multimodal large language models (MLLMs) have made substantial progress in single-image spatial reasoning, multi-image spatial reasoning, which requires integration of information from multiple viewpoints, remains challenging. Cognitive studies suggest that humans address such tasks through two mechanisms: cross-view correspondence, which identifies regions across different views that correspond to the same physical locations, and stepwise viewpoint transformation, which composes relative viewpoint changes sequentially. However, existing studies incorporate these mechanisms only partially and often implicitly, without explicit supervision for both. We propose Human-Aware Training for Cross-view correspondence and viewpoint cHange (HATCH), a training framework with two complementary objectives: (1) Patch-Level Spatial Alignment, which encourages patch representations to align across views for spatially corresponding regions, and (2) Action-then-Answer Reasoning, which requires the model to generate explicit viewpoint transition actions before predicting the final answer. Experiments on three benchmarks demonstrate that HATCH consistently outperforms baselines of comparable size by a clear margin and achieves competitive results against much larger models, while preserving single-image reasoning capabilities.

cs.CV↗

arXiv2026

Autoregressive Direct Preference Optimization

Direct preference optimization (DPO) has emerged as a promising approach for aligning large language models (LLMs) with human preferences. However, the widespread reliance on the response-level Bradley-Terry (BT) model may limit its full potential, as the reference and learnable models are assumed to be autoregressive only after deriving the objective function. Motivated by this limitation, we revisit the theoretical foundations of DPO and propose a novel formulation that explicitly introduces the autoregressive assumption prior to applying the BT model. By reformulating and extending DPO, we derive a novel variant, termed Autoregressive DPO (ADPO), that explicitly integrates autoregressive modeling into the preference optimization framework. Without violating the theoretical foundations, the derived loss takes an elegant form: it shifts the summation operation in the DPO objective outside the log-sigmoid function. Furthermore, through theoretical analysis of ADPO, we show that there exist two length measures to be considered when designing DPO-based algorithms: the token length $μ$ and the feedback length $μ'$. To the best of our knowledge, we are the first to explicitly distinguish these two measures and analyze their implications for preference optimization in LLMs.

cs.AI↗

arXiv2026

Aligning Tree-Search Policies with Fixed Token Budgets in Test-Time Scaling of LLMs

Tree-search decoding is an effective form of test-time scaling for large language models (LLMs), but real-world deployment often imposes a fixed per-query token budget that varies across settings. Existing tree-search policies are largely budget-agnostic, treating the budget merely as a termination condition, thereby risking late-stage over-branching or premature termination. We propose Budget-Guided MCTS (BG-MCTS), a tree-search decoding algorithm that aligns its search policy with the remaining token budget: it starts with broad exploration, then prioritizes refinement and answer completion as the remaining budget decreases while reducing late-stage branching from shallow nodes. BG-MCTS consistently outperforms budget-agnostic tree-search baselines across inference budgets on mathematical reasoning benchmarks and an additional physics reasoning benchmark with open-weight LLMs.

cs.CL↗

arXiv2026

WAON: A Large-Scale Japanese Image-Text Dataset for Cultural Adaptation in Contrastive Vision-Language Models

Contrastive vision-language models have achieved remarkable progress through large-scale pretraining. Recent work has shown that removing English-only caption filters and pretraining on global data is effective for improving multicultural performance. We study whether such global pretraining is sufficient for culture-specific understanding, or whether further adaptation with natively sourced data can boost performance beyond what global pretraining alone achieves. To enable this investigation, we present WAON, the largest publicly available native Japanese image-text dataset constructed from native Japanese web content in Common Crawl, containing approximately 155 million examples. We also introduce WAON-Bench, a manually curated Japanese cultural benchmark spanning 374 classes. Through comparative fine-tuning experiments on multiple Japanese image-text datasets, we observe that models fine-tuned on WAON consistently achieve stronger performance on Japanese cultural benchmarks than those fine-tuned on English-to-Japanese translated data. We release our dataset and code.

cs.CV↗

arXiv2026

HakushoBench: A Japanese Chart and Table VQA Benchmark from Governmental White Papers

Understanding chart and table images is essential for applying vision-language models (VLMs) to real-world document understanding. While English benchmarks have advanced rapidly, non-English counterparts remain scarce, leaving it unclear whether this progress generalizes across languages. A key obstacle is the difficulty of collecting realistic and diverse non-English chart and table images at scale. To address this, we leverage governmental white papers as a scalable source for benchmark construction beyond English, as they contain naturally occurring charts and tables across diverse formats and domains and are freely accessible in many countries. As a first instantiation, we introduce HakushoBench, a challenging Japanese chart and table VQA benchmark built from 33 governmental white papers. HakushoBench contains 2,053 images spanning over 10 image types, with manually annotated QA pairs, designed to assess deep and holistic understanding of charts and tables, rather than local visual cues alone. Experiments across a broad range of VLMs demonstrate that HakushoBench remains challenging for open-weight models: the best open-weight model achieves only 58.6% accuracy, and a 34.9-point gap between open-weight and proprietary models highlights substantial room for improvement in complex chart and table understanding. We release our dataset and code.

cs.CV↗

arXiv2026

Diffusion-State Policy Optimization for Masked Diffusion Language Models

Masked diffusion language models generate text through iterative masked-token filling, but terminal-only rewards on final completions provide coarse credit assignment for the intermediate filling decisions that shape the generation process. We propose Diffusion-State Policy Optimization (DiSPO), a plug-in credit-assignment layer that directly optimizes intermediate filling decisions. At selected intermediate masked states, DiSPO branches by resampling the currently masked positions from rollout-cached logits, scores the resulting completions, and updates only the newly filled tokens, requiring no additional multi-step diffusion rollouts or optimizer steps. We formalize a fixed-state objective for branched completions and derive a policy-gradient estimator that reuses the same rollouts as terminal-feedback policy optimization. Experiments on LLaDA-8B-Instruct show that DiSPO consistently improves terminal-feedback baselines, including diffu-GRPO and SPG, on math and planning benchmarks under matched rollout compute and optimizer steps, supporting its use as a general plug-in for masked diffusion policy optimization. Our project page is available at https://daioba.github.io/dispo .

cs.CL↗

arXiv2026

Drifting Objectives for Refining Discrete Diffusion Language Models

Discrete diffusion language models (DDLMs) generate text by iteratively denoising categorical token sequences, while recent drifting methods for continuous generators suggest that part of this sampling-time correction can instead be absorbed into training through an anti-symmetric fixed-point objective. We study how to transfer this principle to DDLMs, where the main challenge is the interface with discrete text: hard token samples are non-differentiable, and categorical predictions do not directly provide continuous samples to drift. We formulate TokenDrift, a drifting objective that lifts categorical predictions to soft-token features, applies anti-symmetric drifting in a frozen semantic space, and backpropagates the resulting stop-gradient feature target to DDLM logits. In controlled continual-training experiments with masked and uniform-state diffusion backbones, TokenDrift improves fixed-NFE generation quality over matched continuation baselines, reducing Gen.-PPL at 4 NFEs by 89% on MDLM and 86% on DUO. These results suggest that drifting can provide a practical refinement objective for DDLMs.

cs.CL↗

arXiv2026

Stopping Computation for Converged Tokens in Masked Diffusion-LM Decoding

Masked Diffusion Language Models generate sequences via iterative sampling that progressively unmasks tokens. However, they still recompute the attention and feed-forward blocks for every token position at every step -- even when many unmasked tokens are essentially fixed, resulting in substantial waste in compute. We propose SureLock: when the posterior at an unmasked position has stabilized across steps (our sure condition), we lock that position -- thereafter skipping its query projection and feed-forward sublayers -- while caching its attention keys and values so other positions can continue to attend to it. This reduces the dominant per-iteration computational cost from $O(N^2d)$ to $O(MNd)$ where $N$ is the sequence length, $M$ is the number of unlocked token positions, and $d$ is the model dimension. In practice, $M$ decreases as the iteration progresses, yielding substantial savings. On LLaDA-8B, SureLock reduces algorithmic FLOPs by 30--50% relative to the same sampler without locking, while maintaining comparable generation quality. We also provide a theoretical analysis to justify the design rationale of SureLock: monitoring only the local KL at the lock step suffices to bound the deviation in final token probabilities. Our project page is available at https://daioba.github.io/surelock .

cs.CL↗

arXiv2026

ExaGPT: Example-Based Machine-Generated Text Detection for Human Interpretability

Detecting texts generated by Large Language Models (LLMs) could cause grave mistakes due to incorrect decisions, such as undermining students' academic dignity. LLM text detection thus needs to ensure the interpretability of the decision, which can help users judge how reliably correct its prediction is. When humans verify whether a text is human-written or LLM-generated, they intuitively investigate which of them it shares more similar spans with. However, existing interpretable detectors are not aligned with the human decision-making process and fail to offer evidence that users easily understand. To bridge this gap, we introduce ExaGPT, an interpretable detection approach grounded in the human decision-making process for verifying the origin of a text. ExaGPT identifies a text by checking whether it shares more similar spans with human-written vs. with LLM-generated texts from a datastore. This approach can provide similar span examples that contribute to the decision for each span in the text as evidence. Our human evaluation demonstrates that providing similar span examples contributes more effectively to judging the correctness of the decision than existing interpretable methods. Moreover, extensive experiments in four domains and three generators show that ExaGPT massively outperforms prior interpretable detectors by up to +37.0 points of accuracy at a false positive rate of 1%.

cs.CL↗

arXiv2026

From Interpretability to Performance: Optimizing Retrieval Heads for Long-Context Language Models

Advances in mechanistic interpretability have identified special attention heads, known as retrieval heads, that are responsible for retrieving information from the context. However, the role of these retrieval heads in improving model performance remains unexplored. This work investigates whether retrieval heads can be leveraged to enhance the long-context capabilities of LLMs. Specifically, we propose RetMask, a method that generates training signals by contrasting normal model outputs with those from an ablated variant in which the retrieval heads are masked. This mechanism-based approach achieves substantial improvements: +2.28 points on HELMET at 128K for Llama-3.1, with +70% gains on generation with citation and +32% on passage re-ranking, while preserving performance on general tasks. Experiments across four models in three families demonstrate that RetMask consistently improves long-context performance, where gains correlate with the sparsity of the retrieval score distribution: models with sparser distributions, where retrieval capabilities are concentrated in a small set of heads, respond more strongly, while those with less sparse distributions show more modest gains. These results validate the functional role of retrieval heads and show that mechanistic insights can be transformed into performance enhancements.

cs.CL↗

arXiv2026

JAMMEval: A Refined Collection of Japanese Benchmarks for Reliable VLM Evaluation

Reliable evaluation is essential for the development of vision-language models (VLMs). However, Japanese VQA benchmarks have undergone far less iterative refinement than their English counterparts. As a result, many existing benchmarks contain issues such as ambiguous questions, incorrect answers, and instances that can be solved without visual grounding, undermining evaluation reliability and leading to misleading conclusions in model comparisons. To address these limitations, we introduce JAMMEval, a refined collection of Japanese benchmarks for reliable VLM evaluation. It is constructed by systematically refining seven existing Japanese benchmark datasets through two rounds of human annotation, improving both data quality and evaluation reliability. In our experiments, we evaluate open-weight and proprietary VLMs on JAMMEval and analyze the capabilities of recent models on Japanese VQA. We further demonstrate the effectiveness of our refinement by showing that the resulting benchmarks yield evaluation scores that better reflect model capability, exhibit lower run-to-run variance, and improve the ability to distinguish between models of different capability levels. We release our dataset and code to advance reliable evaluation of VLMs.

cs.CV↗

arXiv2026

A Japanese Benchmark for Evaluating Social Bias in Reasoning Based on Attribution Theory

In enhancing the fairness of Large Language Models (LLMs), evaluating social biases rooted in the cultural contexts of specific linguistic regions is essential. However, most existing Japanese benchmarks heavily rely on translating English data, which does not necessarily provide an evaluation suitable for Japanese culture. Furthermore, they only evaluate bias in the conclusion, failing to capture biases lurking in the reasoning. In this study, based on attribution theory in social psychology, we constructed a new dataset, ``JUBAKU-v2,'' which evaluates the bias in attributing behaviors to in-groups and out-groups within reasoning while fixing the conclusion. This dataset consists of 216 examples reflecting cultural biases specific to Japan. Experimental results verified that it can detect performance differences across models more sensitively than existing benchmarks.

cs.CL↗