Takeshi Kojima

Paavo Parmas, Yongmin Kim, Kohsei Matsutani et al.

Policy-gradient methods usually optimize expected return, but many real world applications care about distributional properties of returns: tail risk, outlier robustness, or best-of-K discovery. We introduce OrderGrad, a family of likelihood-ratio and reparameterization gradient estimators for order-statistic objectives. OrderGrad optimizes finite-sample L-statistics, i.e., weighted averages of sorted rewards or costs, recovering objectives such as VaR, CVaR, trimmed means, medians, and top-m/best-of-K criteria by changing only the rank weights. For any fixed sample size and rank-weight vector, OrderGrad provides an unbiased gradient estimator for the corresponding order-statistic objective. The method is implemented as a simple reward transformation that can then be used in an otherwise standard policy-gradient or reparameterized update. We study the resulting estimator's variance behavior and evaluate it on tasks where mean optimization is mismatched to the deployment objective, including LLM math post-training and other tasks. OrderGrad provides a unified, plug-and-play route to risk-averse, robust, and exploratory learning. Code: https://github.com/paavo5/ordergrad

Takeshi Kojima, Shixiang Shane Gu, Machel Reid et al. · deepmind

Pretrained large language models (LLMs) are widely used in many sub-fields of natural language processing (NLP) and generally known as excellent few-shot learners with task-specific exemplars. Notably, chain of thought (CoT) prompting, a recent technique for eliciting complex multi-step reasoning through step-by-step answer examples, achieved the state-of-the-art performances in arithmetics and symbolic reasoning, difficult system-2 tasks that do not follow the standard scaling laws for LLMs. While these successes are often attributed to LLMs' ability for few-shot learning, we show that LLMs are decent zero-shot reasoners by simply adding "Let's think step by step" before each answer. Experimental results demonstrate that our Zero-shot-CoT, using the same single prompt template, significantly outperforms zero-shot LLM performances on diverse benchmark reasoning tasks including arithmetics (MultiArith, GSM8K, AQUA-RAT, SVAMP), symbolic reasoning (Last Letter, Coin Flip), and other logical reasoning tasks (Date Understanding, Tracking Shuffled Objects), without any hand-crafted few-shot examples, e.g. increasing the accuracy on MultiArith from 17.7% to 78.7% and GSM8K from 10.4% to 40.7% with large InstructGPT model (text-davinci-002), as well as similar magnitudes of improvements with another off-the-shelf large model, 540B parameter PaLM. The versatility of this single prompt across very diverse reasoning tasks hints at untapped and understudied fundamental zero-shot capabilities of LLMs, suggesting high-level, multi-task broad cognitive capabilities may be extracted by simple prompting. We hope our work not only serves as the minimal strongest zero-shot baseline for the challenging reasoning benchmarks, but also highlights the importance of carefully exploring and analyzing the enormous zero-shot knowledge hidden inside LLMs before crafting finetuning datasets or few-shot exemplars.

Qi Cao, Takeshi Kojima, Andrew Gambardella et al.

Large language models (LLMs) demonstrate remarkable performance across diverse tasks, but they often generate responses that appear plausible while being factually incorrect. This problem is compounded by the lack of explicit uncertainty estimates, which makes it difficult for users to judge the reliability of model outputs. Existing uncertainty quantification methods typically rely on indirect signals, such as entropy across sampled generations. These signals can be difficult to interpret and do not fully leverage the model's ability to assess its own uncertainty. We propose a simple yet effective self-assessment method for uncertainty quantification in LLMs. Our approach groups sampled generations into semantically distinct clusters, converts them into answer options in a structured multiple-choice question, and uses the probability assigned by the LLM to each option as a confidence estimate. Experiments across multiple models and datasets show that our method consistently outperforms baseline approaches. Notably, it achieves competitive performance with as few as two additional samples, demonstrating both its effectiveness and efficiency.

Kohsei Matsutani, Gouki Minegishi, Takeshi Kojima et al.

Large language models (LLMs) can now solve complex problems through long chain-of-thought (CoT) reasoning, but the trade-off between performance and token cost remains a central challenge. To address this issue, supervised fine-tuning (SFT) often uses compressed reasoning data, where CoT traces are shortened into compact forms. However, the effect of such compressed reasoning data on post-training remains poorly understood. In this paper, we propose a taxonomy of CoT consisting of Explicit CoT, which outputs all operations without aggregation, Composed CoT, which combines multiple operations into a single step, and Implicit CoT, which omits intermediate operations. We construct a synthetic compositional reasoning task that allows controlled variation of difficulty, compression granularity, and data size, and conducted a comprehensive set of experiments across different model families and sizes. Notably, we find that (i) coarser CoT requires more SFT data, (ii) compared with Explicit CoT, Composed CoT and Implicit CoT benefit more from data scaling, while Composed CoT benefits from data repetition and Implicit CoT tends to lead to memorization, (iii) unlike SFT, subsequent reinforcement learning (RL) with verifiable rewards (RLVR) decomposes compressed steps learned during SFT, and (iv) unidirectional CoT ordering shows stronger generalization on longer sequential tasks. Our findings provide implications for CoT design under data resource constraints and offer important insights into the mechanisms of SFT and RL in LLM post-training.

Shota Takashiro, Soichiro Nishimori, Paavo Parmas et al.

Reinforcement learning with verifiable rewards is widely used for post-training reasoning models, but sparse outcome rewards make exploration difficult. A complementary approach is to optimize inference-time objectives such as pass@K and max@K directly, yet existing policy-gradient estimators for these objectives use different signals, baselines, and normalizations, making their relationships unclear. We study this issue through baseline design and advantage centering. Starting from the advantage estimator of a leading method in the field, we show that it is policy-gradient unbiased but yields a non-centered advantage. We then introduce a Leave-Two-Out baseline that preserves policy-gradient unbiasedness while making realized batch advantages exactly centered. The resulting method, MaxPO, has an efficient quadratic-time implementation and integrates naturally into group-based RL for LLM post-training. We further derive the canonical finite-batch advantage for max@K, providing a unified view of existing advantage estimators. Empirically, we verify that the L2O baseline reduces gradient variance and outperforms non-centered alternatives.

Takeshi Kojima, Yutaka Matsuo, Yusuke Iwasawa

Vision Transformer (ViT) is becoming more popular in image processing. Specifically, we investigate the effectiveness of test-time adaptation (TTA) on ViT, a technique that has emerged to correct its prediction during test-time by itself. First, we benchmark various test-time adaptation approaches on ViT-B16 and ViT-L16. It is shown that the TTA is effective on ViT and the prior-convention (sensibly selecting modulation parameters) is not necessary when using proper loss function. Based on the observation, we propose a new test-time adaptation method called class-conditional feature alignment (CFA), which minimizes both the class-conditional distribution differences and the whole distribution differences of the hidden representation between the source and target in an online manner. Experiments of image classification tasks on common corruption (CIFAR-10-C, CIFAR-100-C, and ImageNet-C) and domain adaptation (digits datasets and ImageNet-Sketch) show that CFA stably outperforms the existing baselines on various datasets. We also verify that CFA is model agnostic by experimenting on ResNet, MLP-Mixer, and several ViT variants (ViT-AugReg, DeiT, and BeiT). Using BeiT backbone, CFA achieves 19.8% top-1 error rate on ImageNet-C, outperforming the existing test-time adaptation baseline 44.0%. This is a state-of-the-art result among TTA methods that do not need to alter training phase.

Qi Cao, Takeshi Kojima, Yutaka Matsuo et al.

While Large Language Models (LLMs) have achieved remarkable performance in many tasks, much about their inner workings remains unclear. In this study, we present novel experimental insights into the resilience of LLMs, particularly GPT-4, when subjected to extensive character-level permutations. To investigate this, we first propose the Scrambled Bench, a suite designed to measure the capacity of LLMs to handle scrambled input, in terms of both recovering scrambled sentences and answering questions given scrambled context. The experimental results indicate that most powerful LLMs demonstrate the capability akin to typoglycemia, a phenomenon where humans can understand the meaning of words even when the letters within those words are scrambled, as long as the first and last letters remain in place. More surprisingly, we found that only GPT-4 nearly flawlessly processes inputs with unnatural errors, even under the extreme condition, a task that poses significant challenges for other LLMs and often even for humans. Specifically, GPT-4 can almost perfectly reconstruct the original sentences from scrambled ones, decreasing the edit distance by 95%, even when all letters within each word are entirely scrambled. It is counter-intuitive that LLMs can exhibit such resilience despite severe disruption to input tokenization caused by scrambled text.

Qi Cao, Andrew Gambardella, Takeshi Kojima et al.

Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks. However, the truthfulness of their outputs is not guaranteed, and their tendency toward overconfidence further limits reliability. Uncertainty quantification offers a promising way to identify potentially unreliable outputs, but most existing methods rely on repeated sampling or auxiliary models, introducing substantial computational overhead. To address these limitations, we propose Semantic Token Clustering (STC), an efficient uncertainty quantification method that leverages the semantic information inherently encoded in LLMs. Specifically, we group tokens into semantically consistent clusters using embedding clustering and prefix matching, and quantify uncertainty based on the probability mass aggregated over the corresponding semantic cluster. Our approach requires only a single generation and does not depend on auxiliary models. Experimental results show that STC achieves performance comparable to state-of-the-art baselines while substantially reducing computational overhead.

Gouki Minegishi, Hiroki Furuta, Takeshi Kojima et al.

Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a geometric account based on the geometry of feature superposition. Because features are encoded in overlapping representations, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this effect and empirically test it in multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, GPT-OSS 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach, filtering training samples closest to toxic features, reduces misalignment by 34.5%, substantially outperforming random removal and achieving comparable or slightly lower misalignment than LLM-as-a-judge-based filtering. Our study links emergent misalignment to feature superposition, providing a basis for understanding and mitigating this phenomenon.

Gouki Minegishi, Jingyuan Feng, Hiroki Furuta et al.

Analogy is a central faculty of human intelligence, enabling abstract patterns discovered in one domain to be applied to another. Despite its central role in cognition, the mechanisms by which Transformers acquire and implement analogical reasoning remain poorly understood. In this work, inspired by the notion of functors in category theory, we formalize analogical reasoning as the inference of correspondences between entities across categories. Based on this formulation, we introduce synthetic tasks that evaluate the emergence of analogical reasoning under controlled settings. We find that the emergence of analogical reasoning is highly sensitive to data characteristics, optimization choices, and model scale. Through mechanistic analysis, we show that analogical reasoning in Transformers decomposes into two key components: (1) geometric alignment of relational structure in the embedding space, and (2) the application of a functor within the Transformer. These mechanisms enable models to transfer relational structure from one category to another, realizing analogy. Finally, we quantify these effects and find that the same trends are observed in pretrained LLMs. In doing so, we move analogy from an abstract cognitive notion to a concrete, mechanistically grounded phenomenon in modern neural networks.

Yusuke Watanabe, Yohei Kobashi, Takeshi Kojima et al.

Clinical decisions are often required under incomplete information. Clinical experts must identify whether available information is sufficient for judgment, as both premature conclusion and unnecessary abstention can compromise patient safety. To evaluate this capability of large language models (LLMs), we developed ClinDet-Bench, a benchmark based on clinical scoring systems that decomposes incomplete-information scenarios into determinable and undeterminable conditions. Identifying determinability requires considering all hypotheses about missing information, including unlikely ones, and verifying whether the conclusion holds across them. We find that recent LLMs fail to identify determinability under incomplete information, producing both premature judgments and excessive abstention, despite correctly explaining the underlying scoring knowledge and performing well under complete information. These findings suggest that existing benchmarks are insufficient to evaluate the safety of LLMs in clinical settings. ClinDet-Bench provides a framework for evaluating determinability recognition, leading to appropriate abstention, with potential applicability to medicine and other high-stakes domains, and is publicly available.

Ikuya Yamada, Ryokan Ri, Takeshi Kojima et al.

Dense retrievers often struggle with queries involving less-frequent entities due to their limited entity knowledge. We propose the Knowledgeable Passage Retriever (KPR), a BERT-based retriever enhanced with a context-entity attention layer and dynamically updatable entity embeddings. This design enables KPR to incorporate external entity knowledge without retraining. Experiments on three datasets demonstrate that KPR consistently improves retrieval accuracy, with particularly large gains on the EntityQuestions dataset. When built on the off-the-shelf bge-base retriever, KPR achieves state-of-the-art performance among similarly sized models on two datasets. Models and code are released at https://github.com/knowledgeable-embedding/knowledgeable-embedding.

Andrew Gambardella, Takeshi Kojima, Yusuke Iwasawa et al.

Typical methods for evaluating the performance of language models evaluate their ability to answer questions accurately. These evaluation metrics are acceptable for determining the extent to which language models can understand and reason about text in a general sense, but fail to capture nuanced capabilities, such as the ability of language models to recognize and obey rare grammar points, particularly in languages other than English. We measure the perplexity of language models when confronted with the "first person psych predicate restriction" grammar point in Japanese. Weblab is the only tested open source model in the 7-10B parameter range which consistently assigns higher perplexity to ungrammatical psych predicate sentences than grammatical ones. We give evidence that Weblab's uniformly bad tokenization is a possible root cause for its good performance, and show that Llama 3's perplexity on grammatical psych predicate sentences can be reduced by orders of magnitude (28x difference) by restricting test sentences to those with uniformly well-behaved tokenizations. We show in further experiments on machine translation tasks that language models will use alternative grammar patterns in order to produce grammatical sentences when tokenization issues prevent the most natural sentence from being output.

Takeshi Kojima, Itsuki Okimura, Yusuke Iwasawa et al.

Current decoder-based pre-trained language models (PLMs) successfully demonstrate multilingual capabilities. However, it is unclear how these models handle multilingualism. We analyze the neuron-level internal behavior of multilingual decoder-based PLMs, Specifically examining the existence of neurons that fire ``uniquely for each language'' within decoder-only multilingual PLMs. We analyze six languages: English, German, French, Spanish, Chinese, and Japanese, and show that language-specific neurons are unique, with a slight overlap (< 5%) between languages. These neurons are mainly distributed in the models' first and last few layers. This trend remains consistent across languages and models. Additionally, we tamper with less than 1% of the total neurons in each model during inference and demonstrate that tampering with a few language-specific neurons drastically changes the probability of target language occurrence in text generation.

Kohsei Matsutani, Shota Takashiro, Gouki Minegishi et al.

Large language models (LLMs) are typically trained by reinforcement learning (RL) with verifiable rewards (RLVR) and supervised fine-tuning (SFT) on reasoning traces to improve their reasoning abilities. However, how these methods shape reasoning capabilities remains largely elusive. Going beyond an accuracy-based investigation of how these two components sculpt the reasoning process, this paper introduces a novel analysis framework that quantifies reasoning paths and captures their qualitative changes under each training process (with models of 1.5B, 7B, and 14B parameters on mathematical domains). Specifically, we investigate the reasoning process at two levels of granularity: the trajectory-level, which examines complete reasoning outputs, and the step-level, which analyzes reasoning graphs whose nodes correspond to individual reasoning steps. Notably, clustering of unique reasoning trajectories shows complementary effects: RL compresses incorrect trajectories, whereas SFT expands correct ones. Step-level analysis reveals that RL steepens (about 2.5 times), while SFT flattens (reduced to about one-third), the decay rates of node visitation frequency, degree, and betweenness centrality distributions in the reasoning graph. This indicates that RL concentrates reasoning functionality into a small subset of steps, while SFT homogenizes it across many steps. Furthermore, by evaluating the reasoning graph topologies from multiple perspectives, we delineate the shared and distinct characteristics of RL and SFT. Our work presents a novel reasoning path perspective that explains why the current best practice of two-stage training, with SFT followed by RL, is successful, and offers practical implications for data construction and more efficient learning approaches.

Shota Takashiro, Takeshi Kojima, Shohei Taniguchi et al.

The Mixture of Experts (MoE) selects a few feed-forward networks (FFNs) per token, achieving an effective trade-off between computational cost and performance. In conventional MoE, each expert is treated as entirely independent, and experts are combined in a discrete space. As a result, when the number of experts increases, it becomes difficult to train each expert effectively. To stabilize training while increasing the number of experts, we propose $\infty$-MoE that selects a portion of the parameters of large FFNs based on continuous values sampled for each token. By considering experts in a continuous space, this approach allows for an infinite number of experts while maintaining computational efficiency. Experiments show that a GPT-2 Small-based $\infty$-MoE model, with 129M active and 186M total parameters, achieves comparable performance to a dense GPT-2 Medium with 350M parameters. Adjusting the number of sampled experts at inference time allows for a flexible trade-off between accuracy and speed, with an improvement of up to 2.5\% in accuracy over conventional MoE.

Keno Harada, Lui Yoshida, Takeshi Kojima et al.

The performance of Large Language Models (LLMs) is highly sensitive to the prompts they are given. Drawing inspiration from the field of prompt optimization, this study investigates the potential for enhancing Automated Essay Scoring (AES) by refining the scoring rubrics used by LLMs. Specifically, our approach prompts models to iteratively refine rubrics by reflecting on models' own scoring rationales and observed discrepancies with human scores on sample essays. Experiments on the TOEFL11 and ASAP datasets using GPT-4.1, Gemini-2.5-Pro, and Qwen-3-Next-80B-A3B-Instruct show Quadratic Weighted Kappa (QWK) improvements of up to 0.19 and 0.47, respectively. Notably, even with a simple initial rubric, our approach achieves comparable or better QWK than using detailed human-authored rubrics. Our findings highlight the importance of iterative rubric refinement in LLM-based AES to enhance alignment with human evaluations.

Keno Harada, Yudai Yamazaki, Masachika Taniguchi et al.

As large language models (LLMs) are increasingly applied to real-world scenarios, it becomes crucial to understand their ability to follow multiple instructions simultaneously. To systematically evaluate these capabilities, we introduce two specialized benchmarks for fundamental domains where multiple instructions following is important: Many Instruction-Following Eval (ManyIFEval) for text generation with up to ten instructions, and Style-aware Mostly Basic Programming Problems (StyleMBPP) for code generation with up to six instructions. Our experiments with the created benchmarks across ten LLMs reveal that performance consistently degrades as the number of instructions increases. Furthermore, given the fact that evaluating all the possible combinations of multiple instructions is computationally impractical in actual use cases, we developed three types of regression models that can estimate performance on both unseen instruction combinations and different numbers of instructions which are not used during training. We demonstrate that a logistic regression model using instruction count as an explanatory variable can predict performance of following multiple instructions with approximately 10% error, even for unseen instruction combinations. We show that relatively modest sample sizes (500 for ManyIFEval and 300 for StyleMBPP) are sufficient for performance estimation, enabling efficient evaluation of LLMs under various instruction combinations.

Takeshi Kojima, Yaonan Zhu, Yusuke Iwasawa et al.

Recent Foundation Model-enabled robotics (FMRs) display greatly improved general-purpose skills, enabling more adaptable automation than conventional robotics. Their ability to handle diverse tasks thus creates new opportunities to replace human labor. However, unlike general foundation models, FMRs interact with the physical world, where their actions directly affect the safety of humans and surrounding objects, requiring careful deployment and control. Based on this proposition, our survey comprehensively summarizes robot control approaches to mitigate physical risks by covering all the lifespan of FMRs ranging from pre-deployment to post-accident stage. Specifically, we broadly divide the timeline into the following three phases: (1) pre-deployment phase, (2) pre-incident phase, and (3) post-incident phase. Throughout this survey, we find that there is much room to study (i) pre-incident risk mitigation strategies, (ii) research that assumes physical interaction with humans, and (iii) essential issues of foundation models themselves. We hope that this survey will be a milestone in providing a high-resolution analysis of the physical risks of FMRs and their control, contributing to the realization of a good human-robot relationship.