Yusuke Iwasawa

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.

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

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.

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.

Jiaxian Guo, Bo Yang, Paul Yoo et al. · allen-ai, uw

Unlike perfect information games, where all elements are known to every player, imperfect information games emulate the real-world complexities of decision-making under uncertain or incomplete information. GPT-4, the recent breakthrough in large language models (LLMs) trained on massive passive data, is notable for its knowledge retrieval and reasoning abilities. This paper delves into the applicability of GPT-4's learned knowledge for imperfect information games. To achieve this, we introduce \textbf{Suspicion-Agent}, an innovative agent that leverages GPT-4's capabilities for performing in imperfect information games. With proper prompt engineering to achieve different functions, Suspicion-Agent based on GPT-4 demonstrates remarkable adaptability across a range of imperfect information card games. Importantly, GPT-4 displays a strong high-order theory of mind (ToM) capacity, meaning it can understand others and intentionally impact others' behavior. Leveraging this, we design a planning strategy that enables GPT-4 to competently play against different opponents, adapting its gameplay style as needed, while requiring only the game rules and descriptions of observations as input. In the experiments, we qualitatively showcase the capabilities of Suspicion-Agent across three different imperfect information games and then quantitatively evaluate it in Leduc Hold'em. The results show that Suspicion-Agent can potentially outperform traditional algorithms designed for imperfect information games, without any specialized training or examples. In order to encourage and foster deeper insights within the community, we make our game-related data publicly available.

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.

Xin Zhang, Jiaxian Guo, Paul Yoo et al. · uw

Text-to-image generative models have attracted rising attention for flexible image editing via user-specified descriptions. However, text descriptions alone are not enough to elaborate the details of subjects, often compromising the subjects' identity or requiring additional per-subject fine-tuning. We introduce a new framework called \textit{Paste, Inpaint and Harmonize via Denoising} (PhD), which leverages an exemplar image in addition to text descriptions to specify user intentions. In the pasting step, an off-the-shelf segmentation model is employed to identify a user-specified subject within an exemplar image which is subsequently inserted into a background image to serve as an initialization capturing both scene context and subject identity in one. To guarantee the visual coherence of the generated or edited image, we introduce an inpainting and harmonizing module to guide the pre-trained diffusion model to seamlessly blend the inserted subject into the scene naturally. As we keep the pre-trained diffusion model frozen, we preserve its strong image synthesis ability and text-driven ability, thus achieving high-quality results and flexible editing with diverse texts. In our experiments, we apply PhD to both subject-driven image editing tasks and explore text-driven scene generation given a reference subject. Both quantitative and qualitative comparisons with baseline methods demonstrate that our approach achieves state-of-the-art performance in both tasks. More qualitative results can be found at \url{https://sites.google.com/view/phd-demo-page}.

Gouki Minegishi, Yusuke Iwasawa, Yutaka Matsuo

Grokking is an intriguing phenomenon of delayed generalization, where neural networks initially memorize training data with perfect accuracy but exhibit poor generalization, subsequently transitioning to a generalizing solution with continued training. While factors such as weight norms and sparsity have been proposed to explain this delayed generalization, the influence of network structure remains underexplored. In this work, we link the grokking phenomenon to the lottery ticket hypothesis to investigate the impact of internal network structures. We demonstrate that utilizing lottery tickets obtained during the generalizing phase (termed grokked tickets) significantly reduces delayed generalization across various tasks, including multiple modular arithmetic operations, polynomial regression, sparse parity, and MNIST classification. Through controlled experiments, we show that the mitigation of delayed generalization is not due solely to reduced weight norms or increased sparsity, but rather to the discovery of good subnetworks. Furthermore, we find that grokked tickets exhibit periodic weight patterns, beneficial graph properties such as increased average path lengths and reduced clustering coefficients, and undergo rapid structural changes that coincide with improvements in generalization. Additionally, pruning techniques like the edge-popup algorithm can identify these effective structures without modifying the weights, thereby transforming memorizing networks into generalizing ones. These results underscore the novel insight that structural exploration plays a pivotal role in understanding grokking. The implementation code can be accessed via this link: https://github.com/gouki510/Grokking-Tickets.

So Kuroki, Jiaxian Guo, Tatsuya Matsushima et al. · uw

Due to the inherent uncertainty in their deformability during motion, previous methods in deformable object manipulation, such as rope and cloth, often required hundreds of real-world demonstrations to train a manipulation policy for each object, which hinders their applications in our ever-changing world. To address this issue, we introduce GenDOM, a framework that allows the manipulation policy to handle different deformable objects with only a single real-world demonstration. To achieve this, we augment the policy by conditioning it on deformable object parameters and training it with a diverse range of simulated deformable objects so that the policy can adjust actions based on different object parameters. At the time of inference, given a new object, GenDOM can estimate the deformable object parameters with only a single real-world demonstration by minimizing the disparity between the grid density of point clouds of real-world demonstrations and simulations in a differentiable physics simulator. Empirical validations on both simulated and real-world object manipulation setups clearly show that our method can manipulate different objects with a single demonstration and significantly outperforms the baseline in both environments (a 62% improvement for in-domain ropes and a 15% improvement for out-of-distribution ropes in simulation, as well as a 26% improvement for ropes and a 50% improvement for cloths in the real world), demonstrating the effectiveness of our approach in one-shot deformable object manipulation.

Xiaojie Gu, Sherry T. Tong, Aosong Feng et al.

Reasoning-focused large language models (LLMs) have advanced in many NLP tasks, yet their evaluation remains challenging: final answers alone do not expose the intermediate reasoning steps, making it difficult to determine whether a model truly reasons correctly and where failures occur, while existing multi-hop QA benchmarks lack step-level annotations for diagnosing reasoning failures. To address this gap, we propose Omanic, an open-domain multi-hop QA resource that provides decomposed sub-questions and intermediate answers as structural annotations for analyzing reasoning processes. It contains 10,296 machine-generated training examples (OmanicSynth) and 967 expert-reviewed human-annotated evaluation examples (OmanicBench). Systematic evaluations show that state-of-the-art LLMs achieve only 73.11% multiple-choice accuracy on OmanicBench, confirming its high difficulty. Stepwise analysis reveals that CoT's performance hinges on factual completeness, with its gains diminishing under knowledge gaps and errors amplifying in later hops. Additionally, supervised fine-tuning on OmanicSynth brings substantial transfer gains (7.41 average points) across six reasoning and math benchmarks, validating the dataset's quality and further supporting the effectiveness of OmanicSynth as supervision for reasoning-capability transfer. We release the data at https://huggingface.co/datasets/li-lab/Omanic and the code at https://github.com/XiaojieGu/Omanic.

Hiroki Furuta, Yusuke Iwasawa, Yutaka Matsuo et al.

The rise of generalist large-scale models in natural language and vision has made us expect that a massive data-driven approach could achieve broader generalization in other domains such as continuous control. In this work, we explore a method for learning a single policy that manipulates various forms of agents to solve various tasks by distilling a large amount of proficient behavioral data. In order to align input-output (IO) interface among multiple tasks and diverse agent morphologies while preserving essential 3D geometric relations, we introduce morphology-task graph, which treats observations, actions and goals/task in a unified graph representation. We also develop MxT-Bench for fast large-scale behavior generation, which supports procedural generation of diverse morphology-task combinations with a minimal blueprint and hardware-accelerated simulator. Through efficient representation and architecture selection on MxT-Bench, we find out that a morphology-task graph representation coupled with Transformer architecture improves the multi-task performances compared to other baselines including recent discrete tokenization, and provides better prior knowledge for zero-shot transfer or sample efficiency in downstream multi-task imitation learning. Our work suggests large diverse offline datasets, unified IO representation, and policy representation and architecture selection through supervised learning form a promising approach for studying and advancing morphology-task generalization.

Yue Xun, Junyu Liu, Qian Niu et al.

We introduce JMed48k, a multi-profession Japanese healthcare licensing benchmark for evaluating vision-language models. Built from official PDF materials released by the Japanese Ministry of Health, Labour and Welfare, JMed48k contains 48,862 exam questions and 20,142 images from 11 national licensing examinations between 2005 and 2025, with visual content annotated under an 8-type taxonomy. From this corpus, we derive JMed48k-Eval, a recent five-year evaluation subset with 12,484 scored questions, including 9,905 text-only questions and 2,579 questions with images. We evaluate 21 proprietary, open-source, and medical-specific models, reporting text-only and with-image performance separately. Because these subsets contain different questions, we further introduce a paired image-removal audit that evaluates questions with images before and after removing visual content to explore four answer-transition states. The audit shows that proprietary and open source models gain substantially from images, whereas medical-specific systems show limited observable use of visual evidence, with many correct answers persisting after image removal. Even among proprietary models, the net image-removal effect varies sevenfold across professions, from +5.7 points on Physician questions to +39.8 points on Public Health Nurse questions. We release JMed48k to support reproducible, profession-stratified evaluation of vision-language models in medical licensing settings.

So Kuroki, Jiaxian Guo, Tatsuya Matsushima et al. · uw

Due to the inherent uncertainty in their deformability during motion, previous methods in rope manipulation often require hundreds of real-world demonstrations to train a manipulation policy for each rope, even for simple tasks such as rope goal reaching, which hinder their applications in our ever-changing world. To address this issue, we introduce GenORM, a framework that allows the manipulation policy to handle different deformable ropes with a single real-world demonstration. To achieve this, we augment the policy by conditioning it on deformable rope parameters and training it with a diverse range of simulated deformable ropes so that the policy can adjust actions based on different rope parameters. At the time of inference, given a new rope, GenORM estimates the deformable rope parameters by minimizing the disparity between the grid density of point clouds of real-world demonstrations and simulations. With the help of a differentiable physics simulator, we require only a single real-world demonstration. Empirical validations on both simulated and real-world rope manipulation setups clearly show that our method can manipulate different ropes with a single demonstration and significantly outperforms the baseline in both environments (62% improvement in in-domain ropes, and 15% improvement in out-of-distribution ropes in simulation, 26% improvement in real-world), demonstrating the effectiveness of our approach in one-shot rope manipulation.

Tatsuya Matsushima, Yuki Noguchi, Jumpei Arima et al.

Tidying up a household environment using a mobile manipulator poses various challenges in robotics, such as adaptation to large real-world environmental variations, and safe and robust deployment in the presence of humans.The Partner Robot Challenge in World Robot Challenge (WRC) 2020, a global competition held in September 2021, benchmarked tidying tasks in the real home environments, and importantly, tested for full system performances.For this challenge, we developed an entire household service robot system, which leverages a data-driven approach to adapt to numerous edge cases that occur during the execution, instead of classical manual pre-programmed solutions. In this paper, we describe the core ingredients of the proposed robot system, including visual recognition, object manipulation, and motion planning. Our robot system won the second prize, verifying the effectiveness and potential of data-driven robot systems for mobile manipulation in home environments.

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.

Koya Sakamoto, Taiki Miyanishi, Daichi Azuma et al.

Visual search in 3D environments requires embodied agents to actively explore their surroundings and acquire task-relevant evidence. However, existing visual search and embodied AI benchmarks, including EQA, typically rely on static observations or constrained egocentric motion, and thus do not explicitly evaluate fine-grained viewpoint-dependent phenomena that arise under unrestricted 5-DoF viewpoint control in real-world 3D environments, such as visibility changes caused by vertical viewpoint shifts, revealing contents inside containers, and disambiguating object attributes that are only observable from specific angles. To address this limitation, we introduce {E3VS-Bench}, a benchmark for embodied 3D visual search where agents must control their viewpoints in 5-DoF to gather viewpoint-dependent evidence for question answering. E3VS-Bench consists of 99 high-fidelity 3D scenes reconstructed using 3D Gaussian Splatting and 2,014 question-driven episodes. 3D Gaussian Splatting enables photorealistic free-viewpoint rendering that preserves fine-grained visual details (e.g., small text and subtle attributes) often degraded in mesh-based simulators, thereby allowing the construction of questions that cannot be answered from a single view and instead require active inspection across viewpoints in 5-DoF. We evaluate multiple state-of-the-art VLMs and compare their performance with humans. Despite strong 2D reasoning ability, all models exhibit a substantial gap from humans, highlighting limitations in active perception and coherent viewpoint planning specifically under full 5-DoF viewpoint changes.

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.

Kenji Kubo, Shunsuke Kamiya, Masanori Koyama et al.

Neural network models with latent recurrent processing, where identical layers are recursively applied to the latent state, have gained attention as promising models for performing reasoning tasks. A strength of such models is that they enable test-time scaling, where the models can enhance their performance in the test phase without additional training. Models such as the Hierarchical Reasoning Model (HRM) and Artificial Kuramoto Oscillatory Neurons (AKOrN) can facilitate deeper reasoning by increasing the number of recurrent steps, thereby enabling the completion of challenging tasks, including Sudoku, Maze solving, and AGI benchmarks. In this work, we introduce confidence-based voting (C-voting), a test-time scaling strategy designed for recurrent models with multiple latent candidate trajectories. Initializing the latent state with multiple candidates using random variables, C-voting selects the one maximizing the average of top-1 probabilities of the predictions, reflecting the model's confidence. Additionally, it yields 4.9% higher accuracy on Sudoku-hard than the energy-based voting strategy, which is specific to models with explicit energy functions. An essential advantage of C-voting is its applicability: it can be applied to recurrent models without requiring an explicit energy function. Finally, we introduce a simple attention-based recurrent model with randomized initial values named ItrSA++, and demonstrate that when combined with C-voting, it outperforms HRM on Sudoku-extreme (95.2% vs. 55.0%) and Maze (78.6% vs. 74.5%) tasks.

Rui Yang, Huitao Li, Weihao Xuan et al.

Despite continuous advances in medical technology, the global distribution of health care resources remains uneven. The development of large language models (LLMs) has transformed the landscape of medicine and holds promise for improving health care quality and expanding access to medical information globally. However, existing LLMs are primarily trained on high-resource languages, limiting their applicability in global medical scenarios. To address this gap, we constructed GlobMed, a large multilingual medical dataset, containing over 500,000 entries spanning 12 languages, including four low-resource languages. Building on this, we established GlobMed-Bench, which systematically assesses 56 state-of-the-art proprietary and open-weight LLMs across multiple multilingual medical tasks, revealing significant performance disparities across languages, particularly for low-resource languages. Additionally, we introduced GlobMed-LLMs, a suite of multilingual medical LLMs trained on GlobMed, with parameters ranging from 1.7B to 8B. GlobMed-LLMs achieved an average performance improvement of over 40% relative to baseline models, with a more than threefold increase in performance on low-resource languages. Together, these resources provide an important foundation for advancing the equitable development and application of LLMs globally, enabling broader language communities to benefit from technological advances.

Xinrui Wang, Zhuoru Li, Xiao Zhou et al.

Portrait Stylization aims to imbue portrait photos with vivid artistic effects drawn from style examples. Despite the availability of enormous training datasets and large network weights, existing methods struggle to maintain geometric consistency and achieve satisfactory stylization effects due to the disparity in facial feature distributions between facial photographs and stylized images, limiting the application on rare styles and mobile devices. To alleviate this, we propose to establish meaningful geometric correlations between portraits and style samples to simplify the stylization by aligning corresponding facial characteristics. Specifically, we integrate differentiable Thin-Plate-Spline (TPS) modules into an end-to-end Generative Adversarial Network (GAN) framework to improve the training efficiency and promote the consistency of facial identities. By leveraging inherent structural information of faces, e.g., facial landmarks, TPS module can establish geometric alignments between the two domains, at global and local scales, both in pixel and feature spaces, thereby overcoming the aforementioned challenges. Quantitative and qualitative comparisons on a range of portrait stylization tasks demonstrate that our models not only outperforms existing models in terms of fidelity and stylistic consistency, but also achieves remarkable improvements in 2x training data efficiency and 100x less computational complexity, allowing our lightweight model to achieve real-time inference (30 FPS) at 512*512 resolution on mobile devices.

Shohei Taniguchi, Yusuke Iwasawa, Wataru Kumagai et al.

Markov chain Monte Carlo (MCMC), such as Langevin dynamics, is valid for approximating intractable distributions. However, its usage is limited in the context of deep latent variable models owing to costly datapoint-wise sampling iterations and slow convergence. This paper proposes the amortized Langevin dynamics (ALD), wherein datapoint-wise MCMC iterations are entirely replaced with updates of an encoder that maps observations into latent variables. This amortization enables efficient posterior sampling without datapoint-wise iterations. Despite its efficiency, we prove that ALD is valid as an MCMC algorithm, whose Markov chain has the target posterior as a stationary distribution under mild assumptions. Based on the ALD, we also present a new deep latent variable model named the Langevin autoencoder (LAE). Interestingly, the LAE can be implemented by slightly modifying the traditional autoencoder. Using multiple synthetic datasets, we first validate that ALD can properly obtain samples from target posteriors. We also evaluate the LAE on the image generation task, and show that our LAE can outperform existing methods based on variational inference, such as the variational autoencoder, and other MCMC-based methods in terms of the test likelihood.

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.

Shota Takashiro, Masanori Koyama, Takeru Miyato et al.

We extend the recent latent recurrent modeling to sequential input streams. By interleaving fast, recurrent latent updates with self-organizational ability between slow observation updates, our method facilitates the learning of stable internal structures that evolve alongside the input. This mechanism allows the model to maintain coherent and clustered representations over long horizons, improving out-of-distribution generalization in reinforcement learning and algorithmic tasks compared to sequential baselines such as LSTM, state space models, and Transformer variants.

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.

Hisayuki Yokomizo, Taiki Miyanishi, Yan Gang et al.

Vision-Language Models (VLMs) are increasingly applied to robotic perception and manipulation, yet their ability to infer physical properties required for manipulation remains limited. In particular, estimating the mass of real-world objects is essential for determining appropriate grasp force and ensuring safe interaction. However, current VLMs lack reliable mass reasoning capabilities, and most existing benchmarks do not explicitly evaluate physical quantity estimation under realistic sensing conditions. In this work, we propose PhysQuantAgent, a framework for real-world object mass estimation using VLMs, together with VisPhysQuant, a new benchmark dataset for evaluation. VisPhysQuant consists of RGB-D videos of real objects captured from multiple viewpoints, annotated with precise mass measurements. To improve estimation accuracy, we introduce three visual prompting methods that enhance the input image with object detection, scale estimation, and cross-sectional image generation to help the model comprehend the size and internal structure of the target object. Experiments show that visual prompting significantly improves mass estimation accuracy on real-world data, suggesting the efficacy of integrating spatial reasoning with VLM knowledge for physical inference.

Dingkun Yan, Xinrui Wang, Ru Wang et al.

Sketch colorization is a critical task for automating and assisting in the creation of animations and digital illustrations. Previous research identified the primary difficulty as the distribution shift between semantically aligned training data and highly diverse test data, and focused on mitigating the artifacts caused by the distribution shift instead of fundamentally resolving the problem. In this paper, we present a framework that directly minimizes the distribution shift, thereby achieving superior quality, resolution, and controllability of colorization. We propose a dual-branch framework to explicitly model the data distributions of the training process and inference process with a semantic-aligned branch and a semantic-misaligned branch, respectively. A Gram Regularization Loss is applied across the feature maps of both branches, effectively enforcing cross-domain distribution coherence and stability. Furthermore, we adopt an anime-specific Tagger Network to extract fine-grained attributions from reference images and modulate SDXL's conditional encoders to ensure precise control, and a plugin module to enhance texture transfer. Quantitative and qualitative comparisons, alongside user studies, confirm that our method effectively overcomes the distribution shift challenge, establishing State-of-the-Art performance across both quality and controllability metrics. Ablation study reveals the influence of each component.

Yingjian Chen, Haoran Liu, Yinhong Liu et al.

Large Language Models (LLMs) show strong reasoning ability in open-domain question answering, yet their reasoning processes are typically linear and often logically inconsistent. In contrast, real-world reasoning requires integrating multiple premises and solving subproblems in parallel. Existing methods, such as Chain-of-Thought (CoT), express reasoning in a linear textual form, which may appear coherent but frequently leads to inconsistent conclusions. Recent approaches rely on externally provided graphs and do not explore how LLMs can construct and use their own graph-structured reasoning, particularly in open-domain QA. To fill this gap, we novelly explore graph-structured reasoning of LLMs in general-domain question answering. We propose Self-Graph Reasoning (SGR), a framework that enables LLMs to explicitly represent their reasoning process as a structured graph before producing the final answer. We further construct a graph-structured reasoning dataset that merges multiple candidate reasoning graphs into refined graph structures for model training. Experiments on five QA benchmarks across both general and specialized domains show that SGR consistently improves reasoning consistency and yields a 17.74% gain over the base model. The LLaMA-3.3-70B model fine-tuned with SGR performs comparably to GPT-4o and surpasses Claude-3.5-Haiku, demonstrating the effectiveness of graph-structured reasoning.

Yuta Oshima, Yusuke Iwasawa, Masahiro Suzuki et al.

Video world models have attracted significant attention for their ability to produce high-fidelity future visual observations conditioned on past observations and navigation actions. Temporally- and spatially-consistent, long-term world modeling has been a long-standing problem, unresolved with even recent state-of-the-art models, due to the prohibitively expensive computational costs for long-context inputs. In this paper, we propose WorldPack, a video world model with efficient compressed memory, which significantly improves spatial consistency, fidelity, and quality in long-term generation despite much shorter context length. Our compressed memory consists of trajectory packing and memory retrieval; trajectory packing realizes high context efficiency, and memory retrieval maintains the consistency in rollouts and helps long-term generations that require spatial reasoning. Our performance is evaluated with LoopNav, a benchmark on Minecraft, specialized for the evaluation of long-term consistency, and we verify that WorldPack notably outperforms strong state-of-the-art models.

Bum Jun Kim, Shohei Taniguchi, Makoto Kawano et al.

Training divergence in transformers wastes compute, yet practitioners discover instability only after expensive runs begin. They therefore need an expected probability of failure for a transformer before training starts. Our study of Residual Koopman Spectral Profiling (RKSP) provides such an estimate. From a single forward pass at initialization, RKSP extracts Koopman spectral features by applying whitened dynamic mode decomposition to layer-wise residual snapshots. Our central diagnostic, the near-unit spectral mass, quantifies the fraction of modes concentrated near the unit circle, which captures instability risk. For predicting divergence across extensive configurations, this estimator achieves an AUROC of 0.995, outperforming the best gradient baseline. We further make this diagnostic actionable through Koopman Spectral Shaping (KSS), which reshapes spectra during training. We empirically validate that our method works in practice: RKSP predicts divergence at initialization, and when RKSP flags high risk, turning on KSS successfully prevents divergence. In the challenging high learning rate regime without normalization layers, KSS reduces the divergence rate from 66.7% to 12.5% and enables learning rates that are 50% to 150% higher. These findings generalize to WikiText-103 language modeling, vision transformers on CIFAR-10, and pretrained language models, including GPT-2 and LLaMA-2 up to 7B, as well as emerging architectures such as MoE, Mamba-style SSMs, and KAN.

Ru Wang, Wei Huang, Qi Cao et al.

Test-time reinforcement learning (TTRL) offers a label-free paradigm for adapting models using only synthetic signals at inference, but its success hinges on constructing reliable learning signals. Standard approaches such as majority voting often collapse to spurious yet popular answers. We introduce Self-Harmony, a framework built on a simple intuition: the correct answer should remain stable across both an original question and its paraphrase. Self-Harmony operationalizes this by employing a single model in two complementary roles: a Solver to produce answers and a Reframer to rephrase the input. Based on this, we further propose a pseudo-label method: instead of majority voting, it aggregates answer frequencies across these original and reframed views using the harmonic mean. This is a process that naturally selects for solutions stable under reframing, thereby avoiding the common trap of favoring view-dependent, spurious answers. Crucially, this requires no human supervision or auxiliary models. Across diverse reasoning benchmarks, Self-Harmony achieves state-of-the-art results at the label-free test-time setting, ranking first in 28 of 30 settings across multiple methods. Beyond accuracy, it demonstrates unprecedented robustness, with zero training failures in all experiments, underscoring its stability and reliability.

Kei Akuzawa, Yusuke Iwasawa, Yutaka Matsuo

Agents that can follow language instructions are expected to be useful in a variety of situations such as navigation. However, training neural network-based agents requires numerous paired trajectories and languages. This paper proposes using multimodal generative models for semi-supervised learning in the instruction following tasks. The models learn a shared representation of the paired data, and enable semi-supervised learning by reconstructing unpaired data through the representation. Key challenges in applying the models to sequence-to-sequence tasks including instruction following are learning a shared representation of variable-length mulitimodal data and incorporating attention mechanisms. To address the problems, this paper proposes a novel network architecture to absorb the difference in the sequence lengths of the multimodal data. In addition, to further improve the performance, this paper shows how to incorporate the generative model-based approach with an existing semi-supervised method called a speaker-follower model, and proposes a regularization term that improves inference using unpaired trajectories. Experiments on BabyAI and Room-to-Room (R2R) environments show that the proposed method improves the performance of instruction following by leveraging unpaired data, and improves the performance of the speaker-follower model by 2\% to 4\% in R2R.

Shohei Taniguchi, Keno Harada, Gouki Minegishi et al.

Adam is one of the most popular optimization algorithms in deep learning. However, it is known that Adam does not converge in theory unless choosing a hyperparameter, i.e., $β_2$, in a problem-dependent manner. There have been many attempts to fix the non-convergence (e.g., AMSGrad), but they require an impractical assumption that the gradient noise is uniformly bounded. In this paper, we propose a new adaptive gradient method named ADOPT, which achieves the optimal convergence rate of $\mathcal{O} ( 1 / \sqrt{T} )$ with any choice of $β_2$ without depending on the bounded noise assumption. ADOPT addresses the non-convergence issue of Adam by removing the current gradient from the second moment estimate and changing the order of the momentum update and the normalization by the second moment estimate. We also conduct intensive numerical experiments, and verify that our ADOPT achieves superior results compared to Adam and its variants across a wide range of tasks, including image classification, generative modeling, natural language processing, and deep reinforcement learning. The implementation is available at https://github.com/iShohei220/adopt.

Xinjie Zhao, Moritz Blum, Fan Gao et al.

AGENTiGraph is a user-friendly, agent-driven system that enables intuitive interaction and management of domain-specific data through the manipulation of knowledge graphs in natural language. It gives non-technical users a complete, visual solution to incrementally build and refine their knowledge bases, allowing multi-round dialogues and dynamic updates without specialized query languages. The flexible design of AGENTiGraph, including intent classification, task planning, and automatic knowledge integration, ensures seamless reasoning between diverse tasks. Evaluated on a 3,500-query benchmark within an educational scenario, the system outperforms strong zero-shot baselines (achieving 95.12% classification accuracy, 90.45% execution success), indicating potential scalability to compliance-critical or multi-step queries in legal and medical domains, e.g., incorporating new statutes or research on the fly. Our open-source demo offers a powerful new paradigm for multi-turn enterprise knowledge management that bridges LLMs and structured graphs.

Yuta Oshima, Daiki Miyake, Kohsei Matsutani et al.

Recent text-to-image generation models have acquired the ability of multi-reference generation and editing; the ability to inherit the appearance of subjects from multiple reference images and re-render them under new contexts. However, the existing benchmark datasets often focus on the generation with single or a few reference images, which prevents us from measuring the progress on how model performance advances or pointing out their weaknesses, under different multi-reference conditions. In addition, their task definitions are still vague, typically limited to axes such as "what to edit" or "how many references are given", and therefore fail to capture the intrinsic difficulty of multi-reference settings. To address this gap, we introduce $\textbf{MultiBanana}$, which is carefully designed to assesses the edge of model capabilities by widely covering multi-reference-specific problems at scale: (1) varying the number of references, (2) domain mismatch among references (e.g., photo vs. anime), (3) scale mismatch between reference and target scenes, (4) references containing rare concepts (e.g., a red banana), and (5) multilingual textual references for rendering. Our analysis among a variety of text-to-image models reveals their superior performances, typical failure modes, and areas for improvement. MultiBanana will be released as an open benchmark to push the boundaries and establish a standardized basis for fair comparison in multi-reference image generation. Our data and code are available at https://github.com/matsuolab/multibanana .

Ryosuke Takanami, Petr Khrapchenkov, Shu Morikuni et al.

As robots transition from controlled settings to unstructured human environments, building generalist agents that can reliably follow natural language instructions remains a central challenge. Progress in robust mobile manipulation requires large-scale multimodal datasets that capture contact-rich and long-horizon tasks, yet existing resources lack synchronized force-torque sensing, hierarchical annotations, and explicit failure cases. We address this gap with the AIRoA MoMa Dataset, a large-scale real-world multimodal dataset for mobile manipulation. It includes synchronized RGB images, joint states, six-axis wrist force-torque signals, and internal robot states, together with a novel two-layer annotation schema of sub-goals and primitive actions for hierarchical learning and error analysis. The initial dataset comprises 25,469 episodes (approx. 94 hours) collected with the Human Support Robot (HSR) and is fully standardized in the LeRobot v2.1 format. By uniquely integrating mobile manipulation, contact-rich interaction, and long-horizon structure, AIRoA MoMa provides a critical benchmark for advancing the next generation of Vision-Language-Action models. The first version of our dataset is now available at https://huggingface.co/datasets/airoa-org/airoa-moma .

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.

Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo et al.

Domain generalization (DG) is a difficult transfer learning problem aiming to learn a generalizable model for unseen domains. Recent foundation models (FMs) are robust to many distribution shifts and, therefore, should substantially improve the performance of DG. In this work, we study generic ways to adopt CLIP, a Visual-Language Foundation Model, for DG problems in image classification. While ERM greatly improves the accuracy with bigger backbones and training datasets using standard DG benchmarks, fine-tuning FMs is not practical in many real-world situations. We propose Domain Prompt Learning (DPL) as a novel approach for domain inference in the form of conditional prompt generation. DPL achieved a significant accuracy improvement with only training a lightweight prompt generator (a three-layer MLP), whose parameter is of equivalent scale to the classification projector in the previous DG literature. Combining \dplshort~with CLIP provides surprising performance, raising the accuracy of zero-shot CLIP from 73.7% to 79.3% on several standard datasets, namely PACS, VLCS, OfficeHome, and TerraIncognita. We hope the simplicity and success of our approach lead to broader adoption and analysis of foundation models in the domain generalization field. Our code is available at https://github.com/shogi880/DPLCLIP.

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.

Bum Jun Kim, Makoto Kawano, Yusuke Iwasawa et al.

Normalization layers in neural operators usually compute statistics by uniformly averaging discrete grid values, making the normalization itself discretization-dependent and thereby a source of transfer error across different resolutions or meshes. To enable discretization robustness, we introduce a quadrature normalization family that replaces existing uniform averaging in normalization layers with numerical quadrature: QuadNorm and BlendQuadNorm. On endpoint-inclusive uniform grids, the proposed quadrature moments are $O(h^2)$-consistent across discretizations, meaning that their cross-resolution mismatch decays quadratically with grid spacing. A transfer-error bound then predicts how normalization-induced mismatch scales with both the resolution gap and network depth. The experiments show the same gap- and depth-scaling trends predicted by the transfer-error bound. On Darcy, QuadNorm delivers the best cross-resolution performance at every tested target resolution from $64^2$ to $256^2$; on real-data benchmarks, Transolver with QuadNorm achieves nearly resolution-invariant transfer. The largest gains appear on nonperiodic PDEs and nonspectral architectures, where native-resolution improvements also emerge. We also validate BlendQuadNorm, which stays close to LayerNorm behavior and serves as a conservative default for periodic FNO settings. These results identify normalization as a previously overlooked source of resolution dependence in neural operators.

Weihao Xuan, Rui Yang, Heli Qi et al.

Existing large language model (LLM) evaluation benchmarks primarily focus on English, while current multilingual tasks lack parallel questions that specifically assess cross-linguistic reasoning abilities. This dual limitation makes it challenging to comprehensively assess LLMs' performance in the multilingual setting. To fill this gap, we introduce MMLU-ProX, a comprehensive benchmark covering 29 languages, built on an English benchmark. Each language version consists of 11,829 identical questions, enabling direct cross-linguistic comparisons. Additionally, to meet efficient evaluation needs, we provide a lite version containing 658 questions per language. To ensure the high quality of MMLU-ProX, we employ a rigorous development process that involves multiple powerful LLMs for translation, followed by expert review to ensure accurate expression, consistent terminology, and cultural relevance. Building on this, we systematically evaluate 36 state-of-the-art LLMs, including reasoning-enhanced and multilingual-optimized LLMs. The results reveal significant disparities in the multilingual capabilities of LLMs: While they perform well in high-resource languages, their performance declines markedly in low-resource languages, with gaps of up to 24.3%. Through MMLU-ProX, we aim to advance the development of more inclusive AI systems and promote equitable access to technology across global contexts.

Gouki Minegishi, Hiroki Furuta, Yusuke Iwasawa et al.

Sparse autoencoders (SAEs) have gained a lot of attention as a promising tool to improve the interpretability of large language models (LLMs) by mapping the complex superposition of polysemantic neurons into monosemantic features and composing a sparse dictionary of words. However, traditional performance metrics like Mean Squared Error and L0 sparsity ignore the evaluation of the semantic representational power of SAEs -- whether they can acquire interpretable monosemantic features while preserving the semantic relationship of words. For instance, it is not obvious whether a learned sparse feature could distinguish different meanings in one word. In this paper, we propose a suite of evaluations for SAEs to analyze the quality of monosemantic features by focusing on polysemous words. Our findings reveal that SAEs developed to improve the MSE-L0 Pareto frontier may confuse interpretability, which does not necessarily enhance the extraction of monosemantic features. The analysis of SAEs with polysemous words can also figure out the internal mechanism of LLMs; deeper layers and the Attention module contribute to distinguishing polysemy in a word. Our semantics focused evaluation offers new insights into the polysemy and the existing SAE objective and contributes to the development of more practical SAEs.

Hiroki Furuta, Gouki Minegishi, Yusuke Iwasawa et al.

Grokking has been actively explored to reveal the mystery of delayed generalization and identifying interpretable representations and algorithms inside the grokked models is a suggestive hint to understanding its mechanism. Grokking on modular addition has been known to implement Fourier representation and its calculation circuits with trigonometric identities in Transformers. Considering the periodicity in modular arithmetic, the natural question is to what extent these explanations and interpretations hold for the grokking on other modular operations beyond addition. For a closer look, we first hypothesize that any modular operations can be characterized with distinctive Fourier representation or internal circuits, grokked models obtain common features transferable among similar operations, and mixing datasets with similar operations promotes grokking. Then, we extensively examine them by learning Transformers on complex modular arithmetic tasks, including polynomials. Our Fourier analysis and novel progress measure for modular arithmetic, Fourier Frequency Density and Fourier Coefficient Ratio, characterize distinctive internal representations of grokked models per modular operation; for instance, polynomials often result in the superposition of the Fourier components seen in elementary arithmetic, but clear patterns do not emerge in challenging non-factorizable polynomials. In contrast, our ablation study on the pre-grokked models reveals that the transferability among the models grokked with each operation can be only limited to specific combinations, such as from elementary arithmetic to linear expressions. Moreover, some multi-task mixtures may lead to co-grokking -- where grokking simultaneously happens for all the tasks -- and accelerate generalization, while others may not find optimal solutions. We provide empirical steps towards the interpretability of internal circuits.

Gouki Minegishi, Hiroki Furuta, Shohei Taniguchi et al.

Transformer-based language models exhibit In-Context Learning (ICL), where predictions are made adaptively based on context. While prior work links induction heads to ICL through a sudden jump in accuracy, this can only account for ICL when the answer is included within the context. However, an important property of practical ICL in large language models is the ability to meta-learn how to solve tasks from context, rather than just copying answers from context; how such an ability is obtained during training is largely unexplored. In this paper, we experimentally clarify how such meta-learning ability is acquired by analyzing the dynamics of the model's circuit during training. Specifically, we extend the copy task from previous research into an In-Context Meta Learning setting, where models must infer a task from examples to answer queries. Interestingly, in this setting, we find that there are multiple phases in the process of acquiring such abilities, and that a unique circuit emerges in each phase, contrasting with the single-phases change in induction heads. The emergence of such circuits can be related to several phenomena known in large language models, and our analysis lead to a deeper understanding of the source of the transformer's ICL ability.

Xinjie Zhao, Fan Gao, Xingyu Song et al.

Recent advances in large language models (LLMs) have significantly improved multi-hop question answering (QA) through direct Chain-of-Thought (CoT) reasoning. However, the irreversible nature of CoT leads to error accumulation, making it challenging to correct mistakes in multi-hop reasoning. This paper introduces ReAgent: a Reversible multi-Agent collaborative framework augmented with explicit backtracking mechanisms, enabling reversible multi-hop reasoning. By incorporating text-based retrieval, information aggregation and validation, our system can detect and correct errors mid-reasoning, leading to more robust and interpretable QA outcomes. The framework and experiments serve as a foundation for future work on error-tolerant QA systems. Empirical evaluations across three benchmarks indicate ReAgent's efficacy, yielding average about 6\% improvements against baseline models.

Ru Wang, Wei Huang, Selena Song et al.

Generalization to novel compound tasks under distribution shift is important for deploying transformer-based language models (LMs). This work investigates Chain-of-Thought (CoT) reasoning as a means to enhance OOD generalization. Through controlled experiments across several compound tasks, we reveal three key insights: (1) While QA-trained models achieve near-perfect in-distribution accuracy, their OOD performance degrades catastrophically, even with 10000k+ training examples; (2) the granularity of CoT data strongly correlates with generalization performance; finer-grained CoT data leads to better generalization; (3) CoT exhibits remarkable sample efficiency, matching QA performance with much less (even 80%) data. Theoretically, we demonstrate that compound tasks inherently permit shortcuts in Q-A data that misalign with true reasoning principles, while CoT forces internalization of valid dependency structures, and thus can achieve better generalization. Further, we show that transformer positional embeddings can amplify generalization by emphasizing subtask condition recurrence in long CoT sequences. Our combined theoretical and empirical analysis provides compelling evidence for CoT reasoning as a crucial training paradigm for enabling LM generalization under real-world distributional shifts for compound tasks.

Dingkun Yan, Xinrui Wang, Zhuoru Li et al.

Sketch colorization plays an important role in animation and digital illustration production tasks. However, existing methods still meet problems in that text-guided methods fail to provide accurate color and style reference, hint-guided methods still involve manual operation, and image-referenced methods are prone to cause artifacts. To address these limitations, we propose a diffusion-based framework inspired by real-world animation production workflows. Our approach leverages the sketch as the spatial guidance and an RGB image as the color reference, and separately extracts foreground and background from the reference image with spatial masks. Particularly, we introduce a split cross-attention mechanism with LoRA (Low-Rank Adaptation) modules. They are trained separately with foreground and background regions to control the corresponding embeddings for keys and values in cross-attention. This design allows the diffusion model to integrate information from foreground and background independently, preventing interference and eliminating the spatial artifacts. During inference, we design switchable inference modes for diverse use scenarios by changing modules activated in the framework. Extensive qualitative and quantitative experiments, along with user studies, demonstrate our advantages over existing methods in generating high-qualigy artifact-free results with geometric mismatched references. Ablation studies further confirm the effectiveness of each component. Codes are available at https://github.com/ tellurion-kanata/colorizeDiffusion.

Bo Yang, Jiaxian Guo, Yusuke Iwasawa et al.

Recent studies have increasingly demonstrated that large language models (LLMs) possess significant theory of mind (ToM) capabilities, showing the potential for simulating the tracking of mental states in generative agents. In this study, we propose a novel paradigm called ToM-agent, designed to empower LLMs-based generative agents to simulate ToM in open-domain conversational interactions. ToM-agent disentangles the confidence from mental states, facilitating the emulation of an agent's perception of its counterpart's mental states, such as beliefs, desires, and intentions (BDIs). Using past conversation history and verbal reflections, ToM-Agent can dynamically adjust counterparts' inferred BDIs, along with related confidence levels. We further put forth a counterfactual intervention method that reflects on the gap between the predicted responses of counterparts and their real utterances, thereby enhancing the efficiency of reflection. Leveraging empathetic and persuasion dialogue datasets, we assess the advantages of implementing the ToM-agent with downstream tasks, as well as its performance in both the first-order and the \textit{second-order} ToM. Our findings indicate that the ToM-agent can grasp the underlying reasons for their counterpart's behaviors beyond mere semantic-emotional supporting or decision-making based on common sense, providing new insights for studying large-scale LLMs-based simulation of human social behaviors.