Tetsuro Morimura

Yuu Jinnai, Tetsuro Morimura, Ukyo Honda et al.

Minimum Bayes Risk (MBR) decoding has been shown to be a powerful alternative to beam search decoding in a variety of text generation tasks. MBR decoding selects a hypothesis from a pool of hypotheses that has the least expected risk under a probability model according to a given utility function. Since it is impractical to compute the expected risk exactly over all possible hypotheses, two approximations are commonly used in MBR. First, it integrates over a sampled set of hypotheses rather than over all possible hypotheses. Second, it estimates the probability of each hypothesis using a Monte Carlo estimator. While the first approximation is necessary to make it computationally feasible, the second is not essential since we typically have access to the model probability at inference time. We propose Model-Based MBR (MBMBR), a variant of MBR that uses the model probability itself as the estimate of the probability distribution instead of the Monte Carlo estimate. We show analytically and empirically that the model-based estimate is more promising than the Monte Carlo estimate in text generation tasks. Our experiments show that MBMBR outperforms MBR in several text generation tasks, both with encoder-decoder models and with large language models.

Tetsuro Morimura, Mitsuki Sakamoto, Yuu Jinnai et al.

Reinforcement learning from human feedback (RLHF) plays a crucial role in aligning language models with human preferences. While the significance of dataset quality is generally recognized, explicit investigations into its impact within the RLHF framework, to our knowledge, have been limited. This paper addresses the issue of text quality within the preference dataset by focusing on direct preference optimization (DPO), an increasingly adopted reward-model-free RLHF method. We confirm that text quality significantly influences the performance of models optimized with DPO more than those optimized with reward-model-based RLHF. Building on this new insight, we propose an extension of DPO, termed filtered direct preference optimization (fDPO). fDPO uses a trained reward model to monitor the quality of texts within the preference dataset during DPO training. Samples of lower quality are discarded based on comparisons with texts generated by the model being optimized, resulting in a more accurate dataset. Experimental results demonstrate that fDPO enhances the final model performance. Our code is available at https://github.com/CyberAgentAILab/filtered-dpo.

Yuu Jinnai, Tetsuro Morimura, Ukyo Honda

Beam search and exhaustive search are two extreme ends of text decoding algorithms with respect to the search depth. Beam search is limited in both search width and depth, whereas exhaustive search is a global search that has no such limitations. Surprisingly, beam search is not only computationally cheaper but also performs better than exhaustive search despite its higher search error. Plenty of research has investigated a range of beam widths, from small to large, and reported that a beam width that is neither too large nor too small is desirable. However, in terms of search depth, only the two extreme ends, beam search and exhaustive search are studied intensively. In this paper, we examine a range of search depths between the two extremes to discover the desirable search depth. To this end, we introduce Lookahead Beam Search (LBS), a multi-step lookahead search that optimizes the objective considering a fixed number of future steps. Beam search and exhaustive search are special cases of LBS where the lookahead depth is set to $0$ and $\infty$, respectively. We empirically evaluate the performance of LBS and find that it outperforms beam search overall on machine translation tasks. The result suggests there is room for improvement in beam search by searching deeper. Inspired by the analysis, we propose Lookbehind Heuristic Beam Search, a computationally feasible search algorithm that heuristically simulates LBS with 1-step lookahead. The empirical results show that the proposed method outperforms vanilla beam search on machine translation and text summarization tasks.

Riku Togashi, Tatsushi Oka, Naoto Ohsaka et al.

Excellent tail performance is crucial for modern machine learning tasks, such as algorithmic fairness, class imbalance, and risk-sensitive decision making, as it ensures the effective handling of challenging samples within a dataset. Tail performance is also a vital determinant of success for personalized recommender systems to reduce the risk of losing users with low satisfaction. This study introduces a "safe" collaborative filtering method that prioritizes recommendation quality for less-satisfied users rather than focusing on the average performance. Our approach minimizes the conditional value at risk (CVaR), which represents the average risk over the tails of users' loss. To overcome computational challenges for web-scale recommender systems, we develop a robust yet practical algorithm that extends the most scalable method, implicit alternating least squares (iALS). Empirical evaluation on real-world datasets demonstrates the excellent tail performance of our approach while maintaining competitive computational efficiency.

Yosuke Miyanishi, Tetsuro Morimura

Spatial reasoning requires both location-bound computation and location-invariant structure: agents must make local moves while preserving route, object, or constraint-level plans. We propose interaction locality, a task-geometry-aware framework for measuring whether information flow stays within nearby cells or semantic segments, or crosses them. We instantiate the framework with sparse-autoencoder feature ablations and finite-noise activation patching, with structural Jacobian and attention checks reported in the appendix, and apply it to HRM and TRM, two compact hierarchical and recursive reasoning models, on Maze-Hard, Sudoku Extreme, and ARC-AGI. Across these models, activation patching gives the clearest architectural fingerprint: high-level recurrent states tend to write information within nearby cells or same-segment units, while repeated recursive updates accumulate these local writes into broader solution structure. This pattern holds across maze paths, Sudoku constraints, and ARC-AGI object neighborhoods, with the strongest concentration in TRM. To test whether interaction locality extends beyond toy-yet-challenging grid benchmarks, we also apply it to MTU3D, a large-scale embodied 3D scene-grounding model. In this MTU3D setting, causal spatial locality appears primarily at the transition where visual scene features are handed to the downstream grounding module, rather than uniformly throughout the visual encoder. This contrast suggests that the local-to-global handoff observed in HRM and TRM is tied to explicit recursive reasoning dynamics, while embodied 3D models may concentrate causal spatial structure at module boundaries. Interaction locality turns the intuitive local-execution/global-planning story into a reproducible measurement framework for recursive and embodied spatial reasoning.

Satoshi Hayakawa, Tetsuro Morimura

Reward evaluation of episodes becomes a bottleneck in a broad range of reinforcement learning tasks. Our aim in this paper is to select a small but representative subset of a large batch of episodes, only on which we actually compute rewards for more efficient policy gradient iterations. We build a Gaussian process modeling of discounted returns or rewards to derive a positive definite kernel on the space of episodes, run an ``episodic" kernel quadrature method to compress the information of sample episodes, and pass the reduced episodes to the policy network for gradient updates. We present the theoretical background of this procedure as well as its numerical illustrations in MuJoCo tasks.

Sho Shimoyama, Tetsuro Morimura, Kenshi Abe et al.

Dialog policies, which determine a system's action based on the current state at each dialog turn, are crucial to the success of the dialog. In recent years, reinforcement learning (RL) has emerged as a promising option for dialog policy learning (DPL). In RL-based DPL, dialog policies are updated according to rewards. The manual construction of fine-grained rewards, such as state-action-based ones, to effectively guide the dialog policy is challenging in multi-domain task-oriented dialog scenarios with numerous state-action pair combinations. One way to estimate rewards from collected data is to train the reward estimator and dialog policy simultaneously using adversarial learning (AL). Although this method has demonstrated superior performance experimentally, it is fraught with the inherent problems of AL, such as mode collapse. This paper first identifies the role of AL in DPL through detailed analyses of the objective functions of dialog policy and reward estimator. Next, based on these analyses, we propose a method that eliminates AL from reward estimation and DPL while retaining its advantages. We evaluate our method using MultiWOZ, a multi-domain task-oriented dialog corpus.

Yuu Jinnai, Tetsuro Morimura, Kaito Ariu et al.

Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) to human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking when the accuracy of the reward model is not high enough due to the quality or the quantity of the preference dataset. Because the reward model is an imperfect proxy for the true objective, over-optimizing its value can compromise its performance on the true objective. In this research, we propose MBR-BoN, a variant of BoN that aims to mitigate reward hacking at inference time by incorporating the Minimum Bayes Risk (MBR) objective as a proximity regularization term. We show empirically and analytically that the MBR objective quantifies the proximity of the response to the reference policy, serving as a proximity regularizer. We evaluate MBR-BoN on the AlpacaFarm and Anthropic's hh-rlhf datasets and show that it outperforms both BoN sampling and MBR decoding. We also evaluate MBR-BoN to generate a pairwise preference learning dataset for Direct Preference Optimization (DPO). Empirical results show that models trained on a dataset generated with MBR-BoN outperform those with vanilla BoN. Our code is available at https://github.com/CyberAgentAILab/regularized-bon

Tetsuro Morimura, Kazuhiro Ota, Kenshi Abe et al.

Policy gradient (PG) is a reinforcement learning (RL) approach that optimizes a parameterized policy model for an expected return using gradient ascent. While PG can work well even in non-Markovian environments, it may encounter plateaus or peakiness issues. As another successful RL approach, algorithms based on Monte Carlo Tree Search (MCTS), which include AlphaZero, have obtained groundbreaking results, especially in the game-playing domain. They are also effective when applied to non-Markov decision processes. However, the standard MCTS is a method for decision-time planning, which differs from the online RL setting. In this work, we first introduce Monte Carlo Tree Learning (MCTL), an adaptation of MCTS for online RL setups. We then explore a combined policy approach of PG and MCTL to leverage their strengths. We derive conditions for asymptotic convergence with the results of a two-timescale stochastic approximation and propose an algorithm that satisfies these conditions and converges to a reasonable solution. Our numerical experiments validate the effectiveness of the proposed methods.

Tsunehiko Tanaka, Kenshi Abe, Kaito Ariu et al.

Traditional approaches in offline reinforcement learning aim to learn the optimal policy that maximizes the cumulative reward, also known as return. It is increasingly important to adjust the performance of AI agents to meet human requirements, for example, in applications like video games and education tools. Decision Transformer (DT) optimizes a policy that generates actions conditioned on the target return through supervised learning and includes a mechanism to control the agent's performance using the target return. However, the action generation is hardly influenced by the target return because DT's self-attention allocates scarce attention scores to the return tokens. In this paper, we propose Return-Aligned Decision Transformer (RADT), designed to more effectively align the actual return with the target return. RADT leverages features extracted by paying attention solely to the return, enabling action generation to consistently depend on the target return. Extensive experiments show that RADT significantly reduces the discrepancies between the actual return and the target return compared to DT-based methods. Our code is available at https://github.com/CyberAgentAILab/radt

Yuki Ichihara, Yuu Jinnai, Tetsuro Morimura et al.

Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) with human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking. Since the reward model is an imperfect proxy for the true objective, an excessive focus on optimizing its value can lead to a compromise of its performance on the true objective. Previous work proposes Regularized BoN sampling (RBoN), a BoN sampling with regularization to the objective, and shows that it outperforms BoN sampling so that it mitigates reward hacking and empirically (Jinnai et al., 2024). However, Jinnai et al. (2024) introduce RBoN based on a heuristic and they lack the analysis of why such regularization strategy improves the performance of BoN sampling. The aim of this study is to analyze the effect of BoN sampling on regularization strategies. Using the regularization strategies corresponds to robust optimization, which maximizes the worst case over a set of possible perturbations in the proxy reward. Although the theoretical guarantees are not directly applicable to RBoN, RBoN corresponds to a practical implementation. This paper proposes an extension of the RBoN framework, called Stochastic RBoN sampling (SRBoN), which is a theoretically guaranteed approach to worst-case RBoN in proxy reward. We then perform an empirical evaluation using the AlpacaFarm and Anthropic's hh-rlhf datasets to evaluate which factors of the regularization strategies contribute to the improvement of the true proxy reward. In addition, we also propose another simple RBoN method, the Sentence Length Regularized BoN, which has a better performance in the experiment as compared to the previous methods.

Yuu Jinnai, Ukyo Honda, Tetsuro Morimura et al.

One of the most important challenges in text generation systems is to produce outputs that are not only correct but also diverse. Recently, Minimum Bayes-Risk (MBR) decoding has gained prominence for generating sentences of the highest quality among the decoding algorithms. However, existing algorithms proposed for generating diverse outputs are predominantly based on beam search or random sampling, thus their output quality is capped by these underlying methods. In this paper, we investigate an alternative approach -- we develop diversity-promoting decoding algorithms by enforcing diversity objectives to MBR decoding. We propose two variants of MBR, Diverse MBR (DMBR) and $k$-medoids MBR (KMBR), methods to generate a set of sentences with high quality and diversity. We evaluate DMBR and KMBR on a variety of directed text generation tasks using encoder-decoder models and a large language model with prompting. The experimental results show that the proposed method achieves a better trade-off than the diverse beam search and sampling algorithms.

Atsumoto Ohashi, Ukyo Honda, Tetsuro Morimura et al.

Minimum Bayes-risk (MBR) decoding has recently gained renewed attention in text generation. MBR decoding considers texts sampled from a model as pseudo-references and selects the text with the highest similarity to the others. Therefore, sampling is one of the key elements of MBR decoding, and previous studies reported that the performance varies by sampling methods. From a theoretical standpoint, this performance variation is likely tied to how closely the samples approximate the true distribution of references. However, this approximation has not been the subject of in-depth study. In this study, we propose using anomaly detection to measure the degree of approximation. We first closely examine the performance variation and then show that previous hypotheses about samples do not correlate well with the variation, but our introduced anomaly scores do. The results are the first to empirically support the link between the performance and the core assumption of MBR decoding.

Yuki Ichihara, Yuu Jinnai, Kaito Ariu et al.

Minimum Bayes Risk (MBR) decoding optimizes output selection by maximizing the expected utility value of an underlying human distribution. While prior work has shown the effectiveness of MBR decoding through empirical evaluation, few studies have analytically investigated why the method is effective. As a result of our analysis, we show that, given the size $n$ of the reference hypothesis set used in computation, MBR decoding approaches the optimal solution with high probability at a rate of $O\left(n^{-\frac{1}{2}}\right)$, under certain assumptions, even though the language space $Y$ is significantly larger $|Y|\gg n$. This result helps to theoretically explain the strong performance observed in several prior empirical studies on MBR decoding. In addition, we provide the performance gap for maximum-a-posteriori (MAP) decoding and compare it to MBR decoding. The result of this paper indicates that MBR decoding tends to converge to the optimal solution faster than MAP decoding in several cases.

Yuki Ichihara, Yuu Jinnai, Tetsuro Morimura et al.

Group Relative Policy Optimization (GRPO) has been shown to be an effective algorithm when an accurate reward model is available. However, such a highly reliable reward model is not available in many real-world tasks. In this paper, we particularly focus on multi-objective settings, in which we identify that GRPO is vulnerable to reward hacking, optimizing only one of the objectives at the cost of the others. To address this issue, we propose MO-GRPO, an extension of GRPO with a simple normalization method to reweight the reward functions automatically according to the variances of their values. We first show analytically that MO-GRPO ensures that all reward functions contribute evenly to the loss function while preserving the order of preferences, eliminating the need for manual tuning of the reward functions' scales. Then, we evaluate MO-GRPO experimentally in four domains: (i) the multi-armed bandits problem, (ii) simulated control task (Mo-Gymnasium), (iii) machine translation tasks on the WMT benchmark (En-Ja, En-Zh), and (iv) instruction following task. MO-GRPO achieves stable learning by evenly distributing correlations among the components of rewards, outperforming GRPO, showing MO-GRPO to be a promising algorithm for multi-objective reinforcement learning problems.

Tetsuro Morimura, Tatsushi Oka, Yugo Suzuki et al.

Latent variable models provide a powerful framework for incorporating and inferring unobserved factors in observational data. In causal inference, they help account for hidden factors influencing treatment or outcome, thereby addressing challenges posed by missing or unmeasured covariates. This paper proposes a new framework that integrates latent variable modeling into the double machine learning (DML) paradigm to enable robust causal effect estimation in the presence of such hidden factors. We consider two scenarios: one where a latent variable affects only the outcome, and another where it may influence both treatment and outcome. To ensure tractability, we incorporate latent variables only in the second stage of DML, separating representation learning from latent inference. We demonstrate the robustness and effectiveness of our method through extensive experiments on both synthetic and real-world datasets.

Hao Wang, Tetsuro Morimura, Ukyo Honda et al.

Non-autoregressive (NAR) language models are known for their low latency in neural machine translation (NMT). However, a performance gap exists between NAR and autoregressive models due to the large decoding space and difficulty in capturing dependency between target words accurately. Compounding this, preparing appropriate training data for NAR models is a non-trivial task, often exacerbating exposure bias. To address these challenges, we apply reinforcement learning (RL) to Levenshtein Transformer, a representative edit-based NAR model, demonstrating that RL with self-generated data can enhance the performance of edit-based NAR models. We explore two RL approaches: stepwise reward maximization and episodic reward maximization. We discuss the respective pros and cons of these two approaches and empirically verify them. Moreover, we experimentally investigate the impact of temperature setting on performance, confirming the importance of proper temperature setting for NAR models' training.

Masahiro Kato, Kei Nakagawa, Kenshi Abe et al.

This study investigates the mean-variance (MV) trade-off in reinforcement learning (RL), an instance of the sequential decision-making under uncertainty. Our objective is to obtain MV-efficient policies whose means and variances are located on the Pareto efficient frontier with respect to the MV trade-off; under the condition, any increase in the expected reward would necessitate a corresponding increase in variance, and vice versa. To this end, we propose a method that trains our policy to maximize the expected quadratic utility, defined as a weighted sum of the first and second moments of the rewards obtained through our policy. We subsequently demonstrate that the maximizer indeed qualifies as an MV-efficient policy. Previous studies that employed constrained optimization to address the MV trade-off have encountered computational challenges. However, our approach is more computationally efficient as it eliminates the need for gradient estimation of variance, a contributing factor to the double sampling issue observed in existing methodologies. Through experimentation, we validate the efficacy of our approach.

Kun Zhao, Takayuki Osogami, Tetsuro Morimura

In team-based invasion sports such as soccer and basketball, analytics is important for teams to understand their performance and for audiences to understand matches better. The present work focuses on performing visual analytics to evaluate the value of any kind of event occurring in a sports match with a continuous parameter space. Here, the continuous parameter space involves the time, location, score, and other parameters. Because the spatiotemporal data used in such analytics is a low-level representation and has a very large size, however, traditional analytics may need to discretize the continuous parameter space (e.g., subdivide the playing area) or use a local feature to limit the analysis to specific events (e.g., only shots). These approaches make evaluation impossible for any kind of event with a continuous parameter space. To solve this problem, we consider a whole match as a Markov chain of significant events, so that event values can be estimated with a continuous parameter space by solving the Markov chain with a machine learning model. The significant events are first extracted by considering the time-varying distribution of players to represent the whole match. Then, the extracted events are redefined as different states with the continuous parameter space and built as a Markov chain so that a Markov reward process can be applied. Finally, the Markov reward process is solved by a customized fitted-value iteration algorithm so that the event values with the continuous parameter space can be predicted by a regression model. As a result, the event values can be visually inspected over the whole playing field under arbitrary given conditions. Experimental results with real soccer data show the effectiveness of the proposed system.

Yachiko Obara, Tetsuro Morimura, Hiroki Yanagisawa

Invention involves combination, or more precisely, ratios of composition. According to Thomas Edison, "Genius is one percent inspiration and 99 percent perspiration" is an example. In many situations, researchers and inventors already have a variety of data and manage to create something new by using it, but the key problem is how to select and combine knowledge. In this paper, we propose a new Markov chain Monte Carlo (MCMC) algorithm to generate composition ratios, nonnegative-integer-valued vectors with two properties: (i) the sum of the elements of each vector is constant, and (ii) only a small number of elements is nonzero. These constraints make it difficult for existing MCMC algorithms to sample composition ratios. The key points of our approach are (1) designing an appropriate target distribution by using a condition on the number of nonzero elements, and (2) changing values only between a certain pair of elements in each iteration. Through an experiment on creating a new cocktail, we show that the combination of the proposed method with supervised learning can solve a creative problem.

Tetsuro Morimura, Masashi Sugiyama, Hisashi Kashima et al.

Most conventional Reinforcement Learning (RL) algorithms aim to optimize decision-making rules in terms of the expected returns. However, especially for risk management purposes, other risk-sensitive criteria such as the value-at-risk or the expected shortfall are sometimes preferred in real applications. Here, we describe a parametric method for estimating density of the returns, which allows us to handle various criteria in a unified manner. We first extend the Bellman equation for the conditional expected return to cover a conditional probability density of the returns. Then we derive an extension of the TD-learning algorithm for estimating the return densities in an unknown environment. As test instances, several parametric density estimation algorithms are presented for the Gaussian, Laplace, and skewed Laplace distributions. We show that these algorithms lead to risk-sensitive as well as robust RL paradigms through numerical experiments.