Model-based deep reinforcement learning with world models has gained significant attention as a learning approach for intelligent agents. The performance of the agents strongly depends on the world model that represents environmental dynamics. Contrastive learning is a promising approach to train the world model and can be formulated as maximizing the mutual information between observations and hidden states. However, since mutual information calculation is generally intractable, many conventional approaches have used noise-contrastive estimation (NCE), a tractable lower bound. On the other hand, NCE is known to have an upper bound, which limits the mutual information that can be maximized. Therefore, more effective estimation methods are required. To address this, we propose a contrastive learning method for the world model training that replaces the NCE with a flow-matching-based mutual information estimation. Experiments in multiple image-based simulated environments suggest that the proposed method improves both sample efficiency and the maximum episode rewards in model-based reinforcement learning compared to conventional NCE-based contrastive learning.

Hierarchical world models enable efficient learning of long-term dependencies through temporal information compression. Conventional world models suffer from error accumulation in long-term predictions and difficulties in strategic planning due to sequential state updates at fine-grained resolutions. In this study, we integrate Variational Temporal Abstraction (VTA), which adaptively learns semantic boundaries from observations, into the DreamerV3 framework and evaluate its impact on the Atari100k benchmark. To provide a deeper understanding of the proposed world model, we evaluate the distance between state representations from different timesteps and reveal that the proposed world model selectively captures structural changes in the environment. However, cumulative rewards did not surpass the baseline, likely due to a timescale mismatch between abstract states and low-level policies. Our findings indicate that while adaptive temporal abstraction can effectively learn underlying hierarchical dynamics of the world, we still need to design an established framework for hierarchical policies with multi-timescale representations.

Spatial cognition and its description represent a key challenge in cognitive linguistics, with the Frame of Reference (FoR)—specifically the egocentric/allocentric distinction—serving as a core focus. This paper investigates how this FoR contrast is encoded within text and image vector spaces. Through mean difference vectors, we identify a linear FoR axis in text embeddings. We further demonstrate that this characteristic extends to the CLIP shared embedding space and that the egocentric/allocentric distinction exists as a latent linear feature in DINO image embeddings. Our results indicate that spatial cognition can be captured within a consistent mathematical framework, regardless of whether the input is text or image.

Object instance image retrieval requires accurate identification of the same object instance beyond category-level recognition, which demands stable representation learning. In recent years, self-supervised methods based on non-contrastive learning, such as SimSiam, have attracted considerable attention. However, challenges remain regarding training instability and adaptability to continual learning. Meanwhile, PhiNet has been proposed as a non-contrastive learning inspired by predictive processing mechanisms in the brain. PhiNet extends SimSiam by incorporating two predictors, thereby improving training stability during early learning. In this study, we propose a method that integrates PhiNet into a SimSiam-based method for object instance image retrieval. Experiments on the MIRO and ObjectPI datasets compared the SimSiam-based method. As a result, the proposed method demonstrated improved generalization and identification performances during the early learning.

World-model-based reinforcement learning achieves high sample efficiency by updating policies through imagined rollouts, but intrinsic-reward-driven exploration can become excessive under high uncertainty and destabilize training (the Noisy-TV problem). This study reinterprets intrinsic rewards not only as signals for encouraging exploration but also as control signals for stabilizing and promoting exploitation. Specifically, we use uncertainty-based intrinsic signals to regulate the selection, reuse, and prioritization of experiences from imagined rollouts, aiming to increase updates that contribute to task reward acquisition. Building on DreamerV3, we propose (i) TrendMix, which dynamically adjusts the mixing ratio of exploration- and exploitation-oriented replay based on reward trends (short/long-horizon EMAs) and model uncertainty, and (ii) EMA-Based Policy Shifting, which switches action dispersion according to reward dynamics. Experiments on a subset of Atari-100K tasks reveal tendencies toward improved early reward acquisition and behaviors suggestive of reduced instability from over-exploration. These results indicate that intrinsic signals in world-model RL may be effective not only for exploration but also for exploitation-oriented learning control.

When people share the same documents and observations yet reach different conclusions, the discussion often shifts into judging the other as “defective.” This is better described not as ignorance or bad faith but as non-identifiability inherent in inference operation and learning. We organize it into two levels: (i) θ-level non-identifiability, where outcomes diverge under the same world model W because inference settings differ; and (ii) W-level non-identifiability, where repeated use of θ biases data exposure and update rules, causing the learned world model W itself to diverge. We introduce an inference profile θ=(R,E,S,D)—Reference, Exploration, Stabilization, and Horizon—and show how outputs can split even for the same observation o and the same W. We further explain why disagreements tend to project onto a few bases (abstract/concrete, externalizability, order/freedom) from general constraints (computational, observational, and coordination). We map these axes to deep representation learning (hierarchy, latent-state estimation, regularization–exploration trade-offs) and illustrate the framework with a case study on AI regulation debates.

Planning based on world models learned from state–action trajectories is a promising route to general-purpose physical task solving. Joint-Embedding Predictive World Models optimize plans in a representation space rather than the input space, where abstraction can improve planning efficiency. We propose a language-conditioned latent-planning framework that preserves this non-pixel-generative design. We freeze a pretrained image encoder, latent dynamics model, and text encoder, and train only a lightweight GoalHead that predicts a goal latent z_goal from the initial image I0 and an instruction sentence T. Planning is carried out in latent space via model predictive control; the Cross-Entropy Method samples and refines action sequences to drive predicted latents toward z_goal. On the MetaWorld task suite, we use the standard setting in which z_goal is derived from a goal image as a baseline, and quantify performance degradation under language conditioning by measuring goal-latent estimation error (L2 and cosine distance), task success rate, and steps-to-reach. These results clarify the operating regime and limitations of representation-space planning without reliance on goal images.

In reinforcement learning and robotics, how actions are represented plays an important role in effective control. Many world models focus primarily on predicting future observations, treating actions merely as inputs that drive state transitions. Consequently, the semantic structure of actions is not explicitly modeled. This study proposes a framework that learns latent action representations solely from observation sequences and organizes them into abstract action units grounded in motion semantics. Changes between consecutive observations are encoded into latent representations using an inverse dynamics model. Contrastive learning based on motion categories encourages semantically similar actions to be embedded close to each other across different agents. Category-wise center representations are then extracted and treated as discrete action units. Experiments on the DeepMind Control Suite demonstrate that the learned latent actions are reusable across agents, enabling stable action generation and prediction over discrete action sequences. The results further indicate that incorporating domain information is crucial for achieving stable control across different agents.

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.n A strength of such models is that they enable test-time scaling, allowing them to improve performance at test time without additional training. 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 that maximizes the average of the top-1 prediction probabilities, reflecting the model’s confidence. It yields higher accuracy on Sudoku-hard than the energy-based voting strategy, which is specific to models with explicit energy functions. Finally, we propose ItrSA++, a simple recurrent model, and when combined with C-voting, demonstrates superior performance on both Sudoku-extreme and Maze-hard tasks compared to conventional approaches.

This study proposes a solution to core challenges in the autonomous operation of AI within physical and social environments, advocating for the "essentialization of intelligence through embodiment" modeled after human developmental mechanisms. While embodiment has traditionally been viewed as a control constraint, this research adopts a paradoxical perspective: "intelligence is forced into efficient learning precisely because of embodiment constraints," and develops a learning algorithm to enhance real-world execution.

In socially deployed world models, the key question is whether the decision pathways from model outputs to social consequences can be designed in advance as objects of scientific verification. This paper proposes a design principle that achieves this by topologically, not hierarchically, separating a judgment layer (fixed decision rules) from a computation layer (deep learning processes). This separation preserves computational capacity while fixing responsibility and verification boundaries. Conventional world models embed judgment within mutable deep learning processes, making post-deployment verification fundamentally difficult. In contrast, the proposed principle implements judgment as a fixed structure topologically separated from deep learning, while remaining within the same world model, thereby enabling falsifiable, verifiable, and reproducible outputs after deployment. Here, emergence is defined as a verifiably established update of principled constraints not derivable from prior specifications. Across roughly one year (about 2,500 hours) of examining multiple state-of-the-art LLMs, no autonomous emergence was observed within an LLM alone; emergence leading to patent applications occurred at points of connection with human judgment.

In threshold public goods games, allowing voluntary participation (i.e., giving individuals the option to opt out of the group) increases human cooperation rates, a phenomenon known as the voluntary participation effect. We examine whether six large language models can replicate this effect, evaluating (1) the voluntary participation effect itself, (2) belief-based self-selection, (3) belief updating among remaining participants, and (4) the association between stated beliefs and actual cooperation. None of the models reproduced either the effect or its underlying mechanisms. A central finding is a belief---action dissociation: in humans, reported cooperation beliefs strongly predict behavior regardless of participation condition, whereas GPT-4o matches this predictive power under mandatory participation but drops to chance level when voluntary participation is introduced. Although LLMs can report beliefs, they fail to perform institution-sensitive reasoning when a third option, opting out, is available.

The integration of Large Language Models into Multi-Agent Systems challenges traditional game-theoretic predictions, particularly regarding the user dynamics under different voting rules. This study investigates negotiation protocols as governing mechanisms for LLM behavior, comparing high-capacity and lightweight counterparts under varying voting rules and turn-taking structures. Our results reveal a sharp divergence: lightweight models exhibit premature convergence, rushing into low-quality decisions. While rigid turn-taking exacerbates a last-mover disadvantage for these agents, procedural stochasticity acts as a cognitive prosthetic, mitigating positional biases. Conversely, high-capacity models display alignment resilience, refusing to exploit majority rules to marginalize minorities. Instead, they consistently converge near Pareto, Nash, and Kalai-Smorodinsky optima. We attribute this to a conflict between protocol rationality and RLHF alignment biases. These findings underscore that future mechanism design must be capacity-aware, accounting for the variable cognitive constraints of AI agents.

In job-hunting activities, students are expected to submit high-quality resumes to pass document screening. However, university career support centers often cannot provide sufficient feedback to all students due to limited staff and time. Likewise, in companies, screening a large number of resumes places a considerable burden on hiring personnel. To address these challenges, recent studies have proposed LLM-based methods for automated resume screening. However, these methods are not directly applicable to Japan's unique job-hunting system. Therefore, this study develops an LLM-based automated resume screening framework named AutoScreen-FW, tailored to Japanese new graduate recruitment. Specifically, we employ multiple sample selection strategies to extract a small number of representative resume samples. These samples are then used to perform in-context learning, enabling the LLM to assess unseen resume automatically. Experimental results show that the proposed framework enables an open-source LLM to achieve resume evaluation performance comparable to, or better than, GPT-series models.

Approaches that proactively recommend the next item to purchase or the next action to take based on user behavior prediction have been proposed in recent years; however, because mainstream methods simultaneously optimize recommendation timing and content using only one-step-ahead behavior predictions, generating medium- to long-term sequences of recommended items has been difficult. To address this, we propose a multi-step behavior prediction model that introduces a profile representing users’ latent behavioral tendencies. We envisage an experience in which users can, at their chosen timing, review subsequent proposals all at once, and we focus on the stability of medium- to long-term predictions looking two or more steps ahead. The proposed method combines multi-horizon learning, which jointly learns multiple steps ahead, with latent profile representations to capture medium- to long-term behavioral transitions, aiming to have recommendations prepared when users need them. To validate its effectiveness, we conduct comparative experiments against existing methods based on one-step prediction and evaluate multi-step-ahead prediction performance as well as the evolution of uncertainty in future predictions.

The hotel industry faces the uncertainty of reservations and cancellations. Although dynamic pricing provides appropriate discounts for last-minute cancellations, it is not always possible to fill vacancies unless there is sufficient demand. Therefore, it is important to examine how the accuracy of cancellation forecasts affects the performance of dynamic pricing in practice. In this study, we analyze the relationship between forecast accuracy and expected returns using actual data on room reservations and cancellations from a region in Japan. The results show that an improvement in forecast accuracy leads to an increase in revenue. We also show that this improvement simultaneously increases consumer surplus by lowering the average sales price and improves the load factor. Furthermore, the improvement is more pronounced during the peak season when the number of potential customers is higher.

On two-sided matching platforms such as online dating and recruiting, recommendation algorithms often aim to maximize the total number of matches. However, this objective creates an imbalance, where some users receive too many matches while others receive very few and eventually abandon the platform. Although fairness objectives are sometimes used to mitigate this issue, fairness itself is often not the ultimate goal; in practice, user retention is more critical. In this work, we formalize the new problem setting of maximizing user retention in two-sided matching platforms. We propose a dynamic learning-to-rank (LTR) algorithm called Matching for Retention (MRet). Unlike conventional approaches, MRet models user retention by learning personalized retention curves from user profiles. Using these curves, MRet dynamically adapts recommendations by jointly considering the retention gains of both the recommending and recommended users. Experiments on synthetic datasets demonstrate that MRet achieves higher user retention.

Personalization in Large Language Models (LLMs) creates a risk of sycophancy (conforming to user preferences) and bias reinforcement. This can lead to narrowed perspectives and rigid thinking, which is particularly problematic in creative tasks where divergent thinking and unique viewpoints are essential. Based on the hypothesis that collaborating with an AI trained on others' thinking styles enhances creativity more than one trained on the user's own style, we conducted an experiment involving an advertising copywriting task. Participants collaborated with AI models trained on different thinking styles. Results showed that AI trained on others' styles facilitated a broader range of ideas, a sense of discovery, and higher originality. However, participants reported lower satisfaction with the final output, suggesting a trade-off between creativity and productivity. These findings highlight the importance of strategically incorporating "otherness" into the design of human-AI collaboration for creative work.

Viewing/purchase histories, bookmarking (“favorites”), and recommendation lists on digital platforms strongly reflect users’ preferences. Focusing on anonymity in disclosing such preference information and on preferred disclosure targets, we conducted a questionnaire survey across service domains (video, music, e-commerce, and CGM). About half of respondents did not want to create such lists, and even among those who did, willingness to share was limited. When sharing, close targets (family, close friends) were primarily preferred; under anonymous conditions, willingness to disclose increased toward people with shared interests, users of the same service, and the general public. In contrast, offline acquaintances were avoided both with real names and anonymously, more so than online acquaintances. By domain, music showed comparatively low willingness to disclose histories even anonymously, and lower willingness to disclose to close others. We discuss these findings in terms of anonymity in self-disclosure and the psychological/social distance to targets, and present implications for community features that assume preference sharing.