The choice of action spaces is a critical yet unresolved challenge in developing capable, end-to-end trainable agents. This paper first presents a large-scale, systematic comparison of prominent abstracted action spaces and tokenizers for Vision-Language-Action (VLA) or hierarchical agent models in the open-ended Minecraft. Our analysis reveals that no single action space is universally optimal; instead, the most effective abstraction is highly task-dependent, creating a dilemma for building generalist agents. To resolve this, we introduce Chain of Action (CoA), a novel framework that unifies high-level planning and low-level control within a single, monolithic VLA model. CoA treats an abstracted action not as a command for a separate policy, but as an intermediate reasoning step--akin to a chain of thought--that guides the generation of the final, executable action. Furthermore, we demonstrate that an All-in-One agent trained on a diverse mixture of action spaces using the CoA paradigm learns a more robust and generalizable policy. This unified agent achieves a new state-of-the-art, improving the overall task success rate over strong, specialized baselines. To foster reproducible research, we release the OpenHA (Open Hierarchical Agents) suite, which includes our comprehensive benchmark of over 800 distinct tasks, curated datasets, source code, and all pretrained model checkpoints at https://github.com/CraftJarvis/OpenHA

As Large Language Models (LLMs) are increasingly adopted as automated judges in benchmarking and reward modeling, ensuring their reliability, efficiency, and robustness has become critical. In this work, we present a systematic comparison of "thinking" and "non-thinking" LLMs in the LLM-as-a-judge paradigm using open-source Qwen 3 models of relatively small sizes (0.6B, 1.7B, and 4B parameters). We evaluate both accuracy and computational efficiency (FLOPs) on RewardBench tasks, and further examine augmentation strategies for non-thinking models, including in-context learning, rubric-guided judging, reference-based evaluation, and n-best aggregation. Our results show that despite these enhancements, non-thinking models generally fall short of their thinking counterparts. Our results show that thinking models achieve approximately 10% points higher accuracy with little overhead (under 2x), in contrast to augmentation strategies like few-shot learning, which deliver modest gains at a higher cost (>8x). Bias and robustness analyses further demonstrate that thinking models maintain significantly greater consistency under a variety of bias conditions such as positional, bandwagon, identity, diversity, and random biases (6% higher on average). We further extend our experiments to the multilingual setting and our results confirm that explicit reasoning extends its benefits beyond English. Overall, our work results in several important findings that provide systematic evidence that explicit reasoning offers clear advantages in the LLM-as-a-judge paradigm not only in accuracy and efficiency but also in robustness.

In an era of increasing societal fragmentation, political polarization, and erosion of public trust in institutions, representative deliberative assemblies are emerging as a promising democratic forum for developing effective policy outcomes on complex global issues. Despite theoretical attention, there remains limited empirical work that systematically traces how specific ideas evolve, are prioritized, or are discarded during deliberation to form policy recommendations. Addressing these gaps, this work poses two central questions: (1) How might we trace the evolution and distillation of ideas into concrete recommendations within deliberative assemblies? (2) How does the deliberative process shape delegate perspectives and influence voting dynamics over the course of the assembly? To address these questions, we develop LLM-based methodologies for empirically analyzing transcripts from a tech-enhanced in-person deliberative assembly. The framework identifies and visualizes the space of expressed suggestions. We also empirically reconstruct each delegate's evolving perspective throughout the assembly. Our methods contribute novel empirical insights into deliberative processes and demonstrate how LLMs can surface high-resolution dynamics otherwise invisible in traditional assembly outputs.

Time-series forecasting and causal discovery are central in neuroscience, as predicting brain activity and identifying causal relationships between neural populations and circuits can shed light on the mechanisms underlying cognition and disease. With the rise of foundation models, an open question is how they compare to traditional methods for brain signal forecasting and causality analysis, and whether they can be applied in a zero-shot setting. In this work, we evaluate a foundation model against classical methods for inferring directional interactions from spontaneous brain activity measured with functional magnetic resonance imaging (fMRI) in humans. Traditional approaches often rely on Wiener-Granger causality. We tested the forecasting ability of the foundation model in both zero-shot and fine-tuned settings, and assessed causality by comparing Granger-like estimates from the model with standard Granger causality. We validated the approach using synthetic time series generated from ground-truth causal models, including logistic map coupling and Ornstein-Uhlenbeck processes. The foundation model achieved competitive zero-shot forecasting fMRI time series (mean absolute percentage error of 0.55 in controls and 0.27 in patients). Although standard Granger causality did not show clear quantitative differences between models, the foundation model provided a more precise detection of causal interactions. Overall, these findings suggest that foundation models offer versatility, strong zero-shot performance, and potential utility for forecasting and causal discovery in time-series data.

Navigation is a fundamental capability in embodied AI, representing the intelligence required to perceive and interact within physical environments. To achieve such intelligence, recent advanced works leverage Vision-Language Models (VLMs), which demonstrate strong generalizability and possess a well-suited formulation for navigation. However, these approaches remain largely confined to narrow task settings and embodiment-specific architectures. In this work, we introduce a cross-embodiment and cross-task Navigation Foundation Model (Nav-FoM), trained on eight million navigation samples that encompass quadrupeds, drones, wheeled robots, and vehicles, and spanning diverse tasks such as vision-and-language navigation, object searching, target tracking, and autonomous driving. NavFoM employs a unified architecture that processes multimodal navigation inputs from varying camera configurations and navigation horizons. To accommodate diverse camera setups and temporal horizons, NavFoM incorporates identifier tokens that embed camera view information of embodiments and the temporal context of tasks. Furthermore, to meet the demands of real-world deployment, NavFoM controls all observation tokens using a dynamically adjusted sampling strategy under a limited token length budget. Extensive evaluations on seven public benchmarks demonstrate that our model achieves state-of-the-art or highly competitive performance across different navigation tasks and embodiments without requiring task-specific fine-tuning. Additional real-world experiments further confirm the strong generalizability and practical applicability of our approach. Figure 1: We provide an illustration of architecture (left) alongside real-world experiment results (right). The results include cross-task and cross-embodiment performence of our method. See Sec. 4 for more detials. For both embodied agents and humans, navigation serves as a foundational capability that enables them to move intelligently within physical environments to accomplish specified tasks (Shah et al., 2023a; Bar et al., 2025; Zhang et al., 2024b). Achieving robust navigation requires a deep understanding of environmental context and task instructions, typically presented through visual and linguistic observations, which are reminiscent of Visual Language Models (VLMs).

Memory extraction is crucial for maintaining coherent ultra-long dialogues in human-robot role-playing scenarios. However, existing methods often exhibit uncontrolled memory growth. To address this, we propose MOOM, the first dual-branch memory plugin that leverages literary theory by modeling plot development and character portrayal as core storytelling elements. Specifically, one branch summarizes plot conflicts across multiple time scales, while the other extracts the user's character profile. MOOM further integrates a forgetting mechanism, inspired by the ``competition-inhibition'' memory theory, to constrain memory capacity and mitigate uncontrolled growth. Furthermore, we present ZH-4O, a Chinese ultra-long dialogue dataset specifically designed for role-playing, featuring dialogues that average 600 turns and include manually annotated memory information. Experimental results demonstrate that MOOM outperforms all state-of-the-art memory extraction methods, requiring fewer large language model invocations while maintaining a controllable memory capacity.

Recent Vision-Language-Action models show potential to generalize across embodiments but struggle to quickly align with a new robot's action space when high-quality demonstrations are scarce, especially for bipedal humanoids. We present TrajBooster, a cross-embodiment framework that leverages abundant wheeled-humanoid data to boost bipedal VLA. Our key idea is to use end-effector trajectories as a morphology-agnostic interface. TrajBooster (i) extracts 6D dual-arm end-effector trajectories from real-world wheeled humanoids, (ii) retargets them in simulation to Unitree G1 with a whole-body controller trained via a heuristic-enhanced harmonized online DAgger to lift low-dimensional trajectory references into feasible high-dimensional whole-body actions, and (iii) forms heterogeneous triplets that couple source vision/language with target humanoid-compatible actions to post-pre-train a VLA, followed by only 10 minutes of teleoperation data collection on the target humanoid domain. Deployed on Unitree G1, our policy achieves beyond-tabletop household tasks, enabling squatting, cross-height manipulation, and coordinated whole-body motion with markedly improved robustness and generalization. Results show that TrajBooster allows existing wheeled-humanoid data to efficiently strengthen bipedal humanoid VLA performance, reducing reliance on costly same-embodiment data while enhancing action space understanding and zero-shot skill transfer capabilities. For more details, For more details, please refer to our \href{https://jiachengliu3.github.io/TrajBooster/}.

Vision-language-action (VLA) models finetuned from vision-language models (VLMs) hold the promise of leveraging rich pretrained representations to build generalist robots across diverse tasks and environments. However, direct fine-tuning on robot data often disrupts these representations and limits generalization. We present a framework that better preserves pretrained features while adapting them for robot manipulation. Our approach introduces three components: (i) a dual-encoder design with one frozen vision encoder to retain pretrained features and another trainable for task adaptation, (ii) a string-based action tokenizer that casts continuous actions into character sequences aligned with the model's pretraining domain, and (iii) a co-training strategy that combines robot demonstrations with vision-language datasets emphasizing spatial reasoning and affordances. Evaluations in simulation and on real robots show that our method improves robustness to visual perturbations, generalization to novel instructions and environments, and overall task success compared to baselines.

Background: Emerging reports of "AI psychosis" are on the rise, where user-LLM interactions may exacerbate or induce psychosis or adverse psychological symptoms. Whilst the sycophantic and agreeable nature of LLMs can be beneficial, it becomes a vector for harm by reinforcing delusional beliefs in vulnerable users. Methods: Psychosis-bench is a novel benchmark designed to systematically evaluate the psychogenicity of LLMs comprises 16 structured, 12-turn conversational scenarios simulating the progression of delusional themes(Erotic Delusions, Grandiose/Messianic Delusions, Referential Delusions) and potential harms. We evaluated eight prominent LLMs for Delusion Confirmation (DCS), Harm Enablement (HES), and Safety Intervention(SIS) across explicit and implicit conversational contexts. Findings: Across 1,536 simulated conversation turns, all LLMs demonstrated psychogenic potential, showing a strong tendency to perpetuate rather than challenge delusions (mean DCS of 0.91 $\pm$0.88). Models frequently enabled harmful user requests (mean HES of 0.69 $\pm$0.84) and offered safety interventions in only roughly a third of applicable turns (mean SIS of 0.37 $\pm$0.48). 51 / 128 (39.8%) of scenarios had no safety interventions offered. Performance was significantly worse in implicit scenarios, models were more likely to confirm delusions and enable harm while offering fewer interventions (p < .001). A strong correlation was found between DCS and HES (rs = .77). Model performance varied widely, indicating that safety is not an emergent property of scale alone. Conclusion: This study establishes LLM psychogenicity as a quantifiable risk and underscores the urgent need for re-thinking how we train LLMs. We frame this issue not merely as a technical challenge but as a public health imperative requiring collaboration between developers, policymakers, and healthcare professionals.

The behavior of Large Language Models (LLMs) when facing contextual information that conflicts with their internal parametric knowledge is inconsistent, with no generally accepted explanation for the expected outcome distribution. Recent work has identified in autoregressive transformer models a class of neurons -- called entropy neurons -- that produce a significant effect on the model output entropy while having an overall moderate impact on the ranking of the predicted tokens. In this paper, we investigate the preliminary claim that these neurons are involved in inhibiting context copying behavior in transformers by looking at their role in resolving conflicts between contextual and parametric information. We show that entropy neurons are responsible for suppressing context copying across a range of LLMs, and that ablating them leads to a significant change in the generation process. These results enhance our understanding of the internal dynamics of LLMs when handling conflicting information.