construct behavioral priors

This work addresses the challenge of data scarcity in robot learning by providing a systematic survey of recent advances in transferring manipulation skills from human videos. It introduces the first hierarchical taxonomy tailored to robotic skill acquisition, integrating human-to-robot transfer pathways, data configurations, and learning paradigms across three levels: task, observation, and action. The study further presents a large-scale statistical analysis of existing video datasets, characterizing their scale, structure, and usage trends. By synthesizing developments in policy learning, computer vision, generative modeling, and cross-paradigm coupling methods, this survey comprehensively maps the current landscape, identifies key challenges and limitations, and outlines promising future directions. To foster community progress, the authors also release an open-source collection of relevant papers.

Existing behavioral modeling approaches treat user actions as discrete event sequences, neglecting the contextual information embedded in inter-action time intervals—leading to incomplete behavioral understanding and poor interpretability. To address this, we propose the dual-scale Action-Timing Context (ATC) framework, the first to systematically model *inter-action temporal context* by jointly embedding action types and time intervals within a unified representation space, thereby capturing fine-grained temporal structure. ATC employs dual-scale temporal embedding and low-dimensional action representation learning to yield interpretable and consistent behavioral embeddings. Extensive experiments on multiple real-world digital platform log datasets demonstrate that ATC significantly improves performance in behavioral prediction, post-hoc interpretability, and analysis of sociological mechanisms—including knowledge accumulation and information diffusion—thereby filling a critical gap in temporal structural modeling of human behavior.

This work addresses a fundamental limitation in traditional behavioral measurement, which often relies on passive observation under static or weakly controlled conditions and struggles to disentangle distinct internal mechanisms that produce similar overt behaviors. Treating human behavior as the observable output of a dynamic system, this study introduces— for the first time—the principles of system identification into behavioral science. It proposes a closed-loop experimental framework based on structured perturbations: precise, programmable disturbances are delivered via immersive environments while multimodal behavioral trajectories are simultaneously recorded. These data are integrated with dynamic computational models to enable mechanism-driven, real-time inference. By synergistically combining psychometrics, experimental design, and generative modeling, the approach advances behavioral science from descriptive analysis toward an identifiable, reproducible paradigm centered on generative mechanisms, substantially enhancing both theoretical rigor and causal interpretability in behavioral inference.

Existing AI systems struggle to model the collective dynamics that govern group stability or abrupt transitions, often limited to individual behaviors or post-hoc analyses. This work introduces, for the first time, a formal integration of complex systems theory into group behavior modeling through a novel “behavioral field” framework. It defines structural constructs such as behavioral field bases, tension fields, and critical exponents, enabling unified representation across diverse scenarios. The approach combines graph neural networks with kinematic micro-signals—including position, velocity, and posture—to construct directed interaction graphs that learn the evolving state of groups in real time. Validated in a seven-person negotiation setting, the model demonstrates strong real-time predictive capability, and the calibrated behavioral field generalizes effectively to applications in crowd safety, crisis response, education, and clinical contexts.

This work addresses the challenge of accurately modeling population-level behavioral distributions—such as adoption, churn, and hesitation—under given enterprise decisions. The authors propose Posterior Twins, a memory-anchored digital twin framework that jointly optimizes distributional fidelity and modality accuracy, enabling a reusable evidence system for enterprise decision-making. The approach integrates governance-aware memory, behavior-model routing, and scenario orchestration, with reliability ensured through Wasserstein-1 distance evaluation and a distributional auditing mechanism. Evaluated on a benchmark of 226 samples, TL-Twin Alpha achieves the lowest Wasserstein-1 distance (1.16), while TL-Twin Delta and Gamma demonstrate balanced performance on the Pareto frontier of modality accuracy.

Reinforcement learning policies often exhibit abrupt shifts, oscillations, or degenerate stagnation due to a lack of temporal coherence. To address this, this work proposes Dynamic Prior Reinforcement Learning (DP-RL), which introduces an auxiliary loss derived from external state dynamics—without altering the reward function, environment, or policy architecture—to explicitly incorporate a dynamic prior that embodies evidence accumulation and hysteresis effects as part of the training objective. By solely modifying the optimization target, DP-RL enables control over the temporal geometric properties of an agent’s decision trajectory. Implemented within a policy gradient framework, DP-RL is validated in three minimal environments, demonstrating its ability to systematically shape behaviors with structured temporal characteristics in a task-dependent manner, surpassing the capabilities of generic smoothing approaches.

This work addresses the confounding influence of narrative framing on large language model (LLM) agent behavior, which is often conflated with role prompting. Introducing the novel concept of “narrative priors,” the study employs structurally isomorphic but narratively distinct text-based investigation games to systematically disentangle and quantify the behavioral impact of narrative. Through causal interventions, cross-model variance decomposition, and generalization tests, the authors demonstrate that narrative priors account for 5–31 times more behavioral variance than role specifications alone. They further identify behavioral anchoring as a key mechanism enabling cross-narrative transfer, showing that its removal reduces behavioral consistency by 95%. Leveraging these insights, the proposed role selection method substantially enhances cross-narrative generalization performance.

This study addresses the longstanding challenge of making long-term, personalized prospective predictions of individuals’ future speech behaviors from everyday conversations—a capability essential for timely detection of goal-deviant or potentially inappropriate utterances. For the first time, the authors demonstrate that large language models (LLMs) trained on over 1,000 hours of real-world longitudinal dialogue data collected via wearable devices can effectively forecast specific verbal behaviors in subsequent interactions. Through semi-structured user interviews, the research not only validates the efficacy of LLMs in personalized behavioral prediction but also uncovers their practical potential for proactive behavioral support and intervention. These findings establish a novel paradigm for developing personalized AI systems endowed with predictive foresight.