episodic observation reconstruction (recover briefly-seen items)

Event reconstruction in digital forensics suffers from fragmented perspectives, inconsistent terminology, and methodological fragmentation, lacking a systematic, unifying framework. Method: This paper proposes the first unified temporal event reconstruction framework tailored for digital forensics—adapting classical forensic reconstruction models to the digital domain; constructing a comprehensive, lifecycle-spanning conceptual map of temporal reconstruction; and conducting a systematic literature review (SLR) coupled with conceptual modeling to clarify terminological relationships and process elements. Contribution/Results: The study identifies three core challenges—data scale, temporal distortion, and semantic ambiguity—and establishes an extensible classification system. It delivers a consensus-based terminology set and a standardized process paradigm, thereby providing a rigorous theoretical foundation for the development and evaluation of automated event reconstruction tools.

This paper investigates the fundamental theoretical limits of reconstructing the true interaction structure of complex networks from observational data. It introduces “reconstructibility”—the proportion of structural information recoverable from data—and provides the first rigorous information-theoretic definition. We derive a universal upper bound on the average reconstruction performance of any algorithm, proving it is fully determined by the true data-generating (TDG) model. To enable empirical assessment, we propose a computationally tractable “reconstruction index” as an estimator of reconstructibility, validating its efficacy via Jaccard similarity and error probability modeling. Crucially, we demonstrate that model selection fundamentally governs reconstruction performance. This work establishes the first theoretical benchmark for network reconstruction, revealing intrinsic constraints on structural recoverability and providing principled guidance for method evaluation and model design.

This study addresses the problem of recovering latent discrete states from the evolving weights of models trained on time-varying data streams to characterize non-stationary distributional shifts. The proposed approach trains classifiers over sliding time windows, aligns their weight trajectories, and fits a hidden Markov model (HMM) to these trajectories—enabling, for the first time, the identification of semantically coherent temporal phases solely from weight dynamics. Experiments on the Fakeddit and Yelp datasets demonstrate that transfer performance within the same inferred state significantly outperforms cross-state transfer, and this advantage persists independently of temporal proximity and shifts in class distribution. These findings confirm that model weights encode structural information about data distributions that extends beyond local temporal correlations.

This paper addresses the problem of sequential anomaly detection in a multi-process dynamic system: normal processes remain perpetually in a zero (quiescent) state, whereas anomalous processes evolve their latent states over time according to a Markov chain; observations are obtained only by sequentially probing a subset of processes, and each probe’s outcome depends stochastically on the probed process’s current latent state. Departing from conventional i.i.d. observation assumptions, we introduce, for the first time, a hidden Markov model (HMM) into this sequential search framework. We propose ADHM—an adaptive probing algorithm that jointly models latent-state evolution and observation uncertainty via Bayesian belief updating and statistical evidence accumulation. We establish its asymptotic optimality and derive a fundamental oracle lower bound on detection delay. Simulation results demonstrate that, under strict false-alarm probability constraints, ADHM reduces the average detection time by 32% compared to state-of-the-art methods.

This work investigates how frozen vision-language-action (VLA) models represent and utilize visual history without relying on explicit memory mechanisms. Introducing, for the first time, a training-free temporal deployment auditing framework—comprising hierarchical linear probing, causal swapping interventions, history re-injection, and action prediction analysis—the study systematically compares the historical usage strategies across three VLA architectures. The findings reveal that historical information typically serves as a redundant copy of the current frame and is only invoked under severely degraded input conditions. Moreover, distinct architectures exhibit markedly different patterns of dependence on history, and the manipulability of historical information is determined by its deployment strategy rather than whether it is encoded.

This study addresses the challenge of modeling and reusing recent historical information in longitudinal athlete monitoring by introducing the “latent memory table” as a novel analytical unit. The approach employs a Transformer-based memory operator to map masked temporal windows into finite-dimensional states, which are aggregated into a statistical table amenable to storage, querying, and reuse. This framework unifies and generalizes classical techniques such as moving averages and principal component analysis, while emphasizing six key quality attributes—including restorability, personalization, and temporal consistency—and incorporates uncertainty quantification, Procrustes-based row reliability ensembles, and a composite quality index Q for evaluation. In the SoccerMon case study, the latent memory table achieves a quality index of 0.73, substantially outperforming conventional methods (approximately 0.40) and demonstrating incremental predictive value for certain health indicators.

This work addresses the lack of direct evaluation of long-term memory content in large language model agents, which currently rely solely on downstream behavioral proxies that hinder auditability of retained user states. The authors propose treating long-term memory as an auditable artifact by directly assessing its quality through reconstruction of latent, structured user states. To this end, they introduce MEMPROBE—the first benchmark for memory recovery capability—featuring a synthetic ground-truth repository of hidden user states, simulated user trajectories, controlled information leakage tasks, and balanced state dimensions. Evaluations under both full-storage and top-k retrieval settings reveal a significant gap between task completion performance and memory fidelity: despite high task success rates across 50 users and 1,550 targets, memory recovery rates hover around 0.6 and further decline under top-k retrieval, underscoring a critical disconnect between functional assistance and faithful memory retention.