International Journal of Science, Engineering and TechnologyAuthors: Associate Professor Dr. Pramod Kumar, Assistant Professor Ms. Ambika Tomar, Assistant Professor Mr. Saood Iqbal khan, Assistant Professor Mr. Lakshya Bhardwaj
Abstract: We explored in this paper the performance of standard Markov chain Monte Carlo (MCMC) algorithms as well as of adaptive versions of these algorithms on a set of representative posterior distributions that can be encountered in variational applications to realistic target distributions. Target distributions were generated by a total variation mixture of Gaussian distributions, and they comprised multimodal distributions like the shifted mixture distribution, the Hartman 6D, the LB 4D and the Rosenbrock 6D. In this simulation study, we primarily examined a few basic algorithms, adaptive random walk Metropolis (RWM), adaptive independent Metropolis (AIM), adaptive Gaussian proposal Metropolis (AGPM), and generalized adaptive Metropolis (GAM), on the target distributions created. We observed that although the RWM based MCMC algorithms were capable of capturing most of the probability mass of the multimodal target distributions and generating a good degree of mixing. This however had a significant negative impact on it due to the choice of covariance structure. Further analyses on the performance of AIM, AGPM and GAM indicated that the GAM could yield desired coverage compared to RWM based MCMC but they are vulnerable to decrease the efficiency of sampling. Furthermore, less dimensional targets were required to get a good performance with the AIM than with the low dimensional tests. Overall, our simulation. This paper presents the doctrinal and technical problems that come with using LLM-generated content as an offered, challenged, or investigated piece of evidence. It discusses the existing framework of evidentiary procedures in the US, India and the European Union, and existing incidents of AI hallucination and machine-generated disinformation to bring to the fore the practical implications of the problem. The paper proposes a modified IDAP framework for identification, preservation, acquisition, analysis, and presentation, drawing on the probabilistic, non-deterministic, and cloud-based nature of large language model outputs, in light of existing digital forensic methodologies. Moreover, it argues that there should be a legislative framework that forces the logging of provenance and ensures that records produced by AI are qualified and that cross-border competence requirements for admissibility are harmonised. If there is no attempt at harmonization between the science of Forensics and the evidentiary law, then either AI-generated evidence will be discounted or bogus evidence will be accepted without question. The framework provides a rationale and a scientifically informed guide for legislatures, forensic practitioners, and the courts in a principled approach to this new type of evidence.
