In this paper, basic steps of scientific discovery literature analysis, idea generation, experiment design and paper drafting are discussed as the core problem for the automation of scientific discovery systems via AI. The current deployments of large language models (LLMs) enable each task to be individually executed – they don't currently have the infrastructure to have multiple agents reason cogently and cooperatively as a whole for an end-to-end autonomous research workflow. This paper introduces an innovative, structured multi-agent system, the Hierarchical Autonomous Research Agent (HARA), in which six specialized autonomous agents are coordinated by a central orchestrator, and both use a common semantic memory and a dynamically expanding knowledge graph. The HARA framework rigorously defines three new scoring functions: Novelty Score N_s indicates the importance of the research gap, Hypothesis Confidence H_c indicates the quality of the proposed hypotheses generated, and Research Utility U indicates a composite function from novelty, relevance, and computer cost of the proposed hypotheses generated. HARA outperform all four metrics: 97% novelty detection accuracy, 96% hypothesis accuracy, 98% draft quality and 96% validation score on four metrics against the Standard LLM, RAG baselines, and Multi-Agent baselines on four scientific domains across arXiv metadata, Semantic Scholar and PubMed corpora. After all, the ablation study demonstrates that no single agent contributes more than the sum of its parts: namely, the contribution of the gap detection agent in novelty detection is largest: 5 percentage points.

Keywords : Agentic AI, Hypothesis Generation, LLM, Multi-Agent Frameworks, Scientific Discovery.

Authors : Uma Maheswari Krishnamoorthy

Title : Hierarchical Autonomous Research Agent for Scientific Hypothesis Generation, Experiment Design, and Paper Drafting

Volume/Issue : 2025;2(4 ( October - December ))

Page No : 31 - 41

Embodied Agentic AI systems transcend the capabilities of language-based intelligence systems by incorporating the capacity for sense making in dynamic environments, reasoning over multi-modal representations of agents' context, and taking action sequences in real-time to achieve goals. Current methods have three serious drawbacks: perceptual modules are deployed independent of the planning stack, memories aren't as layered to perform long-horizon task reasoning, and action pipelines cannot adjust to changes in the environment without a complete restart of the task. A unified Embodied Agentic AI framework consisting of five morphologically tightly coupled functional layers, including a cross-modal attention-based fusion engine, a hierarchical reasoning agent that performs a decomposition into a chain of thoughts, a four-level memory hierarchy, from a short-term memory store to a long-term memory store via episodic and semantic stores, and an adaptive action module with corrective feedback functionality are proposed. It embeds a composite reward function R(s,a) = ?P + ?A ? ?C to optimize task performance, action efficiency and environmental costs, along with a state transition function S(t+1) = f(S(t), A(t), E(t)) that models the dynamic environmental coupling. It is evaluated under three standard embodied AI tasks—ALFRED, Habitat and RoboTHOR and achieves task success rates of 98%, action accuracy of 96%, 95% memory retention, and adaptability score of 97%, improving by 17%, 20%, 25% and 24 percentage points over a baseline pipeline method, respectively. The results indicate that full integration of embodied agentic architectures is viable to deploy autonomous tasks in the real-world.

Keywords : Embodied AI, Hierarchical Reasoning, Memory-Augmented Networks, Sensor Fusion, Chain-of-Thought.

Authors : Ujval Mutyala

Title : Embodied Agentic AI for Autonomous Task Planning and Execution in Dynamic Environments

Volume/Issue : 2025;2(4 ( October - December ))

Page No : 22 - 30

The previous accepted convention of largo Language Model (LLM) agents used for autonomous task execution has been that they are static systems; their behaviour when used at inference time is fixed and does not depend on feedback signals they receive during run time. The paper introduces a reinforcement learning (RL) approach to designing a self-improving LLM agent framework that allows agents to selectively improve action selection policies, rather than fine-tuning LLM parameters, by accepting scalar rewards based on the success of their action. The proposed architecture combines a GPT-4/LLaMA task planner with an episodic memory module that looks up previously seen state–action pairs, and an actor-critic policy network trained using proximal policy optimisation (PPO) to concurrently optimise for task accuracy, execution efficiency, and operational cost, represented by a composite reward function R(s,a) = ?A + ?E ? ?C. The framework is tested against a baseline static agent and a memory-augmented agent against three benchmarks: HumanEval, GSM8K, and HotpotQA. Experimental results show the RL agent attains 97% task accuracy, 95% learning efficiency, and 1,650 ms prediction latency on average, due to which 15% high and 20% high improvements are achieved over the standard agent, and prediction latency stays in acceptable range for the conversations when compared with the humans. Learning convergence is reached after 10-12 training episodes, while the policy updater of PPO is stable along the training trajectory due to its KL-divergence (KL-D) clipping mechanism. This paper demonstrates that reinforcement learning is a tractable method for the continual, autonomous optimization of deployed LLM agents..

Keywords : Self-Improving Agents, Reward Shaping, Actor-Critic Networks, Episodic Memory, Q-Learning.

Authors : Ravi Kumar Kottala

Title : Reinforcement Learning-Based Self-Improving LLM Agents for Autonomous Task Optimization

Volume/Issue : 2025;2(4 ( October - December ))

Page No : 14 - 21

Large Language Models (LLMs) are shown to have a strong ability, on various fields of work related to natural language understanding and generation. But their ability to generate hallucinated content – plausible but factually incorrect answers – continues to be a major challenge in high-stakes applications like medicine, law and science. We suggest the Retrieval-Augmented Multi-Agent Architecture (RAMA), an architecture based on collaboration among five specialized agents, with structured communication protocols, each capable of grounding the output of the model. RAMA is a combination of dense vector retrieval for a dynamically updated knowledge base, a graph-augmented reasoning module and a confidence-weighted verification pipeline. Evaluation using HotpotQA, TriviaQA, and Natural Questions benchmarks shows that they hallucinate with a rate of 4%, demonstrate accuracy in question-answering of 97% and AUC of 0.97 (a gain of 14, 13, and 13 percentage points compared to the standard LLM, RAG, and Single-Agent RAG baseline respectively) with a 1700 ms inference latency.

Keywords : LLM, Retrieval-Augmented Generation, Multi-Agent Systems, Vector Databases, Factual Verification, NLP.

Authors : Ravi Samikannu

Title : Retrieval-Augmented Multi-Agent Architecture for Hallucination Reduction in Large Language Models

Volume/Issue : 2025;2(4 ( October - December ))

Page No : 7 - 13

Recent developments of deep learning have enabled important progress in brain-computer interfaces and electroencephalogram-based (EEG) classifications of neural disorders. Progress has been made greatly in brain-computer interfaces and particularly in classification of a neural disorder using its electroencephalogram (EEG) thanks to the use of deep learning. Nevertheless, the Euclidean distance metric used in conventional architectures is based on the covariance-based EEG features that by their nature are defined on a curved, Symmetric Positive Definite (SPD) manifold, thus causing geometric distortion and ultimately compromising the classification accuracy. He proposed a new framework to combine Graph Neural Networks (GNN) with Riemannian geometry in relation to the inter-channel connectivity between brain regions, which has an intrinsic manifold structure. EEG signals are decomposed into five frequency subbands and covariance matrices for each frequency band are computed and projected to the SPD manifold. The functional connectivity is coded as weighted graph edges by Phase Locking Value (PLV) derived adjacency matrices which are then used by a two-layer Graph Convolutional Network (GCN) for the node level message passing. The features from the Riemannian tangent space and the GCN embedding are combined and trained together, using a combined classification loss and connectivity loss. The proposed model achieves the best performance with 98.3% accuracy, with 98.3% accuracy, 0.994 AUC, and macro F1 score of 0.981, outperforming CNN (90.4%), Graph CNN (93.2%) and Transformer (96.1%).

Keywords : Graph Neural Networks, Riemannian Geometry, SPD Manifold, Brain Connectivity, Phase Locking Value.

Authors : Sivaram Murugan

Title : Graph Neural Networks with Riemannian Brain Connectivity Analysis for EEG-Based Neurological Classification

Volume/Issue : 2025;2(4 ( October - December ))

Page No : 1 - 6