Autonomous Research Agent for Multi-Hop Reasoning

This project introduces a highly advanced autonomous AI agent built upon the Gemma-4-E4B-it multimodal foundation model. Optimized using Generative Reward Policy Optimization (GRPO) Reinforcement Learning on high-performance AMD MI300X infrastructure, the agent is specifically engineered to perform deep, multi-hop reasoning tasks. Unlike traditional static LLMs, this agent employs a hierarchical Parallel-Agent Reinforcement Learning (PARL) architecture. This allows it to dynamically delegate sub-tasks, interact with live web environments via Playwright, execute Python code for data analysis, and overcome complex research hurdles autonomously. A key technical achievement of this project is the custom training pipeline, which carefully freezes the native vision encoder to preserve full multimodal capabilities while aggressively fine-tuning the model's reasoning and reporting skills. Equipped with a massive 60,000-token context window, intelligent memory compression, and robust anti-reward-hacking mechanisms, the agent excels at synthesizing scattered, unstructured data into highly detailed, well-structured HTML reports. This makes it a state-of-the-art solution for automated deep-dive research, intelligence gathering, and complex problem-solving.