Pratham Singla

Pioneered the development of advanced AI benchmarks and contributed to reinforcement learning frameworks for vision-language models at a leading research firm.

• Developed a novel benchmark, inspired by ZeroBench, for complex visual reasoning, integrating an automated question-generation pipeline to train and evaluate vision-language models on GPT-level reasoning tasks.
• Contributed to a reinforcement learning-based training framework, specifically designed to enhance models' ability to ground reasoning in visual inputs, reducing reliance on textual cues.
• Pioneered the creation of a sophisticated evaluation system for next-generation AI models, setting a new standard for visual-language understanding.

Explored and optimized Retrieval-Augmented Generation (RAG) architectures, establishing robust local benchmarking setups for Large Language Models (LLMs).

• Explored and analyzed cutting-edge agentic Retrieval-Augmented Generation (RAG) architectures to optimize information retrieval and synthesis for complex queries.
• Established a robust local benchmarking setup using tools like Ollama, enabling efficient deployment and performance evaluation of Large Language Models (LLMs).

Investigated self-awareness in Large Language Models (LLMs) and evaluated fine-tuning methods to understand model behavior and generalization.

• Investigated self-awareness in Large Language Models (LLMs) through in-depth analysis of alignment between internal reasoning and external outputs.
• Evaluated the performance of SFT, DPO, and GRPO tuned models on tasks involving bias, risk, and reward hacking, providing critical insights into model vulnerabilities.
• Designed and executed comprehensive experiments to rigorously test the awareness of learned behaviors and generalization of reasoning across diverse domains.

Actively participated in ACM chapter activities, contributing to initiatives that advance computer science education and community engagement at IIT Roorkee.

• Collaborated with fellow members to promote computer science and foster a vibrant academic community.
• Contributed to the development and dissemination of technical knowledge through participation in group projects and discussions.
• Supported the chapter's mission to provide resources and opportunities for students interested in computing.

Led project planning and execution, fostered an active Deep Learning community, and moderated discussions within the Vision and Language Group at IIT Roorkee.

• Orchestrated the overall planning and moderation of research projects, contributing to strategic direction and successful project outcomes.
• Organized and moderated paper discussions, enhancing knowledge sharing and collaborative research initiatives within the group.
• Actively contributed to the development of an active Deep Learning community on campus, fostering growth and engagement among students.

Developed an innovative Retrieval-Augmented Generation (RAG) framework focused on enhancing retrieval quality through active learning and efficient fine-tuning.

Evaluated AI agent frameworks on QA tasks to assess their reasoning capabilities, response accuracy, and efficiency.

Engineered a voice-interactive AI interview coach utilizing Gemini LLM and Vapi AI to provide real-time feedback and structured scoring for personalized interview preparation.

Developed a neural network-based solution for enhancing low-light images, achieving significant PSNR improvements through advanced image processing techniques.

Analyzed self-awareness and reasoning alignment in Large Language Models (LLMs) across various scenarios, evaluating the impact of SFT, DPO, and GRPO fine-tuning on model behavior, including bias, risk, and reward hacking.

Developed a Multi-Agent urban planning framework using LLMs and Genetic Algorithms for optimization and regional customization, resulting in significant improvements in livability and accessibility.

Developed a novel image steganography model using a neural network-based approach, integrating attention mechanisms (channel and spatial attention) into an autoencoder architecture.

Conducted a study on LLMs' ability to utilize hints in mathematical problem-solving on the MATH dataset, benchmarking zero-shot, few-shot, chain-of-thought, and adversarial hinting against traditional prompting.