Vikas Yadav

AI Researcher · LLMs · Agentic Reasoning · Multilingual AI

My research focuses on building AI systems that are explainable, robust, and scalable. I build agentic AI systems that reason, plan, and collaborate across long-horizon tasks — designing multi-agent architectures with composable reasoning modules, grammar-based coordination protocols, and scalable frameworks that let autonomous agents tackle complex, multi-step problems. My work also focuses on creating rigorous benchmarks to systematically evaluate LLMs and VLMs, probing their capabilities and surfacing critical failure modes.

I am especially interested in specialized agentic systems for vertical AI domains — bringing long-horizon reasoning and multi-agent collaboration to industry-specific challenges where domain expertise, reliability, and safety are paramount.

My recent research has spanned the full LLM and VLM lifecycle — continual pre-training, post-training (DPO variants, safety, abstention), and efficient inference (speculative decoding, dynamic layer slicing, layer-wise quantization) — actively publishing at top-tier conferences including ACL, EMNLP, ICLR, NAACL, AAAI, SIGIR, and COLM. I received my Ph.D. from the University of Arizona and have spent over 6 years in industry research, building and deploying large-scale AI systems across diverse real-world applications.

Get in Touch

Research Focus

My work spans the full LLM lifecycle — from continual pre-training to post-training alignment — with a focus on building scalable, efficient, and aligned AI systems.

Deep Reasoning & Planning

Long-horizon reasoning, chain-of-thought optimization, and multi-step planning systems. Developing benchmarks for evaluating and mitigating over-reasoning in LLMs.

Multi-Agent Systems

Autonomous, scalable agentic systems for complex multi-environment tasks. Grammar-based search and composable reasoning modules for generalizable agent collaboration.

LLM Efficiency & Compression

Speculative decoding, dynamic layer slicing, and layer-wise quantization techniques for accelerating LLM inference while preserving quality.

Alignment & Safety

Preference alignment via DPO variants, adversarial robustness of reasoning models, abstention capabilities, and defense against prompt injection attacks.

Multilingual & Cross-Lingual AI

Enhancing LLM capabilities across non-Latin scripts through phonemic prompting, multilingual instruction alignment, and cross-lingual transfer learning.

Multimodal & Vision-Language

Chart reasoning with visual reinforcement, RAG-powered document QA, and multi-modal dialogue systems integrating vision and language understanding.