Ayush Singh

I am a final-year undergraduate at Indian Institute of Technology Roorkee, pursuing a B.Tech in Materials Engineering. My research interests primarily lie in the domain of Large Language Models, RLHF/Preference Optimization, AI Safety & Red-Teaming, and Multi-Agent Systems.

I have had the pleasure of working as a Research Intern at Adobe MDSR, where I built agentic pipelines for Core Web Vitals optimization, and as an AI Engineer Intern at Prem AI, where I built a multilingual instruction-following benchmark across all 24 EU languages. I've also worked at Repello AI on AI security and red-teaming, at Boston University on LLM alignment research, and at Lossfunk collaborating with Paras Chopra on novel RLHF techniques.

I am a member of the Vision and Language Group (VLG) at IIT Roorkee, where I contribute to research in deep learning and multi-modal AI.

Email  /  Google Scholar  /  LinkedIn  /  GitHub

My research focuses on Large Language Models, RLHF techniques, and efficient fine-tuning methods. Papers with highlighted backgrounds are selected/notable works.

Introduces a tree-informativeness score combining leaf-outcome diversity with policy-reward decorrelation to filter uninformative rollout trees in tree-based RLVR at zero extra compute, reaching 37.5% macro accuracy on Qwen2.5-Math-1.5B across four math benchmarks.

Empirical evaluation of LLM self-awareness across Base, SFT, DPO, and GRPO training methods on 5 behavioral tasks. Finds GRPO lifts bias-induction self-awareness from 3% to 50.5% accuracy but shows the weakest reasoning-output alignment on OOD prompts.

Novel RLHF alternative using LLMs as preference classifiers, reducing dependence on external reward models. Demonstrates self-improving capabilities on RewardBench across Mistral-7B and Llama-1B models with significant improvements.

Efficient parameter reduction technique achieving 20x fewer trainable parameters than vanilla LoRA. Maintains competitive accuracy on GLUE and WMT16 benchmarks across RoBERTa, BERT, and T5 models.

Multi-agent LLM framework for urban planning optimization integrated with genetic algorithms. Achieves substantial improvements in planning efficiency and balanced city development through collaborative coordination.

A retry-aware LangGraph pipeline across planning, financial data collection (yfinance), and hybrid LLM/Python-REPL analytics stages with human-in-the-loop review gates. Generates audit-ready PDF reports with cited evidence and visualizations, served through a FastAPI backend.

An OpenEnv RL environment where LLM agents design steel building frames, scored by a 3D 6-DOF direct-stiffness solver and Eurocode 3 checks. Three graded tasks from a single-story warehouse to a seismic hospital, served via a Dockerized FastAPI server.

Adapting neural-operator forecasters (CNO, FNO, Transolver) for streaming real-world PIV airfoil-wake flow. A bounded adaptation controller (skip / recalibrate / anchored norm-layer update) with uncertainty bounds and LLM-driven action selection under a fail-safe rule fallback.

Reverse-engineered a dataflow guardrail's predicate logic (exfiltration, destructive-write, confused-deputy, untrusted-to-action) to expose exploitable scoring gaps, and built a search harness replaying multi-step tool-call chains against guardrailed GPT-OSS/Gemma agents.

Benchmarked Gemini, PaliGemma, and GPT-4o on 8 low-level geometric reasoning tasks where VLMs underperform despite strong OCR/VQA results. Proposed a keyword-guided captioning pipeline boosting zero-shot accuracy by up to 14pts without any fine-tuning.

Template adapted from Jon Barron.