Divyanshu Goyal
Applied Scientist at Adobe
Applied scientist at Adobe, passionate about multimodal large language models. Over the past 5 years I have trained, aligned, and evaluated these models in highly subjective domains such as brand marketing and photo editing/retouching. I have contributed my work at CVPR, ICML, and AAAI workshops, hold patents in generative AI space, and write about AI through blogs. I am strong believer in first principle thinking and try to bring that to everything I do. I hold an M.S. in Computer Science (Machine Learning) from Georgia Tech.
About Me
Education
M.S. in Computer Science
Georgia Institute of Technology
Machine Learning
B.E. in Computer Science
Birla Institute of Technology and Science, Pilani
M.Sc. in Mathematics
Birla Institute of Technology and Science, Pilani
Expertise
Tools & Languages
Publications & Patents
DistortBench: Benchmarking Vision Language Models on Image Distortion Identification
Divyanshu Goyal, Akhil Eppa, Vanya Bannihatti Kumar
JudgeBreak: Is Taste Hidden in the Layers? Stress-Testing Reward Models and LLM Judges with Representation-Based Subjective Evaluation
Vanya Bannihatti Kumar, Anish Narang, Mayank Anand, Akhil Eppa, Divyanshu Goyal
Curiosity-Driven LLM-as-a-judge for Personalized Creative Judgment
Vanya Bannihatti Kumar, Divyanshu Goyal, Akhil Eppa, Neel Bhandari
Cultural Adaptation of Images Using Generative Artificial Intelligence
Akhil Eppa, Mayank Anand, Divyanshu Goyal
Guiding Language Translation with Translation Documents Using Machine Learning
Divyanshu Goyal, Akhil Eppa, Mayank Anand
Systems and Methods to Provide Parameter-Efficient Fine-Tuned Models
Yuanyou Wang, Naveen Vangala, Mayank Anand, Kunal Kumar Jain, Jose Mathew, Eapen Jose, Divyanshu Goyal, Asmita Chihnara, Arif Abdullah, Anand Dantu
Brand-Aligned Marketing Content Generation Using Structured Brand Data and Generative Models
Divyanshu Goyal, Mayank Anand, Jose Mathew
Experience
Applied Scientist
/Adobe, San Jose, CAJune 2021 – Present
VLM Fine-Tuning with SFT and Reinforcement Learning (Gemma-3-12B)
- Trained Gemma-3-12B (LoRA rank 256, 1.19B trainable parameters) to assess marketing-image quality with distributed SFT, then multi-round GRPO, on 8× H100. Each round samples a group of completions per prompt, scores them with programmatic rewards against ground truth, and updates the policy with a clipped policy-gradient loss. Diagnosed reward hacking (collapse onto hedged, low-information answers) and redesigned the reward scorers.
Multimodal Judge Models
- Built a 1.1M-row training set from 14 human-rated sources with rater disagreement as soft labels, and trained Qwen3.5-4B LoRA judges on mixes sampled from it. After calibrating its score output on crowd ratings, the served judge raised validation SRCC on KonIQ from 0.421 (reference model) to 0.663 (paired 95% CI above 0), measured on a leak-checked 67K-item held-out benchmark suite I built.
Unsupervised Photo Retouching (ATune)
- Led ATune (co-first author, under review): a 54M-parameter model that learns photo retouching from 10M unedited photographs with no expert edits and operates off-the-shelf editors by setting their sliders. With 157 raters, its edits beat GPT-6 Astra using the same editor (54.3% of decisive votes, 95% CI 50.3–58.3; 3,524 answers) and tied with Claude Opus 5 (50.5%, CI 46.8–54.2). On FiveK, PPR10K, and SPAQ, at least 8 of 9 published quality judges prefer it to GPT-6 Astra.
Evaluation Research
- Created DistortBench (first author; CVPR 2026 and ICML 2026 Workshops): 13.5K questions over 27 distortion types and 5 severity levels. Of 18 VLMs, the best (Qwen3.5-27B) scores 61.9% (chance 25%), above the average annotator (60.2%) and below a 3-annotator majority vote (65.7%).
- Co-authored JudgeBreak (ICML 2026 Workshop): linear probes on Qwen3-32B hidden states reach 78.1% pairwise accuracy across 95 subjective settings, vs. 70.1% for the best reward model or LLM judge.
Product
- Prototyped Adobe GenStudio for Performance Marketing (launched October 2024, serving Fortune 500 clients) and built a multimodal Photoshop agent that edits images with vision grounding.
Machine Learning Engineer
/Adobe, Bangalore, IndiaJuly 2016 – August 2019
ML Infrastructure & Serving
- Co-engineered Adobe's ML inference platform: 3,500 QPS/node at 0.3–0.6 ms app-side p99 latency.
- Designed GC-free async event pipeline using LMAX Disruptor, eliminating back-pressure under peak load.
- Built custom model lifecycle management platform: versioned deployment, monitoring, and staged rollout across environments.
Projects
VibeNanoChat: GPT-2 Scale LLM Pretraining from Scratch
A from-scratch framework for pretraining, scaling, and evaluating GPT-2 scale language models, featuring depth-parameterized scaling laws, a custom distributed Muon optimizer, and a 35+ benchmark evaluation suite.
Marketing Quality Vision-Language Models
Fine-tuned Gemma-3-12B with SFT and multi-round GRPO, then trained Qwen3.5-4B judges for marketing attributes, checked on a leak-checked benchmark suite.
Adobe GenStudio for Performance Marketing
Built the founding prototype and pitched it to Adobe leadership, leading to the launch of a generative-AI platform for creating on-brand marketing content at scale.