Divyanshu Goyal
Applied Scientist at Adobe
Applied Scientist at Adobe and multimodal researcher focused on training and evaluating vision-language models. I hold an M.S. in Computer Science (Machine Learning) from Georgia Tech. My work spans VLM training, multimodal reasoning, and LLM pretraining,turning frontier research into systems that ship to production.
About Me
Education
M.S. in Computer Science
Georgia Institute of Technology
Machine Learning
B.E. in Computer Science
Birla Institute of Technology and Sciences, Pilani
M.Sc. in Mathematics
Birla Institute of Technology and Sciences, 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
Vision-Language Model Training & Evaluation
- Led ATune (under review): a 54M-parameter model that learns photo retouching from 10M unedited stock photographs and drives off-the-shelf editors through their sliders. In a study with 157 raters, its edits were preferred over GPT-6 Astra's in 54.3% of votes and tied with Claude Opus 5's (50.5%).
- Fine-tuned Gemma-3-12B with LoRA on 8× H100 (DDP, BF16, FlashAttention-2, Liger kernels) to assess the quality of marketing images.
- Grounded VLM quality reasoning in 14+ computed pixel-level metrics (sharpness, exposure, color balance); the resulting interpretable scorer reaches MAE 0.77 (Pearson r ≈ 0.88) vs. 1.34 for zero-shot GPT-5.4 and 1.95 for Claude Opus 4.7.
- Built the MQCore evaluation benchmark and a fault-tolerant async annotation pipeline (100+ images/min) for the 700K-image KADIS image-quality dataset.
Evaluation Research & Benchmarking
- First author of DistortBench (CVPR 2026 and ICML 2026 Workshops): a 13.5K-question benchmark over 27 distortion types and 5 severity levels. Of 18 VLMs evaluated, the best (Qwen3.5-27B) scores 61.9% (chance 25%), above the average individual annotator (60.2%) but below a 3-annotator majority vote (65.7%).
- Co-authored JudgeBreak (ICML 2026 Workshop): across 95 subjective evaluation settings, linear probes on Qwen3-32B hidden states reach 78.1% mean pairwise accuracy vs. 70.1% for the strongest baseline (the best reward model or a zero-shot GPT-5.4 judge).
- Co-authored a curiosity-driven LLM-as-a-judge that personalizes creative-writing evaluation using an intrinsic curiosity model (AAAI 2026 Workshop, Oral), raising mean F1 from 0.37 to 0.62 and Pearson r from 0.15 to 0.56 over an SFT baseline.
Production Multimodal Systems & Agents
- Built the founding prototype of Adobe GenStudio for Performance Marketing (launched October 2024, now serving Fortune 500 clients including Microsoft and AT&T) and architected its multimodal brand service (brand-DNA extraction, retrieval, VLM content validation) across 3 regions.
- Built a multimodal Photoshop agent that edits images from natural language through a 49-tool observe, think, and act loop with vision grounding and a quality-critic refinement step.
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:Training GPT-2 Scale LLMs 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 Model
Built a complete MLOps system for training vision-language models to assess marketing content quality, featuring novel observable injection and Photoshop edit evaluation capabilities.
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.