Anindya Mondal, PhD
Multimodal Foundation Models · Generative AI · RL Post-Training
I build multimodal foundation models that are reliable and controllable, not just plausible. My research sits where vision-language understanding meets generative AI: unified models that can both perceive and generate, and that stay faithful to what they are asked instead of hallucinating.
Over the past four years I have worked across that problem end to end: RL post-training for alignment (GRPO, process reward models), test-time reasoning with agentic planner-critic loops that steer frozen diffusion models, VLM judges and reward models for evaluating generated content, and open-vocabulary, training-free perception across images and video. This has led to 7 peer-reviewed papers, including first-author work at ACM TOG (SIGGRAPH Asia), TMLR, and AAAI.
I recently completed my Ph.D. in Artificial Intelligence at CVSSP, University of Surrey, advised by Dr. Anjan Dutta, Dr. Xiatian Zhu, and Dr. Joaquin M. Prada, in close collaboration with Dr. Sauradip Nag. Most recently, I was a Research Scientist Intern at Adobe Research.
Email /
CV /
Scholar /
GitHub /
LinkedIn
|
|
On the job market
I defended my Ph.D. in October 2026 and am looking for Research Scientist and Postdoctoral positions in multimodal foundation models, generative AI, RL post-training and alignment, and agentic systems.
Get in touch
Download CV (PDF)
Publications
News
- Oct 2026: New Successfully defended my Ph.D. thesis at the University of Surrey. Now open to Research Scientist and Postdoctoral roles.
- Aug 2026: CountLoop, on agent-guided image generation, accepted at TMLR.
- Jul 2026: ABACUS, on unified count understanding and generation, accepted as an ACM TOG journal paper at SIGGRAPH Asia 2026.
- Jun 2026: Joined Adobe Research as a Research Scientist Intern.
- Jan 2025: OmniCount, on multi-label object counting, accepted at AAAI 2025.
- Dec 2024: Awarded AAAI 2025 Conference Travel Grant ($1,200).
- Oct 2023: Presented actor-agnostic action recognition work at ICCV Workshop 2023 in Paris.
- Oct 2022: Started Ph.D. at the University of Surrey on a fully funded doctoral studentship.
|
|
Adobe Research, Bengaluru, India
Position: Research Scientist Intern (Multimodal Alignment)
GRPO post-training with criterion-level process reward models, and rubric-aware VLM judges that break generative prompts into verifiable criteria.
Jun 2026 - Sep 2026
|
|
University of Surrey (CVSSP), Guildford, UK
Position: Doctoral Researcher
Inference-time agentic reasoning (CountLoop), cycle-consistent GRPO alignment (ABACUS), and training-free open-vocabulary counting (OmniCount) for unified vision-language and generative models.
Oct 2022 - Oct 2026
|
|
Indian Institute of Science, Bengaluru, India
Position: Research Intern
Developed a source-free domain-adaptation framework for image classification.
May 2022 - Aug 2022
|
|
Jadavpur University, Kolkata, India
Position: Undergraduate Research Assistant
Graph learning and event-based vision for signal recovery and moving-object detection.
Oct 2020 - May 2022
|
Research
My work asks how to make multimodal foundation models faithful: able to understand and generate visual content that matches the instruction, and to be checked when it does not. Counting and spatial composition are my main testbed, because they are precise, easy to verify, and still hard for current models. Four threads run through it:
- Unified understanding and generation: a single VLM that both perceives and synthesises, aligned with a cycle-consistent GRPO objective (ABACUS).
- Test-time reasoning and agents: training-free planner-critic loops that spend inference compute to steer frozen diffusion models (CountLoop).
- Reward modeling and evaluation: criterion-level process rewards and rubric-aware VLM judges for generated content (ABACUS, internship work at Adobe Research).
- Open-vocabulary perception: training-free grounding and recognition from foundation-model priors, in images and video (OmniCount, MSQNet).
Representative papers are highlighted. Full list on Google Scholar.
|
|
ABACUS: Adapting Unified Foundation Model for Bridging Image Count Understanding and Generation
Anindya Mondal,
Sauradip Nag,
Anjan Dutta
ACM Transactions on Graphics/ SIGGRAPH ASIA, 2026
PDF /
arXiv /
web
TL;DR▼
Single unified 3B-parameter VLM that jointly solves object counting, crowd counting, referring-expression counting, and count-faithful image generation via density-aware adaptive zooming, MLLM attention-derived objectness maps, and a cycle-consistent GRPO strategy with nested local, boundary, and global rewards, achieving state-of-the-art across seven benchmarks without any benchmark-specific training.
|
|
CountLoop: Iterative Agent Guided High Instance Image Generation
Anindya Mondal,
Ayan Banerjee,
Sauradip Nag,
Josep Llados,
Xiatian Zhu,
Anjan Dutta
TMLR, 2026
PDF /
arxiv /
web
TL;DR▼
Training-free diffusion pipeline with a VLM planner–critic loop: the planner generates structured instance layouts via chain-of-thought reasoning; the critic provides count and spatial feedback; instance-driven cross-attention masking with cumulative attention composition prevents semantic leakage across high-density scenes, reducing counting error by up to 57%.
|
|
OmniCount: Multi-label Object Counting with Semantic-Geometric Priors
Anindya Mondal,
Sauradip Nag,
Xiatian Zhu,
Anjan Dutta
AAAI, 2025
DOI /
arXiv /
web
TL;DR▼
Training-free multi-label counter coupling CLIP semantic embeddings for open-vocabulary category grounding with SAM point-prompt geometric priors for precise instance segmentation; introduces OmniCount-191, the first benchmark with point, bounding-box, and VQA multi-label count annotations.
|
|
Actor-agnostic Multi-label Action Recognition with Multi-modal Query
Anindya Mondal,
Sauradip Nag,
Joaquin M. Prada,
Xiatian Zhu,
Anjan Dutta
ICCVW, 2023
DOI /
code /
web
TL;DR▼
Transformer decoder treating each action class as a multi-modal semantic query (CLIP visual + text embeddings), decoupling action classification from actor-specific pose topology for actor-agnostic multi-label recognition; state-of-the-art across five benchmarks spanning human and animal actions.
|
|
Time-varying Signals Recovery via Graph Neural Networks
John A. Castro-Correa, Jhony H. Giraldo, Anindya Mondal, Mohsen Badiey, Thierry Bouwmans, Fragkiskos D. Malliaros
ICASSP, 2023
PDF /
DOI
TL;DR▼
Encoder-decoder GNN (TimeGNN) trained with a composite loss of MSE and a Sobolev graph smoothness operator, exploiting spatio-temporal correlations across graph topology for robust missing-entry recovery on real sensor and traffic datasets.
|
|
Recovery of Missing Sensor Data by Reconstructing Time-varying Graph Signals
Anindya Mondal, Mayukhmali Das, Aditi Chatterjee, Palaniandavar Venkateswaran
EUSIPCO, 2022
PDF /
DOI /
code
TL;DR▼
Formulates missing wireless sensor data recovery as time-varying graph signal reconstruction minimising a Sobolev norm combining graph Laplacian smoothness across spatial and temporal dimensions, surpassing state-of-the-art by up to 54% at high missing-data rates.
|
|
Moving Object Detection for Event-based Vision using Graph Spectral Clustering
Anindya Mondal, Shashant R, Jhony H. Giraldo, Thierry Bouwmans, Ananda S. Chowdhury
ICCVW, 2021
PDF /
DOI /
code
TL;DR▼
Maps asynchronous event streams to a k-NN graph, applies graph Laplacian spectral decomposition for eigenvector-based clustering (GSCEventMOD), and determines the optimal cluster count via spectral gap analysis, enabling fully unsupervised moving object detection for event cameras.
|
|
University of Surrey, United Kingdom
Degree: Doctor of Philosophy (Ph.D.) in Artificial Intelligence
Thesis: Integrating Auxiliary Information for Visual Scene Understanding and Generation
Oct 2022 - Oct 2026
|
|
Jadavpur University, Kolkata, India
Degree: Bachelor of Engineering (B.E. Hons.) in Electronics & Telecom, First Class with Honours (GPA 8.79/10)
Thesis: Graph Signal Processing and Event-based Vision
Aug 2018 - May 2022
|
Awards, Service & Teaching
Awards & Funding
- University of Surrey Doctoral Studentship (fully funded Ph.D.), 2022–2026
- AAAI-25 Travel Award
- ICCV-23 Attendance Grant
Academic Service
- Reviewer: CVPR, ICCV, ECCV, NeurIPS, AAAI, SIGGRAPH, BMVC, IEEE TSP
Teaching & Mentorship
- Supervision: 10+ MSc Dissertations (University of Surrey)
- Lead Instructor: EEEM076 – Applied Machine Learning (University of Surrey)
- Teaching Assistant:
- EEEM071 – Advanced Topics in Computer Vision and Deep Learning (University of Surrey)
- UKRI Centre for Doctoral Training in AI for Digital Media Inclusion
Teaching reference letter (PDF)
|
|