The research arm of AI Wranglers
optimal* lab works where optimization meets machine learning, and points both at one question: how do we build AI that is safe, trustworthy and aligned with the people it serves? Our long-term mission is a safe and humane planet for all.
optimal*
Optimization · Machine Learning
Research highlights
Pick a theme to see the work behind it. The same filter applies to every publication below.
Diffusion language models EMNLP 2026 · Main
Diffusion language models do not write left to right: they refine a whole draft over many denoising steps. TAP asks how a policy should interface with that process, and proposes policy interfaces built for it.
Safe multi-agent planning AAMAS 2026
Several agents plan together by denoising their trajectories with a diffusion model, with safety treated as part of the plan rather than a check bolted on afterwards.
Safe offline RL AAMAS 2026
Learning to act safely from logged experience alone, with no risky trial and error in the world, using diffusion models as expressive policies.
Vision-language alignment CVPR 2026 · VisCon
Matching image regions to the words that describe them as an optimal transport problem: low-rank and unbalanced, so fine-grained concepts find their partners. A companion paper anchors visual concepts at test time with entropic OT.
World models for robots RSS 2026 · Workshop
A robot that plans inside a learned world model is only as safe as that model is robust. This study judges latent world models by what matters to the planner that relies on them.
Optimization on manifolds SIAM J. Optimization 2023
Min-max problems, the mathematics behind adversarial training and robust learning, posed on curved spaces and solved with Riemannian Hamiltonian methods.
No featured story for this theme yet. Its papers are listed below.
Publications
TAP: Policy Interfaces for Diffusion Language Models
Quantizing Masked Diffusion Language Models: What Compresses, What Breaks, and What Transfers
MA-SafeDiffuser: Safe Multi-Agent Planning with Diffusion Probabilistic Models
Safe Offline Reinforcement Learning using Diffusion Policies
Discrete Diffusion Language Models Are Training-Free Multi-Label Classifiers
Test-Time Visual Concept Anchoring via Entropic Optimal Transport
ConceptOT: Fine-Grained Vision-Language Alignment via Low-Rank Unbalanced Optimal Transport
Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks
Semantically Aware Optimal Transport for Dense Label Transfer
Stable 3D Tokens for Foundation-Model Robot Planning
A Planner-Centered Study of Robust Latent World Models
Alignment and Safety of Diffusion Models via Reinforcement Learning and Reward Modeling: A Survey
A Survey of Safe Reinforcement Learning and Constrained MDPs: A Technical Survey on Single-Agent and Multi-Agent Safety
Hierarchical Sparse Plus Low Rank Compression of LLM
Fast and Efficient Modern BERT based Text-Conditioned Diffusion Model for Medical Image Segmentation
A Gauss-Newton Approach for min-max Problem in GAN
Review of Extreme Multilabel Classification
Qualitative Data Augmentation for Performance Prediction in VLSI circuits
Data Augmentation for Performance Prediction in VLSI circuits
A label-free sensing of creatinine using radio frequency-driven lab-on-chip (LOC) system
A Label-Free Low-Cost Radio Frequency Driven Noninvasive Lab-on-Chip System for Creatinine Detection
Riemannian Hamiltonian methods for min-max optimization on manifolds
alphaElimination: Using deep reinforcement learning for sparse Gaussian Elimination
marl-jax: Multi-agent Reinforcement Learning framework for Social Generalization
Light-weight Deep Extreme Multilabel Classification
Effects of Spectral Normalization in Multi-agent Reinforcement Learning
Angle based dynamic learning rate for gradient descent
Adaptive Consensus Optimization Method for GANs
Enhancing ML model accuracy for Digital VLSI circuits using diffusion models
Hybrid Tokenization and Datasets for Solving Mathematics and Science Problems Using Transformers
A Riemannian Approach to Extreme Classification Problems
Nonnegative Low-Rank Tensor Completion via Dual Formulation with Applications to Image and Video Completion
SCIMAT: Science and Mathematics Dataset
Structured Low-Rank Tensor Learning
A Fast Parameter-Free Preconditioner for Structured Grid Problems
A Deflation Based Fast and Robust Preconditioner for Bundle Adjustment
DXML: Distributed Extreme Multilabel Classification
Multigrid Preconditioned Solver for Bundle Adjustment
Efficient FPGA Implementation of Conjugate Gradient Methods for Laplacian System using HLS
Fast Preconditioned Solver for Truncated Saddle Point Problem in Nonsmooth Cahn-Hilliard Model
Multilevel Communication Optimal Least Squares Solver
High Performance Solvers for Implicit Particle in Cell Simulation
Modified tangential frequency filtering decomposition and its Fourier analysis
Roadmap
Each of our papers, the questions it answers, and the related work around it in the field, kept current as the field moves. The full interactive map is in development.
Coming soonNews
Kiran Ravish named among the Top 100 ML Scientists in India by Amazon
Kiran, a PhD student in the group, has publications at AAMAS, EMNLP and NeurIPS.
Five papers accepted at NeurIPS 2026
Across the Main Track and Workshops.
Two papers accepted at EMNLP 2026
TAP in the Main Conference, and Quantizing Masked Diffusion Language Models in the Industry Track.
Two papers accepted at AAMAS 2026
MA-SafeDiffuser, and Safe Offline Reinforcement Learning using Diffusion Policies.
Paper accepted at SIAM SDM 2026
Discrete Diffusion Language Models Are Training-Free Multi-Label Classifiers.
Two papers at the CVPR 2026 VisCon workshop
ConceptOT, and Test-Time Visual Concept Anchoring via Entropic Optimal Transport.
Two papers at ACL 2026 workshops
Speculative Refinement at GEM, and Semantically Aware Optimal Transport for Dense Label Transfer at ALVR.
Two papers at the RSS 2026 Robot Planning workshop
Stable 3D tokens for foundation-model robot planning, and robust latent world models.
ICML 2026 Silver Reviewer Award, and AAAI 2027 Program Committee
Head of the lab
Pawan leads optimal* lab, with more than twenty publications at CORE A and A* venues across optimization, machine learning and AI safety. He did his PhD at INRIA, France (très honorable), followed by postdoctoral research at KU Leuven, Fraunhofer ITWM and FU Berlin, and is a researcher at IIIT Hyderabad, where he supervises PhD and Masters students.
Best Paper Award, IJCNN 2023 · ICML 2026 Silver Reviewer Award · Marie Curie Fellowship · Qualcomm Faculty Award