Esha Singh

News

Nov 2026 Attending SAIL 2026 in San Diego!

Sep 2026 New preprint: Learning Where to Steer Noise.

Jul 2026 Presenting Divide and Learn at ICML 2026, Seoul!

Jun 2026 Started as Applied Scientist Intern II at AWS AGI Foundations!

Feb 2026 Divide and Learn: Multi-objective Combinatorial Optimization at Scale, accepted at ICML 2026!

Research

My research focuses on building principled and scalable algorithms for generative models and large language models, driven by a core question: how can we design learning algorithms that are simultaneously theoretically grounded, computationally efficient, and practically impactful?

Recently, I have worked on steering flow and diffusion models efficiently, finding the few directions in noise space that matter and developing guidance approaches that are both manifold-preserving and Pareto-aware. Currently, I am exploring one-step diffusion and long-horizon agentic tasks such as RL-trained terminal agents, where efficient generation and decision-making must go hand in hand.

Generative models & test-time steering

Generative modeling for multimodal audio and video, and efficient reasoning.

Learning Where to Steer Noise AmbientNova

Multi-objective & combinatorial optimization

Finding Pareto-optimal trade-offs at scale, with orders-of-magnitude better sample efficiency.

Divide and Learn ICML 2026 Learning Where to Steer Noise

Optimization theory & generalization

Why training dynamics find solutions that generalize, and optimizers that exploit this.

Stable Minima NeurIPS Spotlight MoXCo PMLR 2025

Efficient large language models

Making LLM training and inference cheaper without giving up quality.

Unstructured Sparsity NeurIPS Speculative decoding AWS AGI

Selected Publications

^*Equal contribution

Divide and Learn: Multi-Objective Combinatorial Optimization at Scale

Esha Singh, Dongxia Wu, Chien-Yi Yang, Tajana Rosing, Rose Yu, Yi-An Ma

ICML 2026

Paper Blog Code Poster

Reformulates multi-objective combinatorial optimization as online learning over decomposed decision spaces, achieving 80–98% of specialized solvers with two to three orders of magnitude improvement in sample and computational efficiency.

Learning Where to Steer Noise: Efficient Offline Multi-Objective Optimization with Generative Models

Yuan Lu^*, Esha Singh^*, Yi-An Ma, Yusu Wang

Preprint, 2026

Paper Blog Code

Steers only the initial noise of a generative model toward the Pareto front, using a few sensitive directions estimated once per task with a Recursive Feature Machine. Outranks the best prior generative method at a fraction of its sampling cost.

Stable Minima Cannot Overfit in Univariate ReLU Networks: Generalization by Large Step Sizes

D. Qiao, K. Zhang, Esha Singh, Daniel Soudry, Yu-Xiang Wang

NeurIPS 2024 (Spotlight)

Paper

MoXCo: How I Learned to Stop Exploring and Love My Local Minima

Esha Singh, Shoham Sabach, Yu-Xiang Wang

CPAL 2025 (PMLR) · NeurIPS 2023 (M3L Workshop)

Paper Poster

Adaptive optimizers that expedite exploration with faster convergence while ensuring avoidance of over-exploitation, leading to convergence to good solutions.