Dravyansh Sharma

Recent News and Highlights

★ Organizing a new COLT 2026 tutorial, Learning-Augmented Algorithms: Leveraging Data for Performance with Theoretical Guarantees, with Maryam Aliakbarpour, and Sandeep Silwal. Join us at COLT 2026!
★ Organizing a new workshop Learning in an Agentic World: Foundations and Challenges at COLT 2026 on June 29 at San Diego, with Hedyeh Beyhaghi, Han Shao, and Avrim Blum. Consider submitting to our Call for Abstracts!
★ Organizing a workshop on Learning-driven Algorithms and Machine-aided Proofs (LAMP) at TTIC on August 6–7, 2026, with Sandeep Silwal and Ellen Vitercik. Consider submitting to our call for spotlights/posters/open problems!
★ Presented a tutorial on New Frontiers of Hyperparameter Optimization at NeurIPS 2025, with Nina Balcan and Colin White.
★ Taught a new course TTIC 31290 - Machine Learning for Algorithm Design with Avrim Blum, Fall 2025.
★ Service: Area Chair at ICML 2026, NeurIPS 2026, and NeurIPS 2025.
★ Reviewer (recent): COLT 2026, JMLR 2026, TMLR 2026, UAI 2026, AISTATS 2026, SODA 2026, ITCS 2026, FOCS 2026.
★ Three accepted papers at NeurIPS 2025, including one spotlight.
★ Gave a tutorial at AutoML 2025, Limitations of State-of-the-Art and a New Principled Framework for HPO and Algorithm Selection.
★ Gave a tutorial at UAI 2025, Hyperparameter Optimization and Algorithm Selection: Practical Techniques, Theory, and New Frontiers.

Research

My work spans across learning theory, algorithms, and reliable AI, with a particular recent emphasis on verifier-based reasoning systems and learning in adaptive environments.

LLM verifiers and reasoning systems

I study the learnability of verifiers for LLM Chain-of-Thought reasoning, including online and PAC guarantees, soundness-completeness trade-offs, and how learned verifiers can reliably guide step-by-step reasoning post-training via process supervision.

Online Learnability of Chain-of-Thought Verifiers: Soundness and Completeness Trade-offs. Preprint (under submission), ICML 2026 AI4Math Workshop. With Maria-Florina Balcan, Avrim Blum, Kiriaki Fragkia, and Zhiyuan Li.
On Learning Verifiers and Implications to Chain-of-Thought Reasoning. NeurIPS 2025. With Maria-Florina Balcan, Avrim Blum, and Zhiyuan Li.

Data-driven algorithm design

I study how to learn algorithmic choices, heuristics, hyperparameters, and configurations across problem instances, with guarantees for generalization, online learning, and performance across tasks.

Gradient Descent with Provably Tuned Learning-rate Schedules. AISTATS 2026.
Learning Accurate and Interpretable Decision Trees. UAI 2024, Outstanding Student Paper Award. With Maria-Florina Balcan.
Provably tuning the ElasticNet across instances. NeurIPS 2022. With Maria-Florina Balcan, Mikhail Khodak, and Ameet Talwalkar.
Data Driven Semi-Supervised Learning. NeurIPS 2021, Oral. With Maria-Florina Balcan.

Reliable learning

I work on learning algorithms with formal reliability guarantees under distribution shift, test-time attacks, poisoning, and other adversarial settings.

Robustly-reliable learners under poisoning attacks. COLT 2022. With Maria-Florina Balcan, Avrim Blum, and Steve Hanneke.
Reliable learning in challenging environments. NeurIPS 2023. With Maria-Florina Balcan, Steve Hanneke, and Rattana Pukdee.
An analysis of robustness of non-Lipschitz networks. JMLR 2023. With Maria-Florina Balcan, Avrim Blum and Hongyang Zhang

Learning with agents

I study learning in environments where data is produced by agents who may adapt, improve, or strategically respond to deployed models and algorithms.

PAC Learning with Improvements. ICML 2025. With Idan Attias, Avrim Blum, Keziah Naggita, Donya Saless, and Matthew Walter.
Conservative classifiers do consistently well with improving agents: characterizing statistical and online learning. NeurIPS 2025 Spotlight. With Alec Sun.
Subsidy design for better social outcomes. Pre-print. With Maria-Florina Balcan and Matteo Pozzi.

Publications

2026

[W14] Online Learnability of Chain-of-Thought Verifiers: Soundness and Completeness Trade-offs, AI4Math, ICML 2026

with Maria-Florina Balcan, Avrim Blum, Kiriaki Fragkia and Zhiyuan Li

[C26] Algorithm Design and Stronger Guarantees for the Improving Multi-Armed Bandits Problem, UAI 2026

with Avrim Blum, Marten Garicano and Kavya Ravichandran

[C25] Gradient Descent with Provably Tuned Learning-rate Schedules, AISTATS 2026

2025

[C24] Conservative classifiers do consistently well with improving agents: characterizing statistical and online learning, NeurIPS 2025 Spotlight

with Alec Sun
[C23] On Learning Verifiers and Implications to Chain-of-Thought Reasoning, NeurIPS 2025

with Maria-Florina Balcan, Avrim Blum and Zhiyuan Li
[C22] Sample complexity of data-driven tuning of model hyperparameters in neural networks with structured parameter-dependent dual function, NeurIPS 2025

with Maria-Florina Balcan and Anh Nguyen
[C21] Algorithm Configuration for Structured Pfaffian Settings, AutoML 2025, Hot-off-the-press track; also accepted to TMLR 2025

with Maria-Florina Balcan and Anh Tuan Nguyen
[C20] PAC Learning with Improvements, ICML 2025

with Idan Attias, Avrim Blum, Keziah Naggita, Donya Saless and Matthew Walter
[C19] Tuning Algorithmic and Architectural Hyperparameters in Graph-Based Semi-Supervised Learning with Provable Guarantees, UAI 2025

with Ally Yalei Du and Eric Huang
[C18] Offline-to-online hyperparameter transfer for stochastic bandits, AAAI 2025

with Arun Sai Suggala
[J3] Algorithm Configuration for Structured Pfaffian Settings, TMLR 2025

with Maria-Florina Balcan and Anh Tuan Nguyen
[A1] Learning Accurate and Interpretable Decision Trees, Extended Abstract, IJCAI 2025 Best Papers from Sister Conferences Track

with Maria-Florina Balcan
[A2] An Analysis of Robustness of Non-Lipschitz Networks, Extended Abstract, Workshop on Safe AI, UAI 2025

with Maria-Florina Balcan, Avrim Blum and Hongyang Zhang
[W13] Distribution-dependent Generalization Bounds for Tuning Linear Regression Across Tasks, DiffCoALG: Differentiable Learning of Combinatorial Algorithms, NeurIPS 2025

with Maria-Florina Balcan and Saumya Goyal
[W12] Algorithm design and sharper bounds for improving-bandits, OPT2025: 17th Annual Workshop on Optimization for Machine Learning, NeurIPS 2025

with Avrim Blum, Marten Garicano and Kavya Ravichandran
[W11] Learning reliably under adversarial attacks, distribution shifts and strategic agents, Reliable ML from Unreliable Data, NeurIPS 2025

with Maria-Florina Balcan
[W10] On Learning Verifiers for Chain-of-Thought Reasoning, Workshop on Reliable and Responsible Foundation Models, ICML 2025

with Maria-Florina Balcan, Avrim Blum and Zhiyuan Li
[W9] PAC Learning with Improvements, 2nd Workshop on Social Choice and Learning Algorithms, IJCAI 2025

with Idan Attias, Avrim Blum, Keziah Naggita, Donya Saless and Matthew Walter
[W8] Learning how to step in gradient-based optimization: beyond convexity and smoothness, 3rd Workshop on High-dimensional Learning Dynamics, ICML 2025
[W7] Gradient descent in presence of extreme flatness and steepness, Methods and Opportunities at Small Scale, ICML 2025

2024

[T1] CMU CSD PhD Thesis: Data-driven algorithm design and principled hyperparameter tuning in machine learning
[C17] An Analysis of Robustness of Non-Lipschitz Networks, NeurIPS 2024, Journal-to-conference track

with Maria-Florina Balcan, Avrim Blum and Hongyang Zhang
[C16] Accelerating ERM for data-driven algorithm design using output-sensitive techniques, NeurIPS 2024

with Maria-Florina Balcan and Christopher Seiler
[C15] Subsidy for repair in component maintenance games, EMI/PMC 2024

with Maria-Florina Balcan and Matteo Pozzi
[C14] Learning Accurate and Interpretable Decision Trees, UAI 2024 Outstanding Student Paper Award

with Maria-Florina Balcan
[C13] No Internal Regret with Non-convex Loss Functions, AAAI 2024
[W6] Theoretical Analyses of Hyperparameter Selection in Graph-Based Semi-Supervised Learning, ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling

with Ally Yalei Du and Eric Huang
[W5] Accelerating data-driven algorithm design using output-sensitive techniques, AAAI 2024 Workshop on Learnable Optimization

with Maria-Florina Balcan and Christopher Seiler
[W4] Shifting regret for tuning combinatorial algorithms with applications to clustering, AAAI 2024 Workshop on Learnable Optimization

with Maria-Florina Balcan and Travis Dick

2023

[C12] New Bounds for Hyperparameter Tuning of Regression Problems Across Instances, NeurIPS 2023

with Maria-Florina Balcan and Anh Tuan Nguyen
[C11] Reliable learning in challenging environments, NeurIPS 2023

with Maria-Florina Balcan, Steve Hanneke and Rattana Pukdee
[C10] Efficiently Learning the Graph for Semi-supervised Learning, UAI 2023

with Maxwell Jones
[J2] An analysis of robustness of non-Lipschitz networks, JMLR 2023; earlier version in ICLR 2022 SRML workshop

with Maria-Florina Balcan, Avrim Blum and Hongyang Zhang

2022

[C9] Provably tuning the ElasticNet across instances, NeurIPS 2022 [blog post]

with Maria-Florina Balcan, Mikhail Khodak and Ameet Talwalkar
[C8] Robustly-reliable learners under poisoning attacks, COLT 2022

with Maria-Florina Balcan, Avrim Blum and Steve Hanneke
[W3] On the Power of Abstention and Data-Driven Decision Making for Adversarial Robustness, ICLR 2022 Workshop on Socially Responsible Machine Learning Oral

with Maria-Florina Balcan, Avrim Blum and Hongyang Zhang

2021

[C7] Data driven semi-supervised learning, NeurIPS 2021 Oral, <1%

with Maria-Florina Balcan
[C6] Learning-to-learn non-convex piecewise-Lipschitz functions, NeurIPS 2021

with Maria-Florina Balcan, Mikhail Khodak, Ameet Talwalkar
[W2] Improved pronunciation prediction accuracy using morphology, ACL SIG on Computational Morphology and Phonology, ACL 2021

with Saumya Sahai, Neha Chaudhari, Antoine Bruguier
[W1] Predicting and Explaining French Grammatical Gender, ACL Special Interest Group on Typology, NAACL 2021

with Saumya Sahai

2020 and before

[C5] Learning Piecewise Lipschitz Functions in Changing Environments, AISTATS 2020 [slides]

with Maria-Florina Balcan and Travis Dick
[C4] Better morphology prediction for better speech systems, Interspeech 2019

with Melissa Wilson and Antoine Bruguier
[C3] On Training and Evaluation of Grapheme-to-Phoneme Mappings with Limited Data, Interspeech 2018
[C2] Dictionary Augmented Sequence-to-Sequence Neural Network for Grapheme to Phoneme Prediction, Interspeech 2018

with Antoine Bruguier and Anton Bakhtin
[J1] Some results on a class of mixed van der Waerden numbers, Rocky Mountain J. Math. 2018

with Kaushik Maran, Sai Praneeth Reddy and Amitabha Tripathi
[C1] On greedy maximization of entropy, ICML 2015

with Amit Deshpande and Ashish Kapoor

Tutorials and Teaching

COLT 2026 tutorial: Learning-Augmented Algorithms: Leveraging Data for Performance with Theoretical Guarantees, with Maryam Aliakbarpour, and Sandeep Silwal.
TTIC 31290: Machine Learning for Algorithm Design, with Avrim Blum, Fall 2025.
NeurIPS 2025 tutorial: New Frontiers of Hyperparameter Optimization, with Nina Balcan and Colin White.
AutoML 2025 tutorial: Limitations of State-of-the-Art and a New Principled Framework for HPO and Algorithm Selection.
UAI 2025 tutorial: Hyperparameter Optimization and Algorithm Selection: Practical Techniques, Theory, and New Frontiers.

Workshops and Service

LAW at COLT 2026

Learning in an Agentic World: Foundations and Challenges. June 29, 2026, San Diego.

LAMP at TTIC

Learning-driven Algorithms and Machine-aided Proofs. August 6–7, 2026, TTIC.

Area chair

2026: ICML 2026, NeurIPS 2026
2025: NeurIPS 2025.

Reviewer

2026: COLT, JMLR, TMLR, UAI, AISTATS, SODA, ITCS, FOCS.
Before 2026 (other venues): ICML, NeurIPS, AAAI, ALT, TPAMI, SIMODS, Machine Learning.

Reach Me

Email: dravy [AT] ttic [DOT] edu
Office: 434, Toyota Technological Institute at Chicago, 6045 S Kenwood Ave, Chicago, IL 60637