Scholar
Tarun Suresh
Google Scholar ID: Yxx6B5YAAAAJ
Undergraduate, University of Illinois Urbana-Champaign
Citations & Impact
Citations
283
H-index
8
i10-index
7
Publications
20
Co-authors
5
list available
Publications
Resume (English only)
Academic Achievements
- NeurIPS 2025: DINGO: Constrained Inference for Diffusion LLMs (co-first author) – introduced the first dynamic-programming-based decoding strategy for diffusion LLMs that provably satisfies regex constraints
- ICML 2025: CRANE: Reasoning with Constrained LLM Generation (co-first author) – theoretically analyzed and mitigated reasoning degradation under output constraints, boosting accuracy by up to 10% on symbolic reasoning tasks
- ICLR 2025: IterGen: Iterative Semantic-aware Structured LLM Generation with Backtracking – developed a library enabling grammar-symbol-level backtracking to repair semantic violations, improving SQL accuracy by 18% and preventing privacy leakage
- TMLR 2025: SynCode: LLM Generation with Grammar Augmentation – proposed a sound and complete grammar-guided LLM generation framework, reducing syntax errors by 96–100% and accelerating inference by 1.5–10x
- TMLR 2025: Two-Step Offline Preference-Based Reinforcement Learning with Constrained Actions – co-developed the PRC method that constrains action space to improve stability in offline preference-based RL
- ARLET @ NeurIPS 2025: Learning a Pessimistic Reward Model in RLHF – contributed to PET, a pessimistic reward fine-tuning approach robust against reward hacking in offline RLHF
