VocabTrim: Vocabulary Pruning for Efficient Speculative Decoding in LLMs
R Goel, S Agrawal, M Gagrani, J Park, Y Zao, H Zhang, Y Yang, X Yuan, J Lu, M Lee^, C Lott^
ICML Workshop on Efficient Systems for Foundational Models (ES-FOMO), 2025
Paper

Speculative decoding (SpD) speeds up LLM inference by using smaller draft models to propose token sequences, which are then verified by a larger base model. VOCABTRIM introduces a training-free method to reduce the overhead of drafting by pruning the vocabulary used during draft generation. By retaining only the most frequently accepted tokens from the target model, VOCABTRIM significantly reduces memory-bound latency—especially beneficial for edge devices. Despite a slight drop in acceptance rate, it achieves up to 16% speed-up on Llama-3.2-3B-Instruct in Spec-Bench, making it a practical optimization for real-world deployment.

CAOTE: KV Cache Eviction for LLMs via Attention Output Error-Based Token Selection
R Goel, J Park, M Gagrani, D Jones, M Morse, H Langston, M Lee^, C Lott^
Paper

CAOTE introduces a principled token eviction strategy for long-context LLMs that balances memory efficiency with inference quality. Unlike prior methods that rely solely on attention scores, CAOTE estimates each token's contribution to the attention output using both attention weights and value vectors. This hybrid metric allows CAOTE to explicitly optimize for eviction error, ensuring that removed tokens minimally impact model predictions. The method improves latency and downstream task accuracy, and also serves as a meta-heuristic that enhances existing eviction strategies across diverse models and tasks.

KeyDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference
J park, D Jone, M Morse, R Goel, M Lee^, C Lott^
Paper

KeyDiff introduces a training-free cache eviction method based on key similarity, enabling efficient long-context inference in resource-constrained environments. By identifying geometrically distinctive keys that correlate with high attention scores, KeyDiff retains the most impactful tokens without relying on attention mechanisms—making it compatible with optimized attention implementations like FlashAttention. It achieves near-baseline performance with ~23% KV cache reduction and up to 30% latency reduction, validated across Llama and Qwen models on LongBench and Math500.

On Speculative Decoding for Multimodal Large Language Models
M Gagrani*, R Goel*, W Jeon, J Park, M Lee^, C Lott^
Spotlight (top-4 papers)
CVPR Workshop on Efficient Large Vision Models (eVLM), 2024
Paper

This paper explores the application of speculative decoding to enhance the inference efficiency of multimodal large language models (MLLMs), specifically the LLaVA 7B model. The key contribution is demonstrating that a language-only model can serve as an effective draft model for speculative decoding, achieving significant speedups without the need for image tokens

Direct Alignment of Draft Model for Speculative Decoding with Chat-Fine-Tuned LLMs
R Goel, M Gagrani, W Jeon, J Park, M Lee, C Lott
ICLR Workshop on Understanding of Foundational Models, 2024
Paper

This paper proposes a framework for training draft models directly aligned with chat-fine-tuned large language models (LLMs). The key contribution is the introduction of the Llama 2 Chat Drafter 115M, which achieves up to 2.4× speed-up in inference relative to autoregressive decoding, using a novel Total Variation Distance++ (TVD++) loss for improved alignment

Recursive Speculative Decoding: Accelerating LLM Inference via Sampling Without Replacement
W Jeon, M Gagrani, R Goel, J Park, M Lee, C Lott
ICLR Workshop on LLM Agents, 2024
Paper

This paper presents Recursive Speculative Decoding (RSD), a novel tree-based method that samples draft tokens without replacement to maximize diversity and efficiency. The key contribution is the empirical demonstration that RSD outperforms baseline methods in both fixed draft sequence length and fixed computational budget scenarios, significantly accelerating LLM inference

Composite Adaptive Control for Time-varying Systems with Dual Adaptation
R Goel, SB Roy
IEEE Transaction on Automatic Control (TAC), 2025  
Paper

Introduces a novel control architecture that employs a dual adaptation scheme to handle dynamical systems with time-varying uncertain parameters. Key contributions include the integration of projection and $\sigma$-modification algorithms to achieve global tracking error stability, and the use of a less restrictive initial excitation (IE) condition instead of the traditional persistence of excitation (PE) requirement for parameter estimation

Motion-aware Needle Segmentation in Ultrasound Images
R Goel, C Morales*, M Singh*, A Dubrawski, J Galeotti, H Choset
International Symposium on Bio Medical Imaging (ISBI), 2024 CVPR Workshop (medical vision), 2024
Paper

A novel approach that combines classical Kalman Filter techniques with data-driven learning to improve needle segmentation in 2D ultrasound images. The key contributions include a framework compatible with encoder-decoder architectures, superior performance with a 15% reduction in pixel-wise needle tip error and an 8% reduction in length error, and the implementation of a learnable filter for non-linear needle motion

Autonomous Ultrasound Scanning using Bayesian Optimization and Hybrid Force Control
R Goel*, Abhimanyu*, K Patel, J Galeotti, H Choset
International Conference on Robotics and Automation (ICRA), 2022
Paper

Proposes an innovative robotic ultrasound system that leverages Bayesian Optimization (BO) and hybrid force control to autonomously scan regions for high-quality diagnostic images. Key contributions include the use of Gaussian processes to estimate a quality map based on expert demonstrations, and the integration of deep convolutional neural networks for real-time image quality feedback, achieving high accuracy in probe positioning and force application

Closed-Loop Reference Model Based Distributed MRAC Using Cooperative Initial Excitation and Distributed Reference Input Estimation
R Goel*, T Garg, SB Roy
IEEE Transaction on Control of Network Systems (TCNS), 2022  
Paper

introduces a novel distributed model reference adaptive control (DMRAC) framework for multi-agent systems. Key contributions include the use of a closed-loop reference model (CRM) to enhance transient performance and the implementation of cooperative initial excitation (IE) for improved parameter estimation without the need for persistent excitation (PE) conditions

Closed-loop reference model based distributed model reference adaptive control for multi-agent systems
R Goel*, SB Roy
Letters of Controls and Systems (L-CSS), 2021  
American Control Conference (ACC), 2021  
Paper

presents a distributed control framework that integrates closed-loop reference models (CRM) to enhance the transient performance of multi-agent systems. Key contributions include the use of cooperative initial excitation (IE) for improved parameter estimation without relying on persistent excitation (PE) conditions, and the implementation of distributed reference input estimation to ensure robust and adaptive control across the network

Leader and predator based swarm steering for multiple tasks
R Goel, J Lewis, MA Goodrich, PB Sujit
Inernation Conference on System, Man and Cybernetics (SMC), 2019  
Paper / video

explores the use of leaders and predators to influence robotic swarms in performing various tasks. Key contributions include the analysis of different swarm models (shepherding, Couzin's, and physicomimetic) using Monte-Carlo simulations, and the demonstration that predator-based swarm splitting and steering significantly outperforms other methods, even with large numbers of agents

Dynamic Task Allocation Using Multi-Agent Mobile Robots
Raghavv Goel, Sha Yi, Jaskaran Singh Grover, Katia Sycara
RISS Journal, 2019  
poster / video

We propose to solve task allocation problem in heterogeneous agents using mix integer linear program with collision avoidance and communication breakage constraints