Publications

Contemporary advancements in language model reasoning often depend on computationally intensive reinforcement learning (RL) and massive datasets, creating prohibitive barriers for resource-constrained teams. In this work, we demonstrate that high-quality, iterative training on minimal data can rival modern RL approaches. We introduce a resource-efficient framework that combines Direct Preference Optimization

Speculative decoding is an effective technique for accelerating large language model (LLM) inference by drafting multiple tokens in parallel. However, its practical speedup is often limited by a rigid verification step, which strictly enforces that the accepted token distribution exactly matches that of the target model. This constraint leads to the rejection of many plausible tokens, reducing the acceptance

The integration of Computer-Aided Design (CAD) models into discrete event simulation software is a critical requirement for many simulation projects, particularly those involving the movement of people or vehicles where spatial accuracy directly impacts study outcomes. While importing CAD files and configuring simulation elements is essential for system accuracy, this process is typically time-consuming

The efficient implementation of large language models (LLMs) is crucial for deployment on resource-constrained devices. Low-rank tensor compression techniques, such as tensor-train (TT) networks, have been widely studied for over-parameterized neural networks. However, their applications to compress pre-trained large language models (LLMs) for downstream tasks (post-training) remains challenging due to

Large Language Models (LLMs) perform well on short-horizon tasks but struggle with long-horizon, multimodal scenarios that require multi-step reasoning, perception, and adaptive planning. We identify two key challenges in these settings: the difficulty of long-term coordination between planning and execution within single-agent architectures and the inefficiency of indiscriminate visual grounding. To address

Large Language Models (LLMs) are increasingly deployed for structured data generation, yet output consistency remains critical for production applications. We introduce a comprehensive framework for evaluating and improving consistency in LLM-generated structured outputs. Our approach combines: (1) STED (Semantic Tree Edit Distance), a novel similarity metric balancing semantic flexibility with structural

Conveyors play a crucial role in transporting packages and containers in manufacturing and production facilities. While computer vision has emerged as a promising technology for real-time monitoring of transportation systems, its application in conveyor operations remains in the early stages. This paper introduces an Industrial Internet of Things (IIoT) framework for real-time conveyor monitoring. We first

Simulation plays a central role in the strategic planning and operational evaluation of supply chain networks. Within these networks, order fulfillment traditionally requires solving computationally expensive optimization problems in real-time across multiple constraints. For forward-looking simulations evaluating millions of orders, such optimization becomes prohibitively expensive. We develop a neural

Benchmark design plays a critical role in advancing large language models (LLMs) by providing systematic frameworks to measure, compare, and improve their capabilities across diverse tasks. While existing benchmarks have driven progress in areas such as factual question answering, code synthesis, and scientific problem solving, there remains a notable lack of benchmarks tailored to evaluate LLMs in real-world

Online randomized controlled experiments (A/B tests) measure causal changes in industry. While these experiments use incremental changes to minimize disruption, they often yield statistically insignificant results due to low signal-to-noise ratios. Precision improvement (or reducing standard error) traditionally focuses on trigger observations - where treatment and control outputs differ. Though effective

We present an approach for training language models to interactively prove theorems using the Lean proof assistant. Our approach enables models to propose partial proofs, receive verification feedback, and iteratively refine their proofs. We develop a synthetic data generation pipeline that converts static proof datasets into multi-turn interactive sequences, complete with incremental verification feedback

For music streaming services expanding into audiobooks, cold-start personalization presents a critical challenge: as audiobooks are a newly introduced content type, the vast majority of existing users have no audiobook listening history. This domain-level cold-start scenario differs from traditional item or user cold-start scenarios, since personalization must begin before any behavioral data exists in

We present a protocol for fault-tolerantly implementing the logical quantum random access memory (QRAM) operation, given access to a specialized, noisy QRAM device. For coherently accessing classical memories of size 2^n, our protocol consumes only poly(n) fault-tolerant quantum resources (logical gates, logical qubits, quantum error correction cycles, etc.), avoiding the need to perform active error correction

In this paper, we present CACHE-ED, a novel framework for document entity extraction that combines the power of large language models (LLMs) with graph-based document representations, caching mechanisms, and an actor-critic multi-agent architecture. Our approach addresses the inefficiencies and inaccuracies that are common in extracting structured information from documents, particularly in templated formats

Length control in Large Language Models (LLMs) is a crucial but under-addressed challenge, with applications ranging from voice interfaces requiring concise responses to research summaries needing comprehensive outputs. Current approaches to length control, including Regularized DPO, Length-Instruction Fine-Tuning, and tool-augmented methods, typically require expensive model retrain-ing or complex inference-time