Hitesh Patel @Hitesh_LPatel

SyncMask: Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining This paper proposes SyncMask, a method to address the disparity between image and text information in fashion datasets for VLMs. It generates masks to synchronize attention between image…

DOCMASTER: A Unified Platform for Annotation, Training, & Inference in Document Question-Answering This paper presents DOCMASTER, a platform for annotating PDF documents, model training, and inference, focused on document question-answering. It addresses challenges in working…

WavLLM: Towards Robust and Adaptive Speech Large Language Model This paper introduces WavLLM, a speech LLM with dual encoders and prompt-aware weight adaptation, trained via a curriculum learning approach. It achieves state-of-the-art performance across various speech tasks,…

ARAGOG: Advanced RAG Output Grading The paper addresses the gap in extensive experimental comparisons of RAG methods. The study finds that Hypothetical Document Embedding (HyDE) and LLM reranking enhance retrieval precision significantly, while other methods show mixed results.…

Stable Code The paper introduces Stable Code, a code language model for code completion, reasoning, math, and other software engineering tasks. It includes a variant called stable code instruct for natural language interaction. The report details data, training, and evaluations,…

FABLES: Evaluating faithfulness and content selection in book-length summarization This paper evaluates faithfulness and content selection in summaries generated by LLMs for fictional books. They create a dataset called FABLES, comprising annotations on over 3,000 claims in…

Towards Safety and Helpfulness Balanced Responses via Controllable Large Language Models The paper addresses the challenge of balancing safety and helpfulness in LLMs to enhance user experience. It proposes methods for controlling both attributes without additional human…

BioMedLM: A 2.7B Parameter Language Model Trained On Biomedical Text The paper introduces BioMedLM, a smaller GPT-style model trained exclusively on PubMed data. Despite its smaller size, it achieves competitive results in biomedical question-answering tasks, outperforming…

LONG-FORM FACTUALITY IN LARGE LANGUAGE MODELS The paper introduces SAFE, to evaluate long-form actuality using LLM by breaking down responses into individual facts and verifying them using search queries. It introduces an extended F1 score as a metric for long-form actuality.…

Sorry, Come Again? Prompting – Enhancing Comprehension and Diminishing Hallucination with [PAUSE] -injected Optimal Paraphrasing The paper investigates the impact of formality, readability, and concreteness on hallucination for 21 LLMs. It introduces SCA, an optimal…

MFORT-QA: Multi-hop Few-shot Open Rich Table Question Answering The paper introduces MFORT-QA, a method for Table QA using LLMs. It combines few-shot learning to retrieve relevant tables and contexts, and CoT prompting to decompose complex questions. RAG enhances the process by…

Fine-Tuning Language Models with Reward Learning on Policy The paper proposes Reward Learning on Policy (RLP), an unsupervised framework to refine reward models using policy samples, addressing off-distribution issues in RLHF. It utilizes multi-view learning for robust…

BP4ER: Bootstrap Prompting for Explicit Reasoning in Medical Dialogue Generation The paper introduces BP4ER, a method for medical dialogue generation that explicitly models multi-step reasoning processes without relying on extensive entity annotation. It employs least-to-most…

Gecko: Versatile Text Embeddings Distilled from Large Language Models The paper introduces Gecko, a compact text embedding model that achieves strong retrieval performance by distilling knowledge from LLMs. It utilizes a two-step distillation process, generating synthetic paired…

Jamba: A Hybrid Transformer-Mamba Language Model The paper introduces Jamba, a hybrid large language model combining Transformer and Mamba layers using a MoE architecture. Jamba achieves state-of-the-art performance on language model benchmarks and long-context evaluations…

ReALM: Reference Resolution As Language Modeling The paper explores using large language models (LLMs) for reference resolution, including non-conversational entities like on-screen elements. It demonstrates significant improvements over existing systems, achieving comparable…

Towards Greener LLMs: Bringing Energy-Efficiency to the Forefront of LLM Inference This paper explores the energy-efficient serving of LLMs in data centers, balancing performance with energy consumption. It examines various factors affecting energy usage, latency, and…

ELITR-Bench: A Meeting Assistant Benchmark for Long-Context Language Models This paper introduces ELITR-Bench, a new benchmark for evaluating long-context LLMs in a practical meeting assistant scenario. It addresses challenges posed by noisy and oral data from transcripts,…

ProCQA: A Large-scale Community-based Programming Question Answering Dataset for Code Search The paper introduces ProCQA, a large-scale programming question-answering dataset from Stackoverflow, facilitating naturally structured mixed-modal QA pairs. It proposes a…

Language Rectified Flow: Advancing Diffusion Language Generation with Probabilistic Flows The paper introduces Language Rectified Flow as an alternative to diffusion language models for controlling sentence attributes and structure. LF utilizes neural ordinary differential…