Publications

AS-ES Learning: Towards Efficient CoT Learning in Small Models
In this paper, We introduce AS-ES learning, a method to improve Chain-of-Thought (CoT) capabilities in small models by better utilizing existing CoT data, rather than merely increasing data from Large Language Models. This approach outperforms traditional training methods on tasks requiring extensive CoT, without needing extra data or model modifications, and provides insights into the mechanics of CoT learning.
PDF Cite
AS-ES Learning: Towards Efficient CoT Learning in Small Models
Beyond Direct Diagnosis: LLM-based Multi-Specialist Agent Consultation for Automatic Diagnosis
In this paper, we introduce an AI-based automatic diagnosis framework, Agent-derived Multi-Specialist Consultation (AMSC), which mimics real-world medical consultations by using language models as virtual medical specialists. This approach outperforms previous methods by efficiently combining agents’ predictions on diseases, requiring less training and parameter updates.
PDF Cite
Beyond Direct Diagnosis: LLM-based Multi-Specialist Agent Consultation for Automatic Diagnosis
HuaTuo: Tuning LLaMA Model with Chinese Medical Knowledge
In this report, we introduce the technical details of HuaTuo, a fine-tuned LLaMA Large Language Model, enhanced with generated QA pairs for improved performance in biomedical tasks. This approach equips HuaTuo with more accurate medical knowledge, addressing the expertise gap in standard LLMs.
PDF Cite Code Project
HuaTuo: Tuning LLaMA Model with Chinese Medical Knowledge
From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for LowResource Molecule Discovery
In this paper, we propose a method to enhance in-silico molecule discovery in low-resource settings by using pseudo data generated by Large Language Models (LLMs), demonstrating its efficiency through improved performance, smaller model size, less data, and lower costs, with benefits increasing as more pseudo data is used.
PDF Cite Code
From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for LowResource Molecule Discovery
Manifold-based Verbalizer Space Re-embedding for Tuning-free Prompt-based Classification
In this paper, we introduce a tuning-free method, Locally Linear Embedding with Intra-class Neighborhood Constraint (LLE-INC), for improving prompt-based classification without adjusting parameters. It focuses on preserving local properties within the same class for better classification, showing results on par with tuned methods and enhancing performance by up to 3.2%. LLE-INC also proves efficient for hyper-scale language models like LLaMA-7B&13B.
PDF Cite
Manifold-based Verbalizer Space Re-embedding for Tuning-free Prompt-based Classification
Knowledge-tuning Large Language Models with Structured Medical Knowledge Bases for Reliable Response Generation in Chinese
In this paper, we introduce a method called knowledge-tuning, using structured medical knowledge bases to improve Large Language Models’ (LLMs) understanding and response accuracy in medical contexts, addressing the issue of hallucinated medical facts.
PDF Cite
Knowledge-tuning Large Language Models with Structured Medical Knowledge Bases for Reliable Response Generation in Chinese
UniCoRN: Unified Cognitive Signal ReconstructioN bridging cognitive signals and human language
In this paper, we introduce fMRI2text, a novel open-vocabulary task for converting fMRI time series to human language, and presents UniCoRN, a model that decodes text from fMRI and EEG signals better than the previous baseline, demonstrating the potential for real-world brain-computer interface applications.
PDF Cite Code
UniCoRN: Unified Cognitive Signal ReconstructioN bridging cognitive signals and human language
Global Prompt Cell: A Portable Control Module for Effective Prompt Tuning
In this paper, we introduce a new method to enhance prompt tuning in pre-trained models by using the Global Prompt Cell (GPC), a module that maintains prompt information throughout all encoder layers, leading to a 5.8% performance boost on SuperGLUE datasets.
PDF Cite
Global Prompt Cell: A Portable Control Module for Effective Prompt Tuning
Prompt Combines Paraphrase: Teaching Pre-trained Models to Understand Rare Biomedical Words
In this paper, we propose a method for improving pre-trained models’ performance on biomedical tasks by better incorporating rare biomedical terms during prompt-based fine-tuning, achieving up to a 6% improvement in tasks like biomedical natural language inference without extra resources.
PDF Cite Code
Prompt Combines Paraphrase: Teaching Pre-trained Models to Understand Rare Biomedical Words