arXiv subjects

Fufang Wen

Publications and source records attributed to Fufang Wen.

6 recordsLinked to original sources

arXiv2025

Retention analysis of edited knowledge after fine-tuning

Large language models (LLMs) store vast amounts of knowledge, which often requires updates to correct factual errors, incorporate newly acquired information, or adapt model behavior. Model editing methods have emerged as efficient solutions for such updates, offering localized and precise knowledge modification at significantly lower computational cost than continual training. In parallel, LLMs are frequently fine-tuned for a wide range of downstream tasks. However, the effect of fine-tuning on previously edited knowledge remains poorly understood. In this work, we systematically investigate how different fine-tuning objectives interact with various model editing techniques. Our findings show that edited knowledge is substantially more susceptible to forgetting during fine-tuning than intrinsic knowledge acquired through pre-training. This analysis highlights a key limitation of current editing approaches and suggests that evaluating edit robustness under downstream fine-tuning is critical for their practical deployment. We further find that knowledge retention can be significantly improved by either augmenting edit knowledge with paraphrases or by freezing layers associated with edited content in fine-tuning stage, offering insight for developing more robust editing algorithms.

cs.CL↗

arXiv2025

FinCPRG: A Bidirectional Generation Pipeline for Hierarchical Queries and Rich Relevance in Financial Chinese Passage Retrieval

In recent years, large language models (LLMs) have demonstrated significant potential in constructing passage retrieval datasets. However, existing methods still face limitations in expressing cross-doc query needs and controlling annotation quality. To address these issues, this paper proposes a bidirectional generation pipeline, which aims to generate 3-level hierarchical queries for both intra-doc and cross-doc scenarios and mine additional relevance labels on top of direct mapping annotation. The pipeline introduces two query generation methods: bottom-up from single-doc text and top-down from multi-doc titles. The bottom-up method uses LLMs to disassemble and generate structured queries at both sentence-level and passage-level simultaneously from intra-doc passages. The top-down approach incorporates three key financial elements--industry, topic, and time--to divide report titles into clusters and prompts LLMs to generate topic-level queries from each cluster. For relevance annotation, our pipeline not only relies on direct mapping annotation from the generation relationship but also implements an indirect positives mining method to enrich the relevant query-passage pairs. Using this pipeline, we constructed a Financial Passage Retrieval Generated dataset (FinCPRG) from almost 1.3k Chinese financial research reports, which includes hierarchical queries and rich relevance labels. Through evaluations of mined relevance labels, benchmarking and training experiments, we assessed the quality of FinCPRG and validated its effectiveness as a passage retrieval dataset for both training and benchmarking.

cs.IR↗

arXiv2025

FinBERT2: A Specialized Bidirectional Encoder for Bridging the Gap in Finance-Specific Deployment of Large Language Models

In natural language processing (NLP), the focus has shifted from encoder-only tiny language models like BERT to decoder-only large language models(LLMs) such as GPT-3. However, LLMs' practical application in the financial sector has revealed three limitations: (1) LLMs often perform worse than fine-tuned BERT on discriminative tasks despite costing much higher computational resources, such as market sentiment analysis in financial reports; (2) Application on generative tasks heavily relies on retrieval augmented generation (RAG) methods to provide current and specialized information, with general retrievers showing suboptimal performance on domain-specific retrieval tasks; (3) There are additional inadequacies in other feature-based scenarios, such as topic modeling. We introduce FinBERT2, a specialized bidirectional encoder pretrained on a high-quality, financial-specific corpus of 32b tokens. This represents the largest known Chinese financial pretraining corpus for models of this parameter size. As a better backbone, FinBERT2 can bridge the gap in the financial-specific deployment of LLMs through the following achievements: (1) Discriminative fine-tuned models (Fin-Labelers) outperform other (Fin)BERT variants by 0.4%-3.3% and leading LLMs by 9.7%-12.3% on average across five financial classification tasks. (2) Contrastive fine-tuned models (Fin-Retrievers) outperform both open-source (e.g., +6.8\% avg improvement over BGE-base-zh) and proprietary (e.g., +4.2\% avg improvement over OpenAI's text-embedding-3-large) embedders across five financial retrieval tasks; (3) Building on FinBERT2 variants, we construct the Fin-TopicModel, which enables superior clustering and topic representation for financial titles. Our work revisits financial BERT models through comparative analysis with contemporary LLMs and offers practical insights for effectively utilizing FinBERT in the LLMs era.

cs.IR↗

arXiv2019

Progressive-Growing of Generative Adversarial Networks for Metasurface Optimization

Generative adversarial networks, which can generate metasurfaces based on a training set of high performance device layouts, have the potential to significantly reduce the computational cost of the metasurface design process. However, basic GAN architectures are unable to fully capture the detailed features of topologically complex metasurfaces, and generated devices therefore require additional computationally-expensive design refinement. In this Letter, we show that GANs can better learn spatially fine features from high-resolution training data by progressively growing its network architecture and training set. Our results indicate that with this training methodology, the best generated devices have performances that compare well with the best devices produced by gradient-based topology optimization, thereby eliminating the need for additional design refinement. We envision that this network training method can generalize to other physical systems where device performance is strongly correlated with fine geometric structuring.

physics.comp-ph↗

arXiv2019

Superfluid Onset and Compressibility of $^4$He Films Adsorbed on Carbon Nanotubes

Third sound measurements of superfluid $^4$He thin films adsorbed on 10 nm diameter multiwall carbon nanotubes are used to probe the superfluid onset temperature as a function of the film thickness, and to study the temperature dependence of the film compressibility. The nanotubes provide a highly ordered carbon surface, with layer-by-layer growth of the adsorbed film as shown by oscillation peaks in the third sound velocity at the completion of the third, fourth, and fifth atomic layers, arising from oscillations in the compressibility. In temperature sweeps the third sound velocity at very low temperatures is found to be linear with temperature, but oscillating between positive and negative slope depending on the film thickness. Analysis shows that this can be attributed to a linearly decreasing compressibility of the film with temperature that appears to hold even near zero temperature. The superfluid onset temperature is found to be linear in the film thickness, as predicted by the Kosterlitz-Thouless theory, but the slope is anomalous, a factor of three smaller than the predicted universal value.

cond-mat.other↗

arXiv2017

Event-shape-engineering study of charge separation in heavy-ion collisions

Recent measurements of charge-dependent azimuthal correlations in high-energy heavy-ion collisions have indicated charge-separation signals perpendicular to the reaction plane, and have been related to the chiral magnetic effect (CME). However, the correlation signal is contaminated with the background caused by the collective motion (flow) of the collision system, and an effective approach is needed to remove the flow background from the correlation. We present a method study with simplified Monte Carlo simulations and a multi-phase transport model, and develop a scheme to reveal the true CME signal via the event-shape engineering with the flow vector of the particles of interest.

nucl-th↗