Search

Your search keyword '"Zhou, Zhenhong"' showing total 33 results
Start Over Author "Zhou, Zhenhong"
33 results on '"Zhou, Zhenhong"'

1. Alignment-Enhanced Decoding:Defending via Token-Level Adaptive Refining of Probability Distributions

Author

Liu, Quan, Zhou, Zhenhong, He, Longzhu, Liu, Yi, Zhang, Wei, and Su, Sen

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence
Abstract

Large language models are susceptible to jailbreak attacks, which can result in the generation of harmful content. While prior defenses mitigate these risks by perturbing or inspecting inputs, they ignore competing objectives, the underlying cause of alignment failures. In this paper, we propose Alignment-Enhanced Decoding (AED), a novel defense that employs adaptive decoding to address the root causes of jailbreak issues. We first define the Competitive Index to quantify alignment failures and utilize feedback from self-evaluation to compute post-alignment logits. Then, AED adaptively combines AED and post-alignment logits with the original logits to obtain harmless and helpful distributions. Consequently, our method enhances safety alignment while maintaining helpfulness. We conduct experiments across five models and four common jailbreaks, with the results validating the effectiveness of our approach. Code is available at https://github.com/GIGABaozi/AED.git., Comment: 15 pages, 5 figures

Published
2024

2. Course-Correction: Safety Alignment Using Synthetic Preferences

Author

Xu, Rongwu, Cai, Yishuo, Zhou, Zhenhong, Gu, Renjie, Weng, Haiqin, Liu, Yan, Zhang, Tianwei, Xu, Wei, and Qiu, Han

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence, Computer Science - Machine Learning
Abstract

The risk of harmful content generated by large language models (LLMs) becomes a critical concern. This paper presents a systematic study on assessing and improving LLMs' capability to perform the task of \textbf{course-correction}, \ie, the model can steer away from generating harmful content autonomously. To start with, we introduce the \textsc{C$^2$-Eval} benchmark for quantitative assessment and analyze 10 popular LLMs, revealing varying proficiency of current safety-tuned LLMs in course-correction. To improve, we propose fine-tuning LLMs with preference learning, emphasizing the preference for timely course-correction. Using an automated pipeline, we create \textsc{C$^2$-Syn}, a synthetic dataset with 750K pairwise preferences, to teach models the concept of timely course-correction through data-driven preference learning. Experiments on 2 LLMs, \textsc{Llama2-Chat 7B} and \textsc{Qwen2 7B}, show that our method effectively enhances course-correction skills without affecting general performance. Additionally, it effectively improves LLMs' safety, particularly in resisting jailbreak attacks., Comment: Dataset and script will be available at https://github.com/pillowsofwind/Course-Correction

Published
2024

3. How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States

Author

Zhou, Zhenhong, Yu, Haiyang, Zhang, Xinghua, Xu, Rongwu, Huang, Fei, and Li, Yongbin

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence, Computer Science - Cryptography and Security, Computer Science - Computers and Society
Abstract

Large language models (LLMs) rely on safety alignment to avoid responding to malicious user inputs. Unfortunately, jailbreak can circumvent safety guardrails, resulting in LLMs generating harmful content and raising concerns about LLM safety. Due to language models with intensive parameters often regarded as black boxes, the mechanisms of alignment and jailbreak are challenging to elucidate. In this paper, we employ weak classifiers to explain LLM safety through the intermediate hidden states. We first confirm that LLMs learn ethical concepts during pre-training rather than alignment and can identify malicious and normal inputs in the early layers. Alignment actually associates the early concepts with emotion guesses in the middle layers and then refines them to the specific reject tokens for safe generations. Jailbreak disturbs the transformation of early unethical classification into negative emotions. We conduct experiments on models from 7B to 70B across various model families to prove our conclusion. Overall, our paper indicates the intrinsical mechanism of LLM safety and how jailbreaks circumvent safety guardrails, offering a new perspective on LLM safety and reducing concerns. Our code is available at https://github.com/ydyjya/LLM-IHS-Explanation., Comment: 27 pages

Published
2024

4. Speak Out of Turn: Safety Vulnerability of Large Language Models in Multi-turn Dialogue

Author

Zhou, Zhenhong, Xiang, Jiuyang, Chen, Haopeng, Liu, Quan, Li, Zherui, and Su, Sen

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence
Abstract

Large Language Models (LLMs) have been demonstrated to generate illegal or unethical responses, particularly when subjected to "jailbreak." Research on jailbreak has highlighted the safety issues of LLMs. However, prior studies have predominantly focused on single-turn dialogue, ignoring the potential complexities and risks presented by multi-turn dialogue, a crucial mode through which humans derive information from LLMs. In this paper, we argue that humans could exploit multi-turn dialogue to induce LLMs into generating harmful information. LLMs may not intend to reject cautionary or borderline unsafe queries, even if each turn is closely served for one malicious purpose in a multi-turn dialogue. Therefore, by decomposing an unsafe query into several sub-queries for multi-turn dialogue, we induced LLMs to answer harmful sub-questions incrementally, culminating in an overall harmful response. Our experiments, conducted across a wide range of LLMs, indicate current inadequacies in the safety mechanisms of LLMs in multi-turn dialogue. Our findings expose vulnerabilities of LLMs in complex scenarios involving multi-turn dialogue, presenting new challenges for the safety of LLMs., Comment: working in progress 23pages, 18 figures

Published
2024

5. Quantifying and Analyzing Entity-level Memorization in Large Language Models

Author

Zhou, Zhenhong, Xiang, Jiuyang, Chen, Chaomeng, and Su, Sen

Subjects
Computer Science - Computation and Language
Abstract

Large language models (LLMs) have been proven capable of memorizing their training data, which can be extracted through specifically designed prompts. As the scale of datasets continues to grow, privacy risks arising from memorization have attracted increasing attention. Quantifying language model memorization helps evaluate potential privacy risks. However, prior works on quantifying memorization require access to the precise original data or incur substantial computational overhead, making it difficult for applications in real-world language models. To this end, we propose a fine-grained, entity-level definition to quantify memorization with conditions and metrics closer to real-world scenarios. In addition, we also present an approach for efficiently extracting sensitive entities from autoregressive language models. We conduct extensive experiments based on the proposed, probing language models' ability to reconstruct sensitive entities under different settings. We find that language models have strong memorization at the entity level and are able to reproduce the training data even with partial leakages. The results demonstrate that LLMs not only memorize their training data but also understand associations between entities. These findings necessitate that trainers of LLMs exercise greater prudence regarding model memorization, adopting memorization mitigation techniques to preclude privacy violations., Comment: 9 pages, 7 figures

Published
2023

6. Temporal and Spatial Evolution Characteristics of Carbon Storage in Hefei Ecosystem Based on PLUS and InVEST Models

Author

ZHI Fei, ZHOU Zhenhong, ZHAO Ming, and WANG Shiqi

Subjects
invest model, plus model, carbon stocks, composite index of land use, Environmental sciences, GE1-350, Agriculture
Abstract

[Objective] To seek a new urban development plan of Hefei under the guidance of “double carbon” goal. [Methods] Based on the land use data of Hefei City in the 5th period from 2000 to 2020, under the guidance of territorial spatial Planning of Hefei City, the spatio-temporal evolution law of land use in 2035 was simulated by the PLUS model, and the spatio-temporal change characteristics of carbon storage in Hefei City under multiple scenarios were explored by coupling the InVEST model. And further explore the effect of the comprehensive land use on the carbon storage. [Results] (1) The main characteristics of land use change in Hefei City from 2000 to 2020 were the decrease of cultivated land and forest land, among which cultivated land was the main source of construction land expansion. The law of land change in the natural development and farmland resource protection scenarios was roughly the same, mainly represented by the reduction of cultivated land, forest land and water body. In the green urban development scenario, the area of forest land changed from decreasing to increasing compared with the other two scenarios. (2) Carbon storage in Hefei City decreased year by year from 2000 to 2020, and the carbon loss was the most 2005 to 2010. By 2035, the carbon reserves of natural development scenario, farmland resource protection scenario and green urban development scenario were 138.96×106 t, 140.13×106 t, and 139.81×106 t, respectively. Under the farmland resource protection scenario, the area with the lowest increase of carbon storage significantly slowed down the expansion of the construction land. under the scenario of urban development, forest land changes from carbon loss to carbon sequestration, which was the most potential trend of carbon sequestration. (3) The land utilization rate of green urban development scenario was the highest, which can effectively reduce the threat of comprehensive land use degree to carbon storage loss. [Conclusion] Both farmland resource protection scenario and green accumulation urban development scenario were conducive to urban carbon sequestration development. Implementing parallel policies of ecological protection and urban development, and adjusting comprehensive land use mode were conducive to improving urban carbon loss.

Published
2024
Full Text
View/download PDF

12. Mediating role of psychological resilience in the relationship between self-efficacy and professional identity among nurses.

Author

Yi, Ruonan, Zhou, Zhenhong, Ma, Wenyuan, Yang, Chunling, Wang, Fengling, and Wu, Jianghua

Abstract

To explore the mediating role of resilience in the relationship between general self-efficacy and professional identity of nurses during the COVID-19 pandemic. A cross-sectional design was employed. A total of 982 nurses from four Grade III, class A hospitals in Shandong Province were investigated using general information questionnaire, nurses' professional identity rating scale, general self-efficacy scale (GSES), and Connor-Davidson flexibility scale (CD-RISC). SPSS22.0 and Amos21.0 were used for data analysis and structural equation modeling. p % counseling The nurses had a score of （27.038±5.933） for general self-efficacy score, 38.290±6.234 for psychological resilience, and （114.99±16.209） for professional identity. A positive correlation between general self-efficacy, professional identity, and psychological resilience (<0.01) was found. The SEM analysis shows that psychological resilience plays a mediating role between general self-efficacy and professional identity. The ratio of the effect is 75.155. During the COVID-19 pandemic, the levels of general self-efficacy and professional identity of nurses was medium, while psychological resilience was high. Nurses' general self-efficacy can affect their professional identity through psychological resilience. During the pandemic, the psychological status of nurses should not be ignored. Nursing managers should fully utilize of group and cognitive therapy based on mindfulness to improve nurses' psychological resilience and general self-efficacy, and to promote nurses' professional identity, so as to ensure the lower turnover rate. [ABSTRACT FROM AUTHOR]

Published
2024
Full Text
View/download PDF

13. Residual regression on laser welding of CFRP-metal hybrid structures.

Author

Zhou, Zhenhong, Gao, Xiangdong, Ma, Bo, and Zhang, Yanxi

Subjects
*LASER welding, *HIGH strength steel welding, *CARBON steel
Abstract

The laser welding between high strength steel and carbon fiber reinforced polymer (CFRP) was explored in order to improve the mechanical properties and capacity to reflect physical relationships. The influence of process parameters on the mechanical properties of the joint was analyzed by RSM based on the experimental results. A residual regression method was proposed, using the residual block as the core structure and combined it with the fundamental concept and structure of the residual network. The residual regression model with two-layer residual block structure was compared with the RSM model at the same time, and the feasibility of this method was validated by experiments. The results show that the mechanical properties of the hybrid joint increase with the increase of the input laser energy per unit time. The proposed residual regression model can directly characterize the relationship between process parameters, intermediate physical quantities, and shear force and has better prediction accuracy and more physical relationships than RSM. The optimized parameters are laser power 1100 W, welding speed 9 mm/s and defocus distance 15 mm. [ABSTRACT FROM AUTHOR]

Published
2024
Full Text
View/download PDF

15. Temporal and Spatial Evolution Characteristics of Carbon Storage in Hefei Ecosystem Based on PLUS and In VEST Models.

Author

ZHI Fei, ZHOU Zhenhong, ZHAO Ming, and WANG Shiqi

Abstract

[Objective] To seek a new urban development plan of Hefei under the guidance of "double carbon" goal. [Methods] Based on the land use data of Hefei City in the 5th period from 2000 to 2020, under the guidance of territorial spatial Planning of Hefei City, the spatio-temporal evolution law of land use in 2035 was simulated by the PLUS model, and the spatio-temporal change characteristics of carbon storage in Hefei City under multiple scenarios were explored by coupling the InVEST model. And further explore the effect of the comprehensive land use on the carbon storage. [Results] (1) The main characteristics of land use change in Hefei City from 2000 to 2020 were the decrease of cultivated land and forest land, among which cultivated and was the main source of construction land expansion. The law of land change in the natural development and farmland resource protection scenarios was roughly the same, mainly represented by the reduction of cultivated land, forest land and water body. In the green urban development scenario, the area of forest land hanged from decreasing to increasing compared with the other two scenarios. (2) Carbon storage in Hefei City decreased year by year from 2000 to 2020, and the carbon loss was the most 2005 to 2010. By 2035, the carbon reserves of natural development scenario, farmland resource protection scenario and green urban development scenario were 138.96 × 106 t, 140.13 × 106 t, and 139.81 × 106 t, respectively. Under the farmland resource protection scenario, the area with the lowest increase of carbon storage significantly slowed down the expansion of the construction land. under the scenario of urban development, forest land changes from carbon loss to carbon sequestration, which was the most potential trend of carbon sequestration. (3) The land utilization rate of green urban development scenario was the highest, which can effectively reduce the threat of comprehensive land use degree to carbon storage loss. [Conclusion] Both farmland resource protection scenario and green accumulation urban development scenario were conducive to urban carbon sequestration development. Implementing parallel policies of ecological protection and urban development, and adjusting comprehensive land use mode were conducive to improving urban carbon loss. [ABSTRACT FROM AUTHOR]

Published
2024
Full Text
View/download PDF

19. Development of a rapid neutralization testing system for Rhinovirus C15 based on the enzyme-linked immunospot assay

Author

Zhou, Zhenhong, primary, Zhu, Rui, additional, Yang, Hongwei, additional, Xu, Longfa, additional, Chen, Hao, additional, Wu, Yuanyuan, additional, Yin, Zhichao, additional, Huang, Qiongzi, additional, Zhang, Dongqing, additional, Liu, Che, additional, Que, Yuqiong, additional, Zhang, Jun, additional, Xia, Ningshao, additional, and Cheng, Tong, additional

Published
2022
Full Text
View/download PDF

20. Structural basis for the synergistic neutralization of coxsackievirus B1 by a triple-antibody cocktail

Author

Zheng, Qingbing, primary, Zhu, Rui, additional, Yin, Zhichao, additional, Xu, Longfa, additional, Sun, Hui, additional, Yu, Hai, additional, Wu, Yuanyuan, additional, Jiang, Yichao, additional, Huang, Qiongzi, additional, Huang, Yang, additional, Zhang, Dongqing, additional, Liu, Liqin, additional, Yang, Hongwei, additional, He, Maozhou, additional, Zhou, Zhenhong, additional, Jiang, Yanan, additional, Chen, Zhenqin, additional, Zhao, Huan, additional, Que, Yuqiong, additional, Kong, Zhibo, additional, Zhou, Lizhi, additional, Li, Tingting, additional, Zhang, Jun, additional, Luo, Wenxin, additional, Gu, Ying, additional, Cheng, Tong, additional, Li, Shaowei, additional, and Xia, Ningshao, additional

Published
2022
Full Text
View/download PDF

27. Effects of Sc on laser hot-wire welding performance of 7075 aluminum alloy

Author

Fei Su, Zhou Zhenhong, Wei Xu, Gang Xiao, Jianghua Feng, and Shichun Li

Subjects
Materials science, Polymers and Plastics, Metallurgy, Alloy, Metals and Alloys, chemistry.chemical_element, Welding, engineering.material, Laser, Surfaces, Coatings and Films, Electronic, Optical and Magnetic Materials, law.invention, Biomaterials, chemistry, Aluminium, law, engineering
Abstract

In order to clarify the effects of Sc modification on laser hot-wire welding performance of 7075 high strength aluminum alloys, the changing rules of weld formation, microstructure, element loss, mechanical performances and pores of weld joints were investigated during laser fusion welding, laser hot-wire welding and Sc-modified laser hot-wire welding. Comparing with laser fusion welding, the filler wire in laser hot-wire welding brought welding material and hot wire energy, resulting in better weld formation and sufficient fused weld joint. In laser hot-wire welding, tensile strength and elongation of weld joint were enhanced, element loss of weld joint was dramatically reduced and weld pore was reduced slightly. However, grains in fusion zone were relatively coarse, which caused lower microhardness in laser hot-wire welding. In Sc modified laser hot-wire welding, Sc modification led to smooth and continuous weld joint. During solidification of weld joints, Sc could combine with base metal to form Sc-containing alloy phase, which could lead to the formation of refined and equiaxed grains, thereby enhancing the yield strength of microstructure and increasing the microhardness, tensile strength and elongation of weld joint. In addition, Sc modification could increase the fluidity of molten pool, which thus promoted the escape of bubbles from the molten pool and hindered the formation of weld pores.

Published
2020

28. A New 2-Micrometer Continuous Wave Laser Method for Management of the Distal Ureter in Retroperitoneal Laparoscopic Nephroureterectomy

Author

Xu Zhang, Jun Dong, Zhou Zhenhong, Gang Guo, and Yong Yang

Subjects
Male, medicine.medical_specialty, Urology, Operative Time, Blood Loss, Surgical, Nephrectomy, Disease-Free Survival, Ureter, Blood loss, Humans, Medicine, Kidney Pelvis, Aged, Retrospective Studies, Upper urinary tract, Laparoscopic nephroureterectomy, Carcinoma, Transitional Cell, Ureteral Neoplasms, business.industry, Cystoscopy, Middle Aged, Standard technique, Distal ureter, Kidney Neoplasms, Surgery, medicine.anatomical_structure, Cuff, Operative time, Female, Laparoscopy, Laser Therapy, business
Abstract

Retroperitoneal laparoscopic nephroureterectomy (LNU) is a classic technique for the treatment of patients with urothelial cancer of the upper urinary tract. Management of the distal ureter in patients undergoing this procedure remains controversial, however. We evaluated a new method involving cystoscopic sleeve resection of the distal ureter using a 2-μm continuous wave laser before LNU.LNU was performed in 38 patients at our hospital between January 2008 and January 2012. The distal end of the ureter was managed by using a 2-μm continuous wave laser to evaporate and excise the ureteral orifice and bladder cuff. A standard technique was applied to keep the ureter intact. The ureter was dissected to the entrance of the bladder, and the distal ureter was easily detached from the bladder.The distal ureter was excised completely in all cases using this new technique. The average operative time was 2.4 hours, and the average blood loss was 69.4 mL. At 1 year after surgery, the cause-specific survival rate of the patients was 100%, the bladder recurrence-free rate was 89%, and the extravesical recurrence-free rate was 100%.For managing the distal ureter during LNU, the new technique described proved to be a simple, safe, and minimally invasive method.

Published
2015

29. AB124. A new 2-micrometer continuous wave laser method for management of the distal ureter in retroperitoneal laparoscopic nephroureterectomy

Author

Gang, Guo, Yong, Yang, Jun, Dong, Zhou, Zhenhong, and Xu, Zhang

Subjects
upper urinary tract, Moderated Poster Presentation, Retroperitoneal laparoscopic nephroureterectomy (LNU)
Abstract

Background and purpose Retroperitoneal laparoscopic nephroureterectomy (LNU) is a classic technique for the treatment of patients with urothelial cancer of the upper urinary tract. Management of the distal ureter in patients undergoing this procedure remains controversial, however. We evaluated a new method involving cystoscopic sleeve resection of the distal ureter using a 2-μm continuous wave laser before LNU. Patients and methods LNU was performed in 38 patients at our hospital between January 2008 and January 2012. The distal end of the ureter was managed by using a 2-μm continuous wave laser to evaporate and excise the ureteral orifice and bladder cuff. A standard technique was applied to keep the ureter intact. The ureter was dissected to the entrance of the bladder, and the distal ureter was easily detached from the bladder. Results The distal ureter was excised completely in all cases using this new technique. The average operative time was 2.4 hours, and the average blood loss was 69.4 mL. At 1 year after surgery, the cause-specific survival rate of the patients was 100%, the bladder recurrence-free rate was 89%, and the extravesical recurrence-free rate was 100%. Conclusions For managing the distal ureter during LNU, the new technique described proved to be a simple, safe, and minimally invasive method.

Published
2015

31. Nonlinearity and slow relaxation effects of piezoelectric ceramics

Author

Li Xiaoling, Chen Rong, Li Jing-De, Fan Yangcai, and Zhou Zhenhong

Subjects
Piezoelectric coefficient, Materials science, Electrostriction, Condensed matter physics, Field (physics), Doping, Condensed Matter Physics, Piezoelectricity, Computer Science::Other, Electronic, Optical and Magnetic Materials, Condensed Matter::Materials Science, Nonlinear system, Interferometry, Relaxation (physics)
Abstract

Using an interferometer the strain S of doped PZT ceramics in an applied field E is measured. When E is an alternating field, an asymmetrical piezoelectric loop is observed. The piezoelectric response has a fast component and a slow one. The response is nonlinear with a positive quadtic term, which is different from electrostriction effect. The stochastic and free relaxation effect are observed for a same sample.

Published
1997

32. Time Domain Spectroscopy Function and Line Shape of Slow Responses

Author

Tang Xingui, Fan Yangcai, Li Jing-De, and Zhou Zhenhong

Subjects
Condensed Matter::Materials Science, Materials science, Optics, business.industry, General Physics and Astronomy, High resolution, Function (mathematics), Time domain, Dielectric, business, Spectroscopy, Piezoelectricity, Line (formation)
Abstract

The spectroscopy function and line shape of slow responses are defined in time domain. In dielectric and piezoelectric relaxations experimental results show that this method has high resolution.

Published
1995

Catalog

Books, media, physical & digital resources
See catalog results