Kevin Wang, Alexandre Variengien, Arthur Conmy
et al.
Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior"in the wild"in a language model. We evaluate the reliability of our explanation using three quantitative criteria--faithfulness, completeness and minimality. Though these criteria support our explanation, they also point to remaining gaps in our understanding. Our work provides evidence that a mechanistic understanding of large ML models is feasible, opening opportunities to scale our understanding to both larger models and more complex tasks.
Large Language Models (LLMs) have garnered remarkable advancements across diverse code-related tasks, known as Code LLMs, particularly in code generation that generates source code with LLM from natural language descriptions. This burgeoning field has captured significant interest from both academic researchers and industry professionals due to its practical significance in software development, e.g., GitHub Copilot. Despite the active exploration of LLMs for a variety of code tasks, either from the perspective of Natural Language Processing (NLP) or Software Engineering (SE) or both, there is a noticeable absence of a comprehensive and up-to-date literature review dedicated to LLM for code generation. In this survey, we aim to bridge this gap by providing a systematic literature review that serves as a valuable reference for researchers investigating the cutting-edge progress in LLMs for code generation. We introduce a taxonomy to categorize and discuss the recent developments in LLMs for code generation, covering aspects such as data curation, latest advances, performance evaluation, ethical implications, environmental impact, and real-world applications. In addition, we present a historical overview of the evolution of LLMs for code generation and provide a quantitative and qualitative comparative analysis of experimental results of code LLMs, sourced from their original papers to ensure a fair comparison on the HumanEval, MBPP, and BigCodeBench benchmarks, across various levels of difficulty and types of programming tasks, to highlight the progressive enhancements in LLM capabilities for code generation. We identify critical challenges and promising opportunities regarding the gap between academia and practical development. Furthermore, we have established a dedicated resource GitHub page (https://github.com/juyongjiang/CodeLLMSurvey) to continuously document and disseminate the most recent advances in the field.
As the capabilities of large language models (LLMs) continue to advance, evaluating their performance becomes increasingly crucial and challenging. This paper aims to bridge this gap by introducing CMMLU, a comprehensive Chinese benchmark that covers various subjects, including natural science, social sciences, engineering, and humanities. We conduct a thorough evaluation of 18 advanced multilingual- and Chinese-oriented LLMs, assessing their performance across different subjects and settings. The results reveal that most existing LLMs struggle to achieve an average accuracy of 50%, even when provided with in-context examples and chain-of-thought prompts, whereas the random baseline stands at 25%. This highlights significant room for improvement in LLMs. Additionally, we conduct extensive experiments to identify factors impacting the models' performance and propose directions for enhancing LLMs. CMMLU fills the gap in evaluating the knowledge and reasoning capabilities of large language models within the Chinese context.
To evaluate the bonding behavior between ribbed steel bars and ultra-high-performance concrete (UHPC), several pull-out samples were experimentally investigated and presented. The influences of bar diameter and embedment length on the bond failure mode, ultimate bond stress, stress?slip curve and bond stress distribution were investigated. Compared with that of ordinary reinforced concrete, the ultimate bond stress between steel bars and UHPC was increased by 74% under the same conditions. In addition, the bond-slip curve showed a relatively large slope before the ultimate bond stress was reached, and at the same time, the curve showed better ductility after the ultimate bond stress was exceeded. When the sample experiences steel bar yielding or pull-out failure, the bond stress distribution curve exhibits typical multipeak features. As the bar embedment length increased, the bond stress distribution became increasingly heterogeneous, and the increase in the peak number in the bond stress distribution curve reflected this trend. Under appropriate concrete cover thickness, the critical embedment length was between 4d and 6d.
The adhesion work and force of the five types of asphalt to andesite were measured using surface energy and pull-out tests, respectively. Various tests were operated to estimate the high and low-temperature characteristics of the five types of asphalt mixtures: high-temperature rutting, low-temperature bending, indirect tensile creep, and thermal stress restrained specimen tests. Furthermore, the high and low-temperature evaluation index, adhesion work, and adhesion force of these asphalt mixtures were analyzed using grey correlation analysis. The results show that compared with the adhesion work, the actual contact between mineral aggregate and asphalt in the pull-out test leads to the difference in adhesion force and work measured by SBS-modified, SBR-modified, and extreme cold asphalt. The correlation degree of adhesion work and force to the evaluation indexes of high and low-temperature properties in asphalt mixture is roughly the same, and the adhesion work has a more momentous impact on the high and low-temperature properties in asphalt mixtures. The adhesion work and force correlate best with low-temperature bending test parameters and thermal stress restrained specimen test (TSRST) parameters related to fracture failure of asphalt mixture, followed by the indirect tensile creep test parameters. It shows that the adhesion work and force significantly impact the low-temperature characteristics of asphalt mixture more than their high-temperature stability.
Materials of engineering and construction. Mechanics of materials
CFRP has the potential to replace steel cables in large-span cable-stayed bridges due to its high strength and lightweight material properties. However, the weak lateral force performance of CFRP material creates the challenge of anchoring. This study introduces a new inner cone + arc + straight cylinder bond-type anchorage system to optimize CFRP tendons’ force state. Experimental and finite element analyses verified the new anchoring system’s performance. In static load tensile tests, six groups of seven CFRP tendon anchorage systems with different sleeve grooves were used to study the failure mode and load–strain variation law. The difference in mechanical properties between the new and traditional anchorage is evaluated in the finite element analysis. The results indicate that the new anchorage system can lower the stress concentration in the anchorage zone and enhance anchorage performance. The groove design of the sleeve can effectively increase the anchoring efficiency, where the groove depth is proportional to the anchoring efficiency and the groove spacing is inversely proportional to the anchoring efficiency. The magnitude of the stress inhomogeneity in the multi-tendon anchoring system during tensioning is proportional to the beginning conditions and the load size. When the inner wall of the sleeve becomes more abrasive, the force heterogeneity of the anchorage system reduces. The tests and finite element analysis show that the new anchoring may improve stress distribution and anchorage efficiency. In engineering practice, it can be utilized as a dependable anchorage system.
Abstract Bridge deck condition assessments are typically conducted through visual inspections and by incorporating traditional contact sensors for Non-Destructive Evaluation techniques such as hammer sounding and chain dragging, which require the keen expertise of trained inspectors. The accuracy of these inspections is proportional to the level of deterioration of the bridge deck, as the ability of the inspectors is correlated to the apparent level of damage. This study aims to improve the accuracy of bridge deck inspection processes by utilizing non-destructive evaluation techniques, including analyzing point cloud data gathered via Light Detection and Ranging (LiDAR) as a geometry-capturing tool for identifying surface irregularities. This research aims to evaluate and quantify the effectiveness and efficiency of LiDAR sensors in contributing to the suite of technologies available to perform bridge deck condition assessment. To achieve this, the research proposes to understand the deterioration pattern of New Jersey bridges, evaluate the results gathered from point cloud data collected on a full-scale bridge deck, and quantify the information gained from deploying LiDAR on operating bridges in New Jersey. Two data processing approaches were chosen to measure the gross and fine dimensions of the evaluated bridge decks, such as the Curvature Extraction and Slope Analysis method, and the Least Square Plane Fitting method, resulting in an accuracy of 97.92% in reference to the results gathered from reports generated through the analysis of state-of-the-art NDE technology data and visual inspection.
The history of bridge construction is an important part of historical knowledge. Developments in bridge construction technology reflect not only engineering advances, but also social, economic and cultural aspects of society. Engineers and scientists faced unique challenges when designing and building bridges depending on the technological level of the era, available materials and the needs of society. This process may reflect technological progress, changes in transportation needs, and cultural and social changes. The purpose of this article is to briefly review key moments and stages in the history of metal bridge construction using welding technology in the 20th century. The history of the development of the construction of metal bridges using welding goes back a little over 100 years. The short period from the construction of the first welded bridges to their first disasters led to the need to analyze the possible causes of these destructions. As the analysis performed showed, catastrophic destruction most often occurred under the influence of several factors, as well as a combination of external adverse influences and the internal “unpreparedness” of the structure for them. The above examples indicate that an irrational choice of steel could be both an independent cause causing brittle failure of structures, and an aggravating factor in the presence of structural violations, thermal stresses and welding defects. Over the years, bridge manufacturing technologies have been improved in different countries, and new steels and materials for their welding have been developed. Thanks to the use of carbon, low-alloy and alloy steel, designers abandoned the brutal “railroad-type” beam trusses and today metal bridges with graceful and beautiful silhouettes powerfully stride across the water surface, mountains and valleys. They became real attractions of megacities and country landscapes, and builders were able to successfully solve numerous technical and economic problems. An important contribution to the development of global bridge construction using welding technologies was made by the team of the Institute of Electric Welding of the Academy of Sciences of the Ukrainian SSR under the leadership of Academician Evgeny Oskarovych Paton. The team of the Institute of Electric Welding of the Academy of Sciences of the Ukrainian SSR, introducing welding into bridge construction, carefully checked the results and monitored the behavior of structures. A new grade of steel was created that was resistant to the formation of brittle and fatigue cracks, its welding technology was developed, a technology for installation welding of vertical sheets with forced formation of a seam was developed, and suitable welding materials were selected. At the time of construction in 1953, the Kyiv Evgeny Paton Bridge across the Dnipro River was the largest all-welded bridge in Europe, all seams of which, including assembly ones, were made for the first time using automatic and semi-automatic welding. In addition, the presence of large similar blocks in the design of the Evgeny Paton Bridge made it possible to mechanize assembly and welding operations and organize an in-line method for their production at the factory and installation, which improved the quality of welding work and reduced its labor intensity.
History (General) and history of Europe, Science (General)
Abstract: The level of traffic noise depends mainly on road traffic parameters such as: traffic volume, share of trucks in traffic and vehicles speed. Traffic conditions determined by such parameters as, for example, the average travel speed, average delays at the intersection and the number of vehicles stops are also important. Their particular importance is revealed especially in the case of bad traffic conditions and a significant reduction of traffic flow. The paper presents selected methods of measuring road traffic parameters, and in particular - traffic conditions on the example of research on their impact on the noise level in the vicinity of the street and intersections. The results of these tests are also presented, determining the impact of such parameters on traffic noise as: instantaneous speed, sectional speed, average delays at the intersection entry, the number of vehicle stops, volume to capacity of the entry ratio. Keywords: Traffic noise; Traffic conditions; Traffic parameters; Measurement methods
Highway engineering. Roads and pavements, Bridge engineering
A ballastless track is susceptible to damage and even failure of structural components under the long-term effects of extremely high temperatures. In this paper, considering the influence of different constraint boundaries, a 1 : 4 scaled model of a ballastless track-bridge structural system was produced and placed in a large-sized environmental chamber. The thermal performance of the track structure was studied by carrying out temperature loading tests at extreme temperatures. In combination with the scaled-down model test data, a 3D nonlinear finite element model was established to investigate the damage evolution of broad-narrow joints under temperature gradient loading. The results are shown as follows: (1) the cement asphalt (CA) mortar layer has a hysteretic effect on the vertical temperature transfer. The most unfavorable structural part is between the track slab and the CA mortar layer of the track structure. (2) The constraint conditions accelerate the rate of temperature transfer, creating disturbances to the internal stresses of the track structure and amplifying the internal stresses induced by the environmental temperature increase. (3) Temperature and longitudinal stresses in the track’s structural layers are highly correlated. There is a significant quadratic regression between internal temperature and structural stresses in different extreme high-temperature environments. (4) As the temperature gradient load increases, the damage occurs at the junction of the broad-narrow joint and tends to expand towards the ends, which has little effect on the compression damage of the broad-narrow joint but significantly increases the tension damage. The research could provide useful guidance for the scientific operation and maintenance of the ballastless track in extreme high-temperature environments.
Materials of engineering and construction. Mechanics of materials
Cement-based materials are widely utilized in infrastructure. The main product of hydrated products of cement-based materials is calcium silicate hydrate (C-S-H) gels that are considered as the binding phase of cement paste. C-S-H gels in Portland cement paste account for 60–70% of hydrated products by volume, which has profound influence on the mechanical properties and durability of cement-based materials. The preparation method of C-S-H gels has been well documented, but the quality of the prepared C-S-H affects experimental results; therefore, this review studies the preparation method of C-S-H under different conditions and materials. The progress related to C-S-H microstructure is explored from the theoretical and computational point of view. The fractality of C-S-H is discussed. An evaluation of the mechanical properties of C-S-H has also been included in this review. Finally, there is a discussion of the durability of C-S-H, with special reference to the carbonization and chloride/sulfate attacks.
The accuracy of seismic demand models in seismic vulnerability analysis of structures or components mainly depends on the seismic intensity measures (IMs) and engineering demand parameters (EDPs). This paper proposes a novel method to obtain the optimal seismic demand model for the seismic vulnerability analysis of bridges. The method obtains the IM-EDP combination by matching all IMs and EDPs within a wide range one by one, considering the contribution of multiple IM parameters to the seismic response of the structure and avoiding the blindness of EDP selection. The IM is determined by calculating Pearson correlation coefficient and partial correlation coefficient, controlling the correlation between EDP and IM (or IMs) to a minimum to reduce the multicollinearity within the vector IMs and avoid ill-conditioned models. The optimal seismic demand model is obtained by inspecting the scatter plot and residual plot of suboptimal seismic demand models determined from all combinations by guaranteeing efficiency and sufficiency. The efficiency of seismic demand models is guaranteed by controlling the root mean square error (RMSE) and the coefficient of determination (R2). The sufficiency of models is guaranteed by controlling the slope of fitted line. A continuous rigid frame bridge with double thin-walled piers is used as a case study and a dynamic time-history analysis is performed to obtain the seismic vulnerability of bridge with the proposed method. The results show that the proposed method is feasible and ideally suited for optimizing seismic demand model.
Mauro Coni, Fausto Mistretta, Flavio Stochino
et al.
In this paper, the use of Fast Falling Weight Deflectometer (Fast-FWD) is analyzed as a non-destructive and quick test procedure to evaluate the efficiency of short-span bridges. The Fast-FWD is an instrument that can produce a broadband dynamic force up to an impact value of 120 KN: The impact is constant and replicable, providing accurate action measures of bridge stiffness in a truly short period (30 ms). In this paper, a single-span reinforced concrete bridge is investigated, using the Fast-FWD. The considered bridge, approximately 12.0 m long and 15.5 m wide, was in critical condition. The bridge is in a suburban principal road near to the City of Cagliari in Sardinia (Italy), with an Annual Average Daily Traffic of 13,500 vehicles/day, and was suddenly closed, creating serious problems for urban mobility. In these conditions, the investigation through other standard techniques is time-consuming and labor intensive. For this reason, it is important to introduce methods that can be rapid, accurate and cost-efficient. In this paper, bridge stiffness values obtained during the in situ experimental campaign were compared with finite element models values. The Fast-FWD has the potential to provide engineering information that can help us to better understand bridge condition, in a rapid and cost-effective procedure.