arXiv subjects

G. Tang

Publications and source records attributed to G. Tang.

4 recordsLinked to original sources

arXiv2025

Characterizing Model Behavior Under Synthetic Data Training: An Empirical Study Across Scales and Mixing Ratios

Synthetic data generated by large language models has become integral to modern NLP training pipelines, from bootstrapping reasoning capabilities to augmenting instruction-following datasets. While recent work demonstrates successful applications maintaining high external data ratios, systematic understanding of how synthetic data proportion affects model behavior across different scales remains limited. This paper presents a controlled empirical study examining model performance, calibration, and output characteristics when trained on varying synthetic-to-external data ratios. Using the Pythia model suite (410M-12B parameters) across five diverse tasks, we evaluate models after one to three training iterations with synthetic data proportions ranging from 0-50\%. Our key findings include: models maintain stable performance with up to 20\% synthetic data, but degradation accelerates beyond 30\%; larger models (6.9B-12B) show greater robustness to synthetic data than smaller models (410M-1.4B); calibration degradation precedes accuracy loss, providing an early warning signal; and task characteristics matter, with reasoning tasks degrading faster than retrieval tasks under synthetic data training. Importantly, we find that current best practices, such as those employed in STaR and Self-Instruct systems that maintain greater than 80\% external data, operate well within safe regimes identified by our experiments. We provide practical guidance for practitioners on synthetic data budgets based on model scale and task requirements, alongside detailed comparison with concurrent work including Shumailov et al.'s model collapse findings.

cs.CL↗

arXiv2025

Cognitive Decision Routing in Large Language Models: When to Think Fast, When to Think Slow

Large Language Models (LLMs) face a fundamental challenge in deciding when to rely on rapid, intuitive responses versus engaging in slower, more deliberate reasoning. Inspired by Daniel Kahneman's dual-process theory and his insights on human cognitive biases, we propose a novel Cognitive Decision Routing (CDR) framework that dynamically determines the appropriate reasoning strategy based on query characteristics. Our approach addresses the current limitations where models either apply uniform reasoning depth or rely on computationally expensive methods for all queries. We introduce a meta-cognitive layer that analyzes query complexity through multiple dimensions: correlation strength between given information and required conclusions, domain boundary crossings, stakeholder multiplicity, and uncertainty levels. Through extensive experiments on diverse reasoning tasks, we demonstrate that CDR achieves superior performance while reducing computational costs by 34\% compared to uniform deep reasoning approaches. Our framework shows particular strength in professional judgment tasks, achieving 23\% improvement in consistency and 18\% better accuracy on expert-level evaluations. This work bridges cognitive science principles with practical AI system design, offering a principled approach to adaptive reasoning in LLMs.

cs.CL↗

arXiv2020

TurboPy: A Lightweight Python Framework for Computational Physics

Computational physics problems often have a common set of aspects to them that any particular numerical code will have to address. Because these aspects are common to many problems, having a framework already designed and ready to use will not only speed the development of new codes, but also enhance compatibility between codes. Some of the most common aspects of computational physics problems are: a grid, a clock which tracks the flow of the simulation, and a set of models describing the dynamics of various quantities on the grid. Having a framework that could deal with these basic aspects of the simulation in a common way could provide great value to computational scientists by solving various numerical and class design issues that routinely arise. This paper describes the newly developed computational framework that we have built for rapidly prototyping new physics codes. This framework, called turboPy, is a lightweight physics modeling framework based on the design of the particle-in-cell code turboWAVE. It implements a class (called Simulation) which drives the simulation and manages communication between physics modules, a class (called PhysicsModule) which handles the details of the dynamics of the various parts of the problem, and some additional classes such as a Grid class and a Diagnostic class to handle various ancillary issues that commonly arise.

physics.comp-ph↗

arXiv2002

Measurements of the 64Zn(n,a)61Ni cross section at En=5.0-6.75 MeV

The experiment of determination of the 64Zn(n,a)61Ni reaction cross section in the 5.0-6.75 MeV neutron energy range was performed at the 4.5 MV Van de Graaf accelerator at the Institute of Heavy Ion Physics, Peking University, Beijing. Double section ionization chamber with grid was used for direct registration of the reaction products. The cross sections and angular distributions were extracted from the experimental data. The obtained values were compared with the results of other authors and theoretical estimations. The analysis of the experimental data and model calculations were carried out.

nucl-ex↗