LAAF: A Layered Accountability Architecture Framework for LLM Applications
研究针对大型语言模型应用中的责任归属问题，通过文献综述和多维度分析方法，提出了一种分层责任架构框架LAAF。
0 citationsRead paper
Selected work
Representative Papers
Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity
This study investigates how instruction tuning influences confidence expression and lexical diversity in the generated rationales of language models on question-answering tasks. By systematically comparing matched base and instruction-tuned models across multiple QA benchmarks—using confidence calibration metrics, lexical diversity measures, and controlled analyses—it reveals that instruction tuning consistently induces overconfidence and degrades likelihood calibration, even when accuracy gains are negligible. Furthermore, it significantly reduces semantic diversity in rationales across samples, while surface-level lexical diversity exhibits inconsistent changes. These findings highlight previously underappreciated, implicit effects of instruction tuning on model reasoning behavior, offering a new perspective for its refinement and evaluation.
0 citationsRead paper
FedLBW: A Loss-Based Weighting Strategy for Federated Learning on Non-IID Data in Wireless Networks
This work addresses the challenges of slow convergence and degraded accuracy in federated learning over wireless networks, which stem from non-independent and identically distributed (non-IID) client data and frequent client dropouts. To mitigate these issues, the authors propose FedLBW, a loss-based weighted aggregation method that leverages a small proxy dataset on the server to compute validation losses for each client’s model update. Instead of the conventional data-size-based weighting, FedLBW dynamically assigns weights inversely proportional to these validation losses, thereby prioritizing more reliable updates during aggregation. Experimental results on FashionMNIST, CIFAR-10, and CIFAR-100 demonstrate that FedLBW significantly enhances robustness: under extreme non-IID settings, it improves CIFAR-10 accuracy by up to 7.6%, and consistently outperforms baseline methods such as FedAvg even under high client dropout rates.
0 citationsRead paper
The Minimum Subgraph Complementation Problem
This paper studies the Minimum Subgraph Completion problem: given a graph $G$ and a target graph class $mathcal{C}$, find a smallest vertex subset $S$ such that the completion of the induced subgraph $G[S]$ belongs to $mathcal{C}$. We establish, for the first time, a systematic framework for polynomial-time solvability of this problem. Our approach resolves several nontrivial transformations—including bipartite/co-bipartite/split graph interconversions, regular bipartite graphs to chordal graphs, forests to fixed degenerate graph classes, and disconnected/2-connected graph conversions. Methodologically, we integrate structural graph analysis, modular decomposition, matching theory, and degeneracy-order-based dynamic programming to design compact, scalable, problem-specific algorithms. Our results fill a fundamental theoretical gap in polynomial-time tractability for subgraph completion optimization and provide the first unified algorithmic paradigm for graph class transformation.
0 citationsRead paper
From Text to Returns: Using Large Language Models for Mutual Fund Portfolio Optimization and Risk-Adjusted Allocation
This study investigates the application of large language models (LLMs) to mutual fund portfolio optimization and risk-adjusted asset allocation. Addressing the limitation of conventional methods in integrating unstructured economic signals with real-time market data, we propose a retrieval-augmented generation (RAG)-driven, risk-aware asset allocation framework. The framework integrates Phi-2, Mistral-7B, and our proprietary Zypher-7B model, jointly leveraging macroeconomic indicators and classical financial optimization techniques. Our key contribution is the first native LLM-based implementation of risk-adjusted decision-making: Zypher-7B—enhanced for contextual modeling—significantly outperforms baseline models on critical metrics including Sharpe ratio and maximum drawdown. Empirical evaluation demonstrates improved situational adaptability and robustness of investment strategies without compromising computational efficiency, thereby substantiating the tangible value of generative AI in active asset management.
0 citationsRead paper
Recent publications
Latest Papers
LAAF: A Layered Accountability Architecture Framework for LLM Applications
研究针对大型语言模型应用中的责任归属问题，通过文献综述和多维度分析方法，提出了一种分层责任架构框架LAAF。
0 citationsRead paper
Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity
This study investigates how instruction tuning influences confidence expression and lexical diversity in the generated rationales of language models on question-answering tasks. By systematically comparing matched base and instruction-tuned models across multiple QA benchmarks—using confidence calibration metrics, lexical diversity measures, and controlled analyses—it reveals that instruction tuning consistently induces overconfidence and degrades likelihood calibration, even when accuracy gains are negligible. Furthermore, it significantly reduces semantic diversity in rationales across samples, while surface-level lexical diversity exhibits inconsistent changes. These findings highlight previously underappreciated, implicit effects of instruction tuning on model reasoning behavior, offering a new perspective for its refinement and evaluation.
0 citationsRead paper
FedLBW: A Loss-Based Weighting Strategy for Federated Learning on Non-IID Data in Wireless Networks
This work addresses the challenges of slow convergence and degraded accuracy in federated learning over wireless networks, which stem from non-independent and identically distributed (non-IID) client data and frequent client dropouts. To mitigate these issues, the authors propose FedLBW, a loss-based weighted aggregation method that leverages a small proxy dataset on the server to compute validation losses for each client’s model update. Instead of the conventional data-size-based weighting, FedLBW dynamically assigns weights inversely proportional to these validation losses, thereby prioritizing more reliable updates during aggregation. Experimental results on FashionMNIST, CIFAR-10, and CIFAR-100 demonstrate that FedLBW significantly enhances robustness: under extreme non-IID settings, it improves CIFAR-10 accuracy by up to 7.6%, and consistently outperforms baseline methods such as FedAvg even under high client dropout rates.
0 citationsRead paper
The Minimum Subgraph Complementation Problem
This paper studies the Minimum Subgraph Completion problem: given a graph $G$ and a target graph class $mathcal{C}$, find a smallest vertex subset $S$ such that the completion of the induced subgraph $G[S]$ belongs to $mathcal{C}$. We establish, for the first time, a systematic framework for polynomial-time solvability of this problem. Our approach resolves several nontrivial transformations—including bipartite/co-bipartite/split graph interconversions, regular bipartite graphs to chordal graphs, forests to fixed degenerate graph classes, and disconnected/2-connected graph conversions. Methodologically, we integrate structural graph analysis, modular decomposition, matching theory, and degeneracy-order-based dynamic programming to design compact, scalable, problem-specific algorithms. Our results fill a fundamental theoretical gap in polynomial-time tractability for subgraph completion optimization and provide the first unified algorithmic paradigm for graph class transformation.
0 citationsRead paper
From Text to Returns: Using Large Language Models for Mutual Fund Portfolio Optimization and Risk-Adjusted Allocation
This study investigates the application of large language models (LLMs) to mutual fund portfolio optimization and risk-adjusted asset allocation. Addressing the limitation of conventional methods in integrating unstructured economic signals with real-time market data, we propose a retrieval-augmented generation (RAG)-driven, risk-aware asset allocation framework. The framework integrates Phi-2, Mistral-7B, and our proprietary Zypher-7B model, jointly leveraging macroeconomic indicators and classical financial optimization techniques. Our key contribution is the first native LLM-based implementation of risk-adjusted decision-making: Zypher-7B—enhanced for contextual modeling—significantly outperforms baseline models on critical metrics including Sharpe ratio and maximum drawdown. Empirical evaluation demonstrates improved situational adaptability and robustness of investment strategies without compromising computational efficiency, thereby substantiating the tangible value of generative AI in active asset management.
0 citationsRead paper