Selected work

Representative Papers

Machine Unlearning using Forgetting Neural Networks

Oct 29, 2024arXiv.org

To address the privacy requirement of compliantly erasing specific user data from AI models without costly retraining, this paper proposes a novel machine unlearning method based on Forgetting Neural Networks (FNNs). We introduce the first learnable and verifiable forgetting framework grounded in FNNs, incorporating four distinct forgetting layers—gated, masked, perturbed, and reconstructed—to emulate cognitive forgetting mechanisms. Empirical evaluation on MNIST and Fashion-MNIST demonstrates that our approach significantly reduces membership inference attack success rates (average reduction >40%) while maintaining model utility, with post-forgetting accuracy degradation under 2%. This work provides an efficient, interpretable, and retraining-free solution for model-level data unlearning, advancing trustworthy AI systems.

3 citationsRead paper

KGCaRe: Explainable Complex Conditional Question Answering using Automatic Knowledge Graph Construction and Context Retrieval with LLMs

Aug 10, 2026

This work addresses the challenge that large language models (LLMs) and conventional retrieval-augmented generation approaches struggle to accurately answer complex conditional questions in domain-specific settings. The authors propose a hybrid architecture integrating neural retrieval with symbolic reasoning, which automatically constructs knowledge graphs from documents via a multi-prompt strategy and stores them in a graph database while simultaneously building a vector index for semantic retrieval. They introduce an LLM-guided iterative graph traversal mechanism coupled with a clue-entity re-traversal strategy to jointly retrieve textual passages and graph triple paths, thereby generating answers that are both interpretable and highly accurate. Experiments demonstrate that the method significantly outperforms baselines—including vanilla LLMs, Think-on-Graph, and HybridContextQA—across two complex conditional question answering benchmarks, with consistent gains observed using diverse LLMs such as Mistral, Mixtral, GPT-3.5, and GPT-4o. The complete software pipeline has been open-sourced.

0 citationsRead paper

SciPhy Reinforcement Learning for Portfolio Optimization

Jul 16, 2026

This study addresses the multi-period dynamic asset allocation problem faced by large institutional investors under trading costs and market impact. The authors propose a signal-driven robust optimization framework that formulates a continuous-time reinforcement learning model, innovatively projecting the Hamilton–Jacobi–Bellman (HJB) equation onto observed price paths and solving it offline in a single pass using physics-informed neural networks (PINNs), thereby circumventing traditional iterative procedures. The approach further incorporates a microstructure-based quadratic price impact model and a discrete target-holding mechanism tailored to short-term trading decisions. Empirical evaluation on a portfolio of 14 ETFs demonstrates that the method significantly improves out-of-sample Sharpe ratios while effectively controlling volatility and turnover, outperforming both static and myopic baseline strategies.

0 citationsRead paper

BoolXLLM: LLM-Assisted Explainability for Boolean Models

May 12, 2026

This work addresses the challenge non-technical users face in interpreting semantic feature selection and logical rules within Boolean models. To bridge the gap between formal logic and human-readable explanations, the study presents the first end-to-end integration of large language models (LLMs) into the entire Boolean rule learning pipeline. The proposed approach leverages LLMs to support three critical stages: feature selection, recommendation of discretization thresholds for numerical features, and rule compression with natural language explanations. By synergistically combining the symbolic reasoning of BoolXAI classifiers with the language generation capabilities of LLMs, the method maintains competitive predictive performance while substantially enhancing interpretability for non-expert users, effectively reconciling rigorous logical formalism with intuitive, accessible explanations.

0 citationsRead paper

LegalDrill: Diagnosis-Driven Synthesis for Legal Reasoning in Small Language Models

Apr 26, 2026

This work addresses the limited performance of small language models on complex legal reasoning tasks, primarily due to the scarcity of high-quality, fine-grained reasoning trajectory data. To overcome this challenge, the authors propose LegalDrill, a novel framework featuring diagnosis-driven data synthesis and self-reflective sample selection. LegalDrill employs fine-grained prompts to extract and iteratively refine reasoning trajectories from a strong teacher model, then leverages self-reflection to automatically identify high-value training samples without human annotation. By integrating supervised fine-tuning with direct preference optimization, the method significantly enhances the legal reasoning capabilities of small models across multiple benchmarks, outperforming baselines that rely on manual labeling or standard sampling strategies.

0 citationsRead paper

Recent publications

Latest Papers

KGCaRe: Explainable Complex Conditional Question Answering using Automatic Knowledge Graph Construction and Context Retrieval with LLMs

Aug 10, 2026

This work addresses the challenge that large language models (LLMs) and conventional retrieval-augmented generation approaches struggle to accurately answer complex conditional questions in domain-specific settings. The authors propose a hybrid architecture integrating neural retrieval with symbolic reasoning, which automatically constructs knowledge graphs from documents via a multi-prompt strategy and stores them in a graph database while simultaneously building a vector index for semantic retrieval. They introduce an LLM-guided iterative graph traversal mechanism coupled with a clue-entity re-traversal strategy to jointly retrieve textual passages and graph triple paths, thereby generating answers that are both interpretable and highly accurate. Experiments demonstrate that the method significantly outperforms baselines—including vanilla LLMs, Think-on-Graph, and HybridContextQA—across two complex conditional question answering benchmarks, with consistent gains observed using diverse LLMs such as Mistral, Mixtral, GPT-3.5, and GPT-4o. The complete software pipeline has been open-sourced.

0 citationsRead paper

SciPhy Reinforcement Learning for Portfolio Optimization

Jul 16, 2026

This study addresses the multi-period dynamic asset allocation problem faced by large institutional investors under trading costs and market impact. The authors propose a signal-driven robust optimization framework that formulates a continuous-time reinforcement learning model, innovatively projecting the Hamilton–Jacobi–Bellman (HJB) equation onto observed price paths and solving it offline in a single pass using physics-informed neural networks (PINNs), thereby circumventing traditional iterative procedures. The approach further incorporates a microstructure-based quadratic price impact model and a discrete target-holding mechanism tailored to short-term trading decisions. Empirical evaluation on a portfolio of 14 ETFs demonstrates that the method significantly improves out-of-sample Sharpe ratios while effectively controlling volatility and turnover, outperforming both static and myopic baseline strategies.

0 citationsRead paper

BoolXLLM: LLM-Assisted Explainability for Boolean Models

May 12, 2026

This work addresses the challenge non-technical users face in interpreting semantic feature selection and logical rules within Boolean models. To bridge the gap between formal logic and human-readable explanations, the study presents the first end-to-end integration of large language models (LLMs) into the entire Boolean rule learning pipeline. The proposed approach leverages LLMs to support three critical stages: feature selection, recommendation of discretization thresholds for numerical features, and rule compression with natural language explanations. By synergistically combining the symbolic reasoning of BoolXAI classifiers with the language generation capabilities of LLMs, the method maintains competitive predictive performance while substantially enhancing interpretability for non-expert users, effectively reconciling rigorous logical formalism with intuitive, accessible explanations.

0 citationsRead paper

LegalDrill: Diagnosis-Driven Synthesis for Legal Reasoning in Small Language Models

Apr 26, 2026

This work addresses the limited performance of small language models on complex legal reasoning tasks, primarily due to the scarcity of high-quality, fine-grained reasoning trajectory data. To overcome this challenge, the authors propose LegalDrill, a novel framework featuring diagnosis-driven data synthesis and self-reflective sample selection. LegalDrill employs fine-grained prompts to extract and iteratively refine reasoning trajectories from a strong teacher model, then leverages self-reflection to automatically identify high-value training samples without human annotation. By integrating supervised fine-tuning with direct preference optimization, the method significantly enhances the legal reasoning capabilities of small models across multiple benchmarks, outperforming baselines that rely on manual labeling or standard sampling strategies.

0 citationsRead paper

Transfer Learning from Foundational Optimization Embeddings to Unsupervised SAT Representations

Apr 16, 2026

This work investigates whether foundation embeddings trained for optimization problems can generalize to decision problems such as Boolean satisfiability (SAT), enabling cross-task transfer. By mapping conjunctive normal form (CNF) formulas onto the same bipartite constraint-variable graph structure used in mixed-integer programming, the authors directly transfer a pretrained embedding model without architectural modifications or supervised fine-tuning, achieving unsupervised representation learning. This study provides the first evidence that foundation optimization embeddings can be effectively transferred to constraint satisfaction problems, advancing a unified representational framework for optimization and decision tasks. Experiments demonstrate that the approach successfully captures structural regularities in SAT instances, supporting unsupervised clustering and distribution identification, thereby validating the feasibility and efficacy of cross-domain transfer.

0 citationsRead paper