The Answer Is Not the Argument
研究探讨了在AI监管中提供参考答案是否有助于验证推理过程。通过对比有无答案情况下的监控效果，发现提供答案主要提高了结论一致性检查而非独立验证论证过程的有效性。
0 citationsRead paper
Selected work
Representative Papers
Air-Plan: Query-Optimized Topology Selection for Over-the-Air Decentralized Federated Learning
This work addresses the absence of communication topology selection mechanisms tailored to the integration of over-the-air (OTA) computing and decentralized federated learning (DFL). It formulates OTA-DFL for the first time as a distributed query processing problem, establishing an equivalence between this setting and cost-based query optimization. To enable topology-aware cost modeling under privacy constraints, the authors propose a privacy-preserving Count-Min Sketch–based statistical method. By integrating graph-aware cost estimation, topology-aware sparsification, and wireless multi-access superposition computation, the approach achieves near-optimal topologies in 91.4% of evaluated workloads while incurring less than 1.8% additional overhead. The study further provides theoretical guarantees that highly connected topologies exhibit greater robustness to aggressive model compression.
0 citationsRead paper
Patch-Effect Graph Kernels for LLM Interpretability
This work addresses the challenge of systematically comparing high-dimensional, unstructured data generated by large-scale activation patching, which hinders scalable mechanistic interpretability of causal circuits in large language models. The authors propose modeling activation patching results as graph structures and introduce three graph construction methods based on causal mediation, partial correlation, and co-influence. To distinguish task-specific signals from general causal evidence, they incorporate rigorous baseline controls. Evaluating on the Indirect Object Identification (IOI) task in GPT-2 Small using graph kernels and feature extraction, they find that local edge-slot features outperform global graph descriptors in classification performance. Moreover, edges selected via causal inference (CI) and partial correlation (PC) criteria significantly surpass random or low-rank baselines, demonstrating that graph representations effectively compress and reveal structured causal signals.
0 citationsRead paper
From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features
Existing analyses of sparse autoencoder (SAE) features predominantly rely on activation tokens or decoder weights, which struggle to capture higher-order co-occurrence structures among features. This work addresses this limitation by modeling SAE features as token co-occurrence graphs and introduces a frequency-binned Weisfeiler–Lehman graph kernel to measure their structural similarity. The proposed approach uncovers coherent relational patterns that remain invisible under conventional analytical perspectives. Applied to SAE features derived from GPT-2 Small, the method successfully clusters semantically consistent motifs—such as punctuation-dense sequences, linguistic scripts, and code templates—demonstrating robustness across varying hyperparameters and random seeds, thereby validating both its effectiveness and reliability.
0 citationsRead paper
Diffusion-Based Solver for CNF Placement on the Cloud-Continuum
This work addresses the joint deployment of cloud-native network functions (CNFs) across the cloud continuum—spanning cloud, edge, and end devices—in 5G/6G cloud-edge-end collaborative scenarios. Method: We propose the first diffusion-based generative approach for constrained network function placement, integrating denoising diffusion probabilistic models (DDPMs) with graph neural networks (GNNs). A heterogeneous graph unifies service chains and distributed infrastructure, casting CNF placement as an end-to-end denoising process from Gaussian noise to feasible solutions; embedded constraint losses explicitly enforce combinatorial constraints—including resource capacity, bandwidth, and latency. Contribution/Results: Our method achieves 100% feasibility across diverse topologies, with inference speedups of 2–3 orders of magnitude over state-of-the-art mixed-integer nonlinear programming (MINLP) solvers. Results demonstrate the efficacy and generalizability of diffusion models for complex, constraint-aware network embedding tasks.
0 citationsRead paper
Recent publications
Latest Papers
The Answer Is Not the Argument
研究探讨了在AI监管中提供参考答案是否有助于验证推理过程。通过对比有无答案情况下的监控效果，发现提供答案主要提高了结论一致性检查而非独立验证论证过程的有效性。
0 citationsRead paper
Air-Plan: Query-Optimized Topology Selection for Over-the-Air Decentralized Federated Learning
This work addresses the absence of communication topology selection mechanisms tailored to the integration of over-the-air (OTA) computing and decentralized federated learning (DFL). It formulates OTA-DFL for the first time as a distributed query processing problem, establishing an equivalence between this setting and cost-based query optimization. To enable topology-aware cost modeling under privacy constraints, the authors propose a privacy-preserving Count-Min Sketch–based statistical method. By integrating graph-aware cost estimation, topology-aware sparsification, and wireless multi-access superposition computation, the approach achieves near-optimal topologies in 91.4% of evaluated workloads while incurring less than 1.8% additional overhead. The study further provides theoretical guarantees that highly connected topologies exhibit greater robustness to aggressive model compression.
0 citationsRead paper
Patch-Effect Graph Kernels for LLM Interpretability
This work addresses the challenge of systematically comparing high-dimensional, unstructured data generated by large-scale activation patching, which hinders scalable mechanistic interpretability of causal circuits in large language models. The authors propose modeling activation patching results as graph structures and introduce three graph construction methods based on causal mediation, partial correlation, and co-influence. To distinguish task-specific signals from general causal evidence, they incorporate rigorous baseline controls. Evaluating on the Indirect Object Identification (IOI) task in GPT-2 Small using graph kernels and feature extraction, they find that local edge-slot features outperform global graph descriptors in classification performance. Moreover, edges selected via causal inference (CI) and partial correlation (PC) criteria significantly surpass random or low-rank baselines, demonstrating that graph representations effectively compress and reveal structured causal signals.
0 citationsRead paper
From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features
Existing analyses of sparse autoencoder (SAE) features predominantly rely on activation tokens or decoder weights, which struggle to capture higher-order co-occurrence structures among features. This work addresses this limitation by modeling SAE features as token co-occurrence graphs and introduces a frequency-binned Weisfeiler–Lehman graph kernel to measure their structural similarity. The proposed approach uncovers coherent relational patterns that remain invisible under conventional analytical perspectives. Applied to SAE features derived from GPT-2 Small, the method successfully clusters semantically consistent motifs—such as punctuation-dense sequences, linguistic scripts, and code templates—demonstrating robustness across varying hyperparameters and random seeds, thereby validating both its effectiveness and reliability.
0 citationsRead paper
Diffusion-Based Solver for CNF Placement on the Cloud-Continuum
This work addresses the joint deployment of cloud-native network functions (CNFs) across the cloud continuum—spanning cloud, edge, and end devices—in 5G/6G cloud-edge-end collaborative scenarios. Method: We propose the first diffusion-based generative approach for constrained network function placement, integrating denoising diffusion probabilistic models (DDPMs) with graph neural networks (GNNs). A heterogeneous graph unifies service chains and distributed infrastructure, casting CNF placement as an end-to-end denoising process from Gaussian noise to feasible solutions; embedded constraint losses explicitly enforce combinatorial constraints—including resource capacity, bandwidth, and latency. Contribution/Results: Our method achieves 100% feasibility across diverse topologies, with inference speedups of 2–3 orders of magnitude over state-of-the-art mixed-integer nonlinear programming (MINLP) solvers. Results demonstrate the efficacy and generalizability of diffusion models for complex, constraint-aware network embedding tasks.
0 citationsRead paper