arXiv subjects

Fengfei Yu

Publications and source records attributed to Fengfei Yu.

4 recordsLinked to original sources

arXiv2026

Calibrating LLMs with Semantic-level Reward

As large language models (LLMs) are deployed in consequential settings such as medical question answering and legal reasoning, the ability to estimate when their outputs are likely to be correct is essential for safe and reliable use, requiring well-calibrated uncertainty. Standard reinforcement learning with verifiable rewards (RLVR) trains models with a binary correctness reward that is indifferent to confidence, providing no penalty for confident but wrong predictions and thereby degrading calibration. Recent work addresses this by training models to produce verbalized confidence scores alongside answers and rewarding agreement with correctness. However, verbalized confidence is calibrated at the token level and thus exhibits inconsistency across textual variations with same semantic meaning. We propose \textbf{Calibration with Semantic Reward (CSR)}, a framework that calibrates language models directly in semantic space without a verbalized confidence interface. CSR combines the correctness reward with a novel semantic calibration reward that encourages exploitation among correct rollouts by promoting semantic agreement, and exploration among incorrect ones by discouraging spurious consistency. Experiments across three model families on HotpotQA (in-distribution) and TriviaQA, MSMARCO, and NQ-Open (out-of-distribution) show that CSR consistently achieves lower ECE and higher AUROC than verbalized-confidence baselines across nearly all settings, reducing ECE by up to $40\%$ and improving AUROC by up to $31\%$ over verbalized-confidence baselines, with calibration behavior generalizing robustly across all four evaluation settings.

cs.CL↗

arXiv2025

GraphCue for SDN Configuration Code Synthesis

We present GraphCue, a topology-grounded retrieval and agent-in-the-loop framework for automated SDN configuration. Each case is abstracted into a JSON graph and embedded using a lightweight three-layer GCN trained with contrastive learning. The nearest validated reference is injected into a structured prompt that constrains code generation, while a verifier closes the loop by executing the candidate configuration and feeding failures back to the agent. On 628 validation cases, GraphCue achieves an 88.2 percent pass rate within 20 iterations and completes 95 percent of verification loops within 9 seconds. Ablation studies without retrieval or structured prompting perform substantially worse, indicating that topology-aware retrieval and constraint-based conditioning are key drivers of performance.

cs.SE↗

arXiv2025

Generative Model Inversion Through the Lens of the Manifold Hypothesis

Model inversion attacks (MIAs) aim to reconstruct class-representative samples from trained models. Recent generative MIAs utilize generative adversarial networks to learn image priors that guide the inversion process, yielding reconstructions with high visual quality and strong fidelity to the private training data. To explore the reason behind their effectiveness, we begin by examining the gradients of inversion loss with respect to synthetic inputs, and find that these gradients are surprisingly noisy. Further analysis reveals that generative inversion implicitly denoises these gradients by projecting them onto the tangent space of the generator manifold, filtering out off-manifold components while preserving informative directions aligned with the manifold. Our empirical measurements show that, in models trained with standard supervision, loss gradients often exhibit large angular deviations from the data manifold, indicating poor alignment with class-relevant directions. This observation motivates our central hypothesis: models become more vulnerable to MIAs when their loss gradients align more closely with the generator manifold. We validate this hypothesis by designing a novel training objective that explicitly promotes such alignment. Building on this insight, we further introduce a training-free approach to enhance gradient-manifold alignment during inversion, leading to consistent improvements over state-of-the-art generative MIAs.

cs.LG↗

arXiv2024

Model Inversion Attacks: A Survey of Approaches and Countermeasures

Deep neural networks have enabled numerous studies and applications on both Euclidean data, such as images and text, and non-Euclidean data, such as graphs. Because these networks may process private data, their deployment raises concerns about privacy leakage. Model inversion attacks (MIAs) exploit access to a trained model to reconstruct training examples or infer privacy-sensitive characteristics represented by the model. The effectiveness of MIAs has been demonstrated in various domains, including images, text, and graphs. These attacks highlight the vulnerability of neural networks and raise awareness about the risk of privacy leakage within the research community. This survey provides a threat-model- and assumption-aware synthesis of attacks and defenses. We compare exposed interfaces, attacker knowledge, reconstruction priors and targets, failure modes, deployment constraints, and privacy-utility trade-offs, while highlighting modeling principles, optimization challenges, and future directions. We also maintain an evolving repository of relevant research at https://github.com/AndrewZhou924/Awesome-model-inversion-attack.

cs.LG↗