Selected work

Representative Papers

The Third VoicePrivacy Challenge: Preserving Emotional Expressiveness and Linguistic Content in Voice Anonymization

Jan 17, 2026

This work addresses the challenge of voice anonymization by preserving linguistic content and emotional expression while concealing speaker identity. It introduces the first systematic evaluation framework that explicitly incorporates emotional fidelity as a core assessment dimension, establishing a multi-objective optimization paradigm that jointly optimizes privacy protection, semantic preservation, and emotional consistency. By integrating techniques such as speaker embedding perturbation, voice conversion, and generative modeling, and by introducing objective metrics based on adversarial attack models, the framework enables comprehensive evaluation of various baseline and submitted anonymization systems. Experimental results demonstrate that the proposed approach effectively balances privacy guarantees with speech utility, offering a new benchmark and guiding direction for future research in voice privacy.

1 citationsRead paper

SmartGen: Seamless Disaggregated LLM Inference with Selective KV Cache Transfer

Jul 30, 2026

This work addresses the substantial inter-node communication overhead in decoupled large language model (LLM) inference within cloud environments, where separating prefill and decode stages incurs high bandwidth costs due to full KV cache transmission. To mitigate this, the authors propose a selective KV cache transfer mechanism that transmits only critical cache entries. The approach leverages profiling-driven importance prediction and integrates three data pathways: proactive transfer, parallel on-demand fetching, and speculative completion. As the first selective KV transmission scheme tailored for decoupled LLM inference, it achieves up to a 4.3× reduction in time-to-second-token latency while preserving generation accuracy and decode throughput, significantly alleviating cross-node communication demands.

0 citationsRead paper

StageGuard: Physiologically Constrained Sleep Staging

Jul 25, 2026

This study addresses a critical limitation of current automatic sleep staging models: despite high classification accuracy, they often produce hypnograms that violate physiological constraints—such as implausible state transitions or excessive fragmentation—leading to biased downstream metrics. To resolve this, the authors propose StageGuard, a plug-and-play, backbone-agnostic structured inference framework that jointly enforces state transition rules and minimum duration constraints during both training and decoding via a differentiable soft transition penalty and a semi-Markovian constrained decoder. By seamlessly integrating physiological priors as soft constraints into deep staging pipelines, StageGuard significantly enhances output plausibility while preserving rare pathological events. Experiments across six backbone architectures and four datasets demonstrate that StageGuard reduces transition violations to physiologically plausible levels, decreases fragmentation by 56–62%, maintains or slightly improves classification accuracy, and cuts errors in derived clinical metrics by 59–79%, thereby more accurately capturing subgroup effects.

0 citationsRead paper

Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI

Jul 24, 2026

Current agent benchmarks often yield misleading evaluation scores due to invalid protocols, primarily stemming from reward hacking or assessment vulnerabilities. This work presents the first systematic formalization of “protocol validity” and introduces Mislead Gap—a quantitative metric—and HackDetect, a posterior auditing framework. By integrating trajectory auditing, exposure point identification, and intent-exploitation score comparison, the framework uniformly detects and quantifies the impact of reward hacking. Empirical analysis across 15 benchmarks and 2,385 agent trajectories reveals that 66.7%–67.0% of evaluations exhibit exposure to or active engagement in reward hacking, inflating scores by 0.45–1.00. These findings demonstrate that prevailing benchmarks generally fail to validate agents’ true capabilities.

0 citationsRead paper

Investigating Codec-Internal Latent Audio Watermarking for Neural Codec Robustness

Jul 23, 2026

This work addresses the vulnerability of traditional audio watermarking schemes to neural audio codecs by proposing a method that embeds 32-bit messages directly into the continuous latent space of a codec-like speech autoencoder, prior to neural decoding. The approach significantly enhances robustness against neural codec transformations through a SEANet-style encoder-decoder architecture, a Conformer-based message embedder, RVQ-guided latent space decomposition, and a latent-domain detector. The system is trained under joint perturbations from both conventional signal processing and neural codecs. Evaluated on 48 kHz speech and tested after perceptual training with EnCodec, the method improves bit accuracy from 78.8% to 95.6% and 97.1% under EnCodec-24k compression, with only minor PESQ degradation to 3.514 and 3.427, respectively.

0 citationsRead paper

Recent publications

Latest Papers

SmartGen: Seamless Disaggregated LLM Inference with Selective KV Cache Transfer

Jul 30, 2026

This work addresses the substantial inter-node communication overhead in decoupled large language model (LLM) inference within cloud environments, where separating prefill and decode stages incurs high bandwidth costs due to full KV cache transmission. To mitigate this, the authors propose a selective KV cache transfer mechanism that transmits only critical cache entries. The approach leverages profiling-driven importance prediction and integrates three data pathways: proactive transfer, parallel on-demand fetching, and speculative completion. As the first selective KV transmission scheme tailored for decoupled LLM inference, it achieves up to a 4.3× reduction in time-to-second-token latency while preserving generation accuracy and decode throughput, significantly alleviating cross-node communication demands.

0 citationsRead paper

StageGuard: Physiologically Constrained Sleep Staging

Jul 25, 2026

This study addresses a critical limitation of current automatic sleep staging models: despite high classification accuracy, they often produce hypnograms that violate physiological constraints—such as implausible state transitions or excessive fragmentation—leading to biased downstream metrics. To resolve this, the authors propose StageGuard, a plug-and-play, backbone-agnostic structured inference framework that jointly enforces state transition rules and minimum duration constraints during both training and decoding via a differentiable soft transition penalty and a semi-Markovian constrained decoder. By seamlessly integrating physiological priors as soft constraints into deep staging pipelines, StageGuard significantly enhances output plausibility while preserving rare pathological events. Experiments across six backbone architectures and four datasets demonstrate that StageGuard reduces transition violations to physiologically plausible levels, decreases fragmentation by 56–62%, maintains or slightly improves classification accuracy, and cuts errors in derived clinical metrics by 59–79%, thereby more accurately capturing subgroup effects.

0 citationsRead paper

Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI

Jul 24, 2026

Current agent benchmarks often yield misleading evaluation scores due to invalid protocols, primarily stemming from reward hacking or assessment vulnerabilities. This work presents the first systematic formalization of “protocol validity” and introduces Mislead Gap—a quantitative metric—and HackDetect, a posterior auditing framework. By integrating trajectory auditing, exposure point identification, and intent-exploitation score comparison, the framework uniformly detects and quantifies the impact of reward hacking. Empirical analysis across 15 benchmarks and 2,385 agent trajectories reveals that 66.7%–67.0% of evaluations exhibit exposure to or active engagement in reward hacking, inflating scores by 0.45–1.00. These findings demonstrate that prevailing benchmarks generally fail to validate agents’ true capabilities.

0 citationsRead paper

Investigating Codec-Internal Latent Audio Watermarking for Neural Codec Robustness

Jul 23, 2026

This work addresses the vulnerability of traditional audio watermarking schemes to neural audio codecs by proposing a method that embeds 32-bit messages directly into the continuous latent space of a codec-like speech autoencoder, prior to neural decoding. The approach significantly enhances robustness against neural codec transformations through a SEANet-style encoder-decoder architecture, a Conformer-based message embedder, RVQ-guided latent space decomposition, and a latent-domain detector. The system is trained under joint perturbations from both conventional signal processing and neural codecs. Evaluated on 48 kHz speech and tested after perceptual training with EnCodec, the method improves bit accuracy from 78.8% to 95.6% and 97.1% under EnCodec-24k compression, with only minor PESQ degradation to 3.514 and 3.427, respectively.

0 citationsRead paper

Graph Neural Network-based Algorithm Selection for the Traveling Salesman Problem: A Systematic Study of Cost and Rank Losses under Distinct Budget Regimes

Jul 20, 2026

This work addresses the challenge that Traveling Salesman Problem (TSP) solver performance is highly instance-dependent by proposing GNNAS-TSP, a framework that—without manual feature engineering—leverages graph neural networks to learn instance representations directly from raw graph structures. It formulates algorithm selection as a joint cost prediction and ranking task. The study systematically evaluates various combinations of loss functions, including MSE, MAE, Huber for regression and RankNet, ListNet, LambdaRank for learning-to-rank, under different computational budgets. Experimental results demonstrate that GNNAS-TSP consistently outperforms the single best solver (SBS) at both 10-second and 60-second time limits, achieving statistically significant and robust improvements in solution quality, particularly within the stringent 10-second budget.

0 citationsRead paper