Selected work

Representative Papers

COMPASS: A Framework for Evaluating Organization-Specific Policy Alignment in LLMs

Jan 05, 2026arXiv.org

This work addresses a critical gap in the safety evaluation of large language models (LLMs), which has predominantly focused on general risks while neglecting systematic assessment of compliance with organization-specific policies such as allowlists and blocklists. To bridge this gap, we propose COMPASS, the first evaluation framework tailored for organizational policy alignment. COMPASS encompasses 5,920 test queries spanning eight industries, integrating policy-driven query generation, adversarial edge cases, human validation, and multi-model benchmarking. Experiments across seven mainstream LLMs reveal that while models correctly fulfill over 95% of permissible requests, they fail to reject 60%–87% of prohibited adversarial queries, exposing significant vulnerabilities in high-stakes policy enforcement scenarios. This study thus fills a crucial void in enterprise-grade AI safety evaluation.

1 citationsRead paper

Mind the Approximation: Fisher-Weighted SVD Compression for ViTs

Sep 07, 2026

研究解决了Vision Transformers压缩时精度损失问题，提出FACTS方法，通过改进Fisher加权SVD和引入快速约束秩搜索来优化模型压缩效果。

0 citationsRead paper

Beyond Overt Reactions: Analyzing Subtle User Emotional Response to Unexpected In-Vehicle System Behavior

Aug 15, 2026

This study addresses the challenges of recognizing subtle user emotions and data scarcity associated with unexpected autonomous driving behaviors. Leveraging a driving simulator, we collected video, audio, and heart rate signals to construct the first multimodal dataset capturing subtle in-vehicle emotional responses during unexpected events. By integrating affective computing with behavioral analysis, this work elucidates the characteristics of users' subtle emotional reactions to sudden incidents, thereby bridging a critical data gap in the field. The findings validate the necessity for vehicles to perceive and adapt to user states, providing essential empirical evidence and theoretical foundations for enhancing the safety and adaptive capabilities of human-machine interaction in autonomous driving systems.

0 citationsRead paper

Safeguards for Speech2Speech LLM-Assistants: A Case Study in Automotive Applications

Jul 23, 2026

This study addresses the challenges of integrating programmable, domain-specific safety guardrails into in-vehicle end-to-end speech-to-speech (S2S) dialogue systems, where real-time performance and reliability are critical. It presents the first systematic evaluation of two prevailing approaches—transcript-based post-hoc filtering and tool-call-mediated safety mechanisms—in realistic automotive settings. By combining large S2S language models with precise latency measurements and behavioral analysis, the work reveals that both methods introduce delays ranging from 0 to 1.4 seconds and exhibit non-deterministic behaviors. These characteristics render them ill-suited for automotive applications demanding stringent low-latency and high-determinism requirements, thereby exposing fundamental limitations of current safety guardrail designs in real-time spoken interaction scenarios.

0 citationsRead paper

Long-Context Fine-Tuning with Limited VRAM

Jul 16, 2026

This work addresses the high GPU memory cost of dense attention in fine-tuning large language models with long contexts. The authors propose an efficient training method that integrates hierarchical global attention (HGA), segmented backpropagation, and hierarchical KV caching: only the current segment resides in GPU memory, while historical KV caches are offloaded to host memory or NVMe storage and selectively loaded via HGA to provide each query block with a bounded window of relevant past tokens. This approach enables, for the first time, 16K-context QLoRA fine-tuning on a 16GB GPU (Quadro RTX 5000) with a peak memory usage of 15.28 GB, achieving comparable training quality to dense attention (2.7405 vs. 2.7383 nats). Moreover, it supports inference scaling up to 131K context length, substantially alleviating the memory bottleneck in long-context model adaptation.

0 citationsRead paper

Recent publications

Latest Papers

Mind the Approximation: Fisher-Weighted SVD Compression for ViTs

Sep 07, 2026

研究解决了Vision Transformers压缩时精度损失问题，提出FACTS方法，通过改进Fisher加权SVD和引入快速约束秩搜索来优化模型压缩效果。

0 citationsRead paper

Beyond Overt Reactions: Analyzing Subtle User Emotional Response to Unexpected In-Vehicle System Behavior

Aug 15, 2026

This study addresses the challenges of recognizing subtle user emotions and data scarcity associated with unexpected autonomous driving behaviors. Leveraging a driving simulator, we collected video, audio, and heart rate signals to construct the first multimodal dataset capturing subtle in-vehicle emotional responses during unexpected events. By integrating affective computing with behavioral analysis, this work elucidates the characteristics of users' subtle emotional reactions to sudden incidents, thereby bridging a critical data gap in the field. The findings validate the necessity for vehicles to perceive and adapt to user states, providing essential empirical evidence and theoretical foundations for enhancing the safety and adaptive capabilities of human-machine interaction in autonomous driving systems.

0 citationsRead paper

Safeguards for Speech2Speech LLM-Assistants: A Case Study in Automotive Applications

Jul 23, 2026

This study addresses the challenges of integrating programmable, domain-specific safety guardrails into in-vehicle end-to-end speech-to-speech (S2S) dialogue systems, where real-time performance and reliability are critical. It presents the first systematic evaluation of two prevailing approaches—transcript-based post-hoc filtering and tool-call-mediated safety mechanisms—in realistic automotive settings. By combining large S2S language models with precise latency measurements and behavioral analysis, the work reveals that both methods introduce delays ranging from 0 to 1.4 seconds and exhibit non-deterministic behaviors. These characteristics render them ill-suited for automotive applications demanding stringent low-latency and high-determinism requirements, thereby exposing fundamental limitations of current safety guardrail designs in real-time spoken interaction scenarios.

0 citationsRead paper

Long-Context Fine-Tuning with Limited VRAM

Jul 16, 2026

This work addresses the high GPU memory cost of dense attention in fine-tuning large language models with long contexts. The authors propose an efficient training method that integrates hierarchical global attention (HGA), segmented backpropagation, and hierarchical KV caching: only the current segment resides in GPU memory, while historical KV caches are offloaded to host memory or NVMe storage and selectively loaded via HGA to provide each query block with a bounded window of relevant past tokens. This approach enables, for the first time, 16K-context QLoRA fine-tuning on a 16GB GPU (Quadro RTX 5000) with a peak memory usage of 15.28 GB, achieving comparable training quality to dense attention (2.7405 vs. 2.7383 nats). Moreover, it supports inference scaling up to 131K context length, substantially alleviating the memory bottleneck in long-context model adaptation.

0 citationsRead paper

HiFi-LLP: High-Fidelity, Low-Cost Latency Predictors with Confidence for Robust HW-NAS

Jul 13, 2026

This work addresses the inefficiency of hardware-aware neural architecture search (HW-NAS), which traditionally relies on costly hardware-in-the-loop (HIL) latency measurements, and the limitations of existing latency predictors—namely high sample requirements and insufficient accuracy. To overcome these challenges, we propose HiFi-LLP, a high-fidelity, low-cost latency predictor based on graph attention networks, featuring a novel confidence estimation mechanism. Within a hybrid NAS framework, predictions with low confidence are selectively validated via HIL, substantially reducing HIL dependency while preserving robustness. Experiments demonstrate that HiFi-LLP achieves a Spearman correlation coefficient of 0.996 across six hardware platforms in LatBench using only a small number of samples, outperforms state-of-the-art methods by up to 9 percentage points in accuracy within 10% error tolerance, accelerates NAS by up to 8.6×, and maintains superior Pareto-front performance.

0 citationsRead paper