Zeyu Shi

Ziming Wang, Zeyu Shi, Haoyi Zhou et al.

Fine-tuned Large Language Models (LLMs) often demonstrate poor calibration, with their confidence scores misaligned with actual performance. While calibration has been extensively studied in models trained from scratch, the impact of LLMs' prior knowledge on calibration during fine-tuning remains understudied. Our research reveals that LLMs' prior knowledge causes potential poor calibration due to the ubiquitous presence of known data in real-world fine-tuning, which appears harmful for calibration. Specifically, data aligned with LLMs' prior knowledge would induce overconfidence, while new knowledge improves calibration. Our findings expose a tension: LLMs' encyclopedic knowledge, while enabling task versatility, undermines calibration through unavoidable knowledge overlaps. To address this, we propose CogCalib, a cognition-aware framework that applies targeted learning strategies according to the model's prior knowledge. Experiments across 7 tasks using 3 LLM families prove that CogCalib significantly improves calibration while maintaining performance, achieving an average 57\% reduction in ECE compared to standard fine-tuning in Llama3-8B. These improvements generalize well to out-of-domain tasks, enhancing the objectivity and reliability of domain-specific LLMs, and making them more trustworthy for critical human-AI interaction applications.

Zeyu Shi, Ziming Wang, Tianyu Chen et al.

The honesty of Large Language Models (LLMs) is increasingly important for safe deployment in high-stakes domains. However, this crucial trait is severely undermined by supervised fine-tuning (SFT), a common technique for model specialization. Existing recovery methods rely on data-intensive global parameter adjustments, implicitly assuming that SFT deeply corrupts the models' ability to recognize their knowledge boundaries. However, we observe that fine-tuned LLMs still preserve this ability; what is damaged is their capacity to faithfully express that awareness. Building on this, we propose Honesty-Critical Neurons Restoration (HCNR) to surgically repair this suppressed capacity. HCNR identifies and restores key expression-governing neurons to their pre-trained state while harmonizing them with task-oriented neurons via Hessian-guided compensation. Experiments on four QA tasks and five LLM families demonstrate that HCNR effectively recovers 33.25% of the compromised honesty while achieving at least 2.23x speedup with over 10x less data compared to baseline methods, offering a practical solution for trustworthy LLM deployment.

Tao Zhou, Kai Ye, Zeyu Shi et al.

Numerous remarkable advancements have been made in accuracy, speed, and parallelism for solving the Unmanned Aerial Vehicle Route Planing (UAVRP). However, existing UAVRP solvers face challenges when attempting to scale effectively and efficiently for larger instances. In this paper, we present a generalization framework that enables current UAVRP solvers to robustly extend their capabilities to larger instances, accommodating up to 10,000 points, using widely recognized test sets. The UAVRP under a large number of patrol points is a typical large-scale TSP problem.Our proposed framework comprises three distinct steps. Firstly, we employ Delaunay triangulation to extract subgraphs from large instances while preserving global features. Secondly, we utilize an embedded TSP solver to obtain sub-results, followed by graph fusion. Finally, we implement a decoding strategy customizable to the user's requirements, resulting in high-quality solutions, complemented by a warming-up process for the heatmap. To demonstrate the flexibility of our approach, we integrate two representative TSP solvers into our framework and conduct a comprehensive comparative analysis against existing algorithms using large TSP benchmark datasets. The results unequivocally demonstrate that our framework efficiently scales existing TSP solvers to handle large instances and consistently outperforms state-of-the-art (SOTA) methods. Furthermore, since our proposed framework does not necessitate additional training or fine-tuning, we believe that its generality can significantly advance research on end-to-end UAVRP solvers, enabling the application of a broader range of methods to real-world scenarios.