arXiv subjects

Chun Tao

Publications and source records attributed to Chun Tao.

4 recordsLinked to original sources

arXiv2026

A Workflow-Aware Serving Layer for Agentic Applications

Agentic AI applications form an emerging serving workload in which a request creates a workflow: a directed acyclic graph of LLM and tool calls that exposes per-node model choices and optional quality operators such as verifiers. This workload falls between two existing layers. Model-serving engines execute individual calls efficiently but cannot see workflow structure, while agent frameworks fix the workflow but cannot see backend load, so neither jointly chooses each node's model, verifier, and backend under serving-time conditions. We present Dyserve, a workflow-aware serving layer that fills this gap. Dyserve compiles each workflow's per-node model and verifier choices in one integer linear program (ILP) over a heterogeneous backend pool, priced by skill-conditioned offline profiles that transfer across workflows. This couples with hardware entering only through per-model throughput sweeps, and is weighted to concentrate strong models and verification on the nodes whose errors propagate the furthest. Because no single latency-quality preference fits every workload mix, Dyserve pre-solves the program at several pressure levels at admission and shifts a workflow's uncommitted suffix among these strategies under load, keeping the solver off the load-shift path; a failed tool call triggers a one-time residual re-solve that preserves committed work.

cs.DC↗

arXiv2026

The Fine-Tuning Trap: Evaluating Negative Transfer and the Role of PEFT in Sub-1B Mathematical Reasoning

Deploying Small Language Models (SLMs) on edge devices requires efficient fine-tuning strategies that adapt models to new tasks without degrading their general capabilities. In this study, we benchmark five sub-1B models (135M-1B) on mathematical reasoning tasks and uncover a critical vulnerability: Full Fine-Tuning (Full FT) actively harms performance in models under 300M parameters, often dropping accuracy below zero-shot baselines. This "negative transfer" makes Parameter-Efficient Fine-Tuning (PEFT) not just an efficiency preference, but a stability requirement. We find that while Low-Rank Adaptation (LoRA) and Weight-Decomposed LoRA (DoRA) perform comparably, their strengths vary by task; DoRA excels in complex reasoning (GSM8K), while LoRA dominates pattern matching (OrcaMath). In particular, Full FT is outperformed by LoRA on aligned models (Qwen2.5-0.5B) and even by simple 5-shot In-Context Learning on the smallest architectures (SmolLM2-135M). Based on these findings, we recommend defaulting to PEFT for all aligned sub-1B models and caution against Full FT for any architecture smaller than 500M parameters to prevent catastrophic forgetting. Reproduction of this work can be found at https://github.com/gulguluu/tiny-slm-finetune-compare.

cs.LG↗

arXiv2024

Semantic-Syntactic Discrepancy in Images (SSDI): Learning Meaning and Order of Features from Natural Images

Despite considerable progress in image classification tasks, classification models seem unaffected by the images that significantly deviate from those that appear natural to human eyes. Specifically, while human perception can easily identify abnormal appearances or compositions in images, classification models overlook any alterations in the arrangement of object parts as long as they are present in any order, even if unnatural. Hence, this work exposes the vulnerability of having semantic and syntactic discrepancy in images (SSDI) in the form of corruptions that remove or shuffle image patches or present images in the form of puzzles. To address this vulnerability, we propose the concept of "image grammar", comprising "image semantics" and "image syntax". Image semantics pertains to the interpretation of parts or patches within an image, whereas image syntax refers to the arrangement of these parts to form a coherent object. We present a semi-supervised two-stage method for learning the image grammar of visual elements and environments solely from natural images. While the first stage learns the semantic meaning of individual object parts, the second stage learns how their relative arrangement constitutes an entire object. The efficacy of the proposed approach is then demonstrated by achieving SSDI detection rates ranging from 70% to 90% on corruptions generated from CelebA and SUN-RGBD datasets. Code is publicly available at: https://github.com/ChunTao1999/SSDI/

cs.CV↗

arXiv2021

On the Noise Stability and Robustness of Adversarially Trained Networks on NVM Crossbars

Applications based on Deep Neural Networks (DNNs) have grown exponentially in the past decade. To match their increasing computational needs, several Non-Volatile Memory (NVM) crossbar based accelerators have been proposed. Recently, researchers have shown that apart from improved energy efficiency and performance, such approximate hardware also possess intrinsic robustness for defense against adversarial attacks. Prior works quantified this intrinsic robustness for vanilla DNNs trained on unperturbed inputs. However, adversarial training of DNNs is the benchmark technique for robustness, and sole reliance on intrinsic robustness of the hardware may not be sufficient. In this work, we explore the design of robust DNNs through the amalgamation of adversarial training and intrinsic robustness of NVM crossbar-based analog hardware. First, we study the noise stability of such networks on unperturbed inputs and observe that internal activations of adversarially trained networks have lower Signal-to-Noise Ratio (SNR), and are sensitive to noise compared to vanilla networks. As a result, they suffer on average 2x performance degradation due to the approximate computations on analog hardware. Noise stability analyses show the instability of adversarially trained DNNs. On the other hand, for adversarial images generated using Square Black Box attacks, ResNet-10/20 adversarially trained on CIFAR-10/100 display a robustness gain of 20-30%. For adversarial images generated using Projected-Gradient-Descent (PGD) White-Box attacks, adversarially trained DNNs present a 5-10% gain in robust accuracy due to underlying NVM crossbar when $\epsilon_{attack}$ is greater than $\epsilon_{train}$. Our results indicate that implementing adversarially trained networks on analog hardware requires careful calibration between hardware non-idealities and $\epsilon_{train}$ for optimum robustness and performance.

cs.LG↗