arXiv subjects

Kezhi Mao

Publications and source records attributed to Kezhi Mao.

At least 19 recordsLinked to original sources

arXiv2026

Understanding the Limits of Agentic ICD Coding

ICD-10-CM codes are alphanumeric codes used in the US to classify diagnoses and injuries for medical billing and epidemiological reporting. Standard ICD-10-CM benchmarks report aggregate metrics that obscure performance on complex coding scenarios. We evaluate neural, workflow, and agentic systems on a rarity-stratified set of MIMIC-IV discharge summaries and identify two orthogonal failure modes. Neural classifiers exhibit a 0.43 micro-F1 gap between rare and common codes. Workflow systems handle rare codes well but score near zero on injury and external cause codes that require multi-step guideline following. A tool-augmented agentic configuration with structured access to official ICD-10-CM reference materials recovers up to 0.34 micro-F1 on this subset. No single system dominates across all conditions.

cs.CL↗

arXiv2026

Fine-Grained Activation Steering: Steering Less, Achieving More

Activation steering has emerged as a cost-effective paradigm for modifying large language model (LLM) behaviors. Existing methods typically intervene at the block level, steering the bundled activations of selected attention heads, feedforward networks, or residual streams. However, we reveal that block-level activations are inherently heterogeneous, entangling beneficial, irrelevant, and harmful features, thereby rendering block-level steering coarse, inefficient, and intrusive. To investigate the root cause, we decompose block activations into fine-grained atomic unit (AU)-level activations, where each AU-level activation corresponds to a single dimension of the block activation, and each AU denotes a slice of the block weight matrix. Steering an AU-level activation is thus equivalent to steering its associated AU. Our theoretical and empirical analysis show that heterogeneity arises because different AUs or dimensions control distinct token distributions in LLM outputs. Hence, block-level steering inevitably moves helpful and harmful token directions together, which reduces efficiency. Restricting intervention to beneficial AUs yields more precise and effective steering. Building on this insight, we propose AUSteer, a simple and efficient method that operates at a finer granularity of the AU level. AUSteer first identifies discriminative AUs globally by computing activation momenta on contrastive samples. It then assigns adaptive steering strengths tailored to diverse inputs and selected AU activations. Comprehensive experiments on multiple LLMs and tasks show that AUSteer consistently surpasses advanced baselines while steering considerably fewer activations, demonstrating that steering less achieves more.

cs.CL↗

arXiv2026

Rethinking Prompt Optimizers: From Prompt Merits to Optimization

Prompt optimization (PO) provides a practical way to improve response quality when users lack the time or expertise to manually craft effective prompts. Existing methods typically rely on LLMs' self-generation ability to optimize prompts. However, due to limited downward compatibility, the instruction-heavy prompts generated by advanced LLMs can overwhelm lightweight inference models and degrade response quality, while also lacking interpretability due to implicit optimization. In this work, we rethink prompt optimization through the lens of explicit and interpretable design. We first identify a set of model-agnostic prompt quality merits and empirically validate their effectiveness in enhancing prompt and response quality. We then introduce MePO, a merit-guided, locally deployable prompt optimizer trained on our merit-guided prompt preference dataset generated by a lightweight LLM. MePO avoids online optimization, reduces privacy concerns, and, by learning clear, interpretable merits, generalizes effectively to both large-scale and lightweight inference models. Experiments demonstrate that MePO achieves better results across diverse tasks and model types, offering a scalable and robust solution for real-world deployment. The code, model and dataset can be found in https://github.com/MidiyaZhu/MePO

cs.CL↗

arXiv2025

Restoring Pruned Large Language Models via Lost Component Compensation

Pruning is a widely used technique to reduce the size and inference cost of large language models (LLMs), but it often causes performance degradation. To mitigate this, existing restoration methods typically employ parameter-efficient fine-tuning (PEFT), such as LoRA, to recover the pruned model's performance. However, most PEFT methods are designed for dense models and overlook the distinct properties of pruned models, often resulting in suboptimal recovery. In this work, we propose a targeted restoration strategy for pruned models that restores performance while preserving their low cost and high efficiency. We observe that pruning-induced information loss is reflected in attention activations, and selectively reintroducing components of this information can significantly recover model performance. Based on this insight, we introduce RestoreLCC (Restoring Pruned LLMs via Lost Component Compensation), a plug-and-play method that contrastively probes critical attention heads via activation editing, extracts lost components from activation differences, and finally injects them back into the corresponding pruned heads for compensation and recovery. RestoreLCC is compatible with structured, semi-structured, and unstructured pruning schemes. Extensive experiments demonstrate that RestoreLCC consistently outperforms state-of-the-art baselines in both general and task-specific performance recovery, without compromising the sparsity or inference efficiency of pruned models.

cs.LG↗

arXiv2025

Domain Lexical Knowledge-based Word Embedding Learning for Text Classification under Small Data

Pre-trained language models such as BERT have been proved to be powerful in many natural language processing tasks. But in some text classification applications such as emotion recognition and sentiment analysis, BERT may not lead to satisfactory performance. This often happens in applications where keywords play critical roles in the prediction of class labels. Our investigation found that the root cause of the problem is that the context-based BERT embedding of the keywords may not be discriminative enough to produce discriminative text representation for classification. Motivated by this finding, we develop a method to enhance word embeddings using domain-specific lexical knowledge. The knowledge-based embedding enhancement model projects the BERT embedding into a new space where within-class similarity and between-class difference are maximized. To implement the knowledge-based word embedding enhancement model, we also develop a knowledge acquisition algorithm for automatically collecting lexical knowledge from online open sources. Experiment results on three classification tasks, including sentiment analysis, emotion recognition and question answering, have shown the effectiveness of our proposed word embedding enhancing model. The codes and datasets are in https://github.com/MidiyaZhu/KVWEFFER.

cs.CL↗

arXiv2025

Beyond the Next Token: Towards Prompt-Robust Zero-Shot Classification via Efficient Multi-Token Prediction

Zero-shot text classification typically relies on prompt engineering, but the inherent prompt brittleness of large language models undermines its reliability. Minor changes in prompt can cause significant discrepancies in model performance. We attribute this prompt brittleness largely to the narrow focus on nexttoken probabilities in existing methods. To address this, we propose Placeholding Parallel Prediction (P3), a novel approach that predicts token probabilities across multiple positions and simulates comprehensive sampling of generation paths in a single run of a language model. Experiments show improved accuracy and up to 98% reduction in the standard deviation across prompts, boosting robustness. Even without a prompt, P3 maintains comparable performance, reducing the need for prompt engineering.

cs.CL↗

arXiv2024

UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation

Large language models (LLMs) have demonstrated impressive capabilities in various tasks using the in-context learning (ICL) paradigm. However, their effectiveness is often compromised by inherent bias, leading to prompt brittleness, i.e., sensitivity to design settings such as example selection, order, and prompt formatting. Previous studies have addressed LLM bias through external adjustment of model outputs, but the internal mechanisms that lead to such bias remain unexplored. Our work delves into these mechanisms, particularly investigating how feedforward neural networks (FFNs) and attention heads result in the bias of LLMs. By Interpreting the contribution of individual FFN vectors and attention heads, we identify the biased LLM components that skew LLMs' prediction toward specific labels. To mitigate these biases, we introduce UniBias, an inference-only method that effectively identifies and eliminates biased FFN vectors and attention heads. Extensive experiments across 12 NLP datasets demonstrate that UniBias significantly enhances ICL performance and alleviates prompt brittleness of LLMs.

cs.CL↗

arXiv2024

LLMs Learn Task Heuristics from Demonstrations: A Heuristic-Driven Prompting Strategy for Document-Level Event Argument Extraction

In this study, we investigate in-context learning (ICL) in document-level event argument extraction (EAE) to alleviate the dependency on large-scale labeled data for this task. We introduce the Heuristic-Driven Link-of-Analogy (HD-LoA) prompting to address the challenge of example selection and to develop a prompting strategy tailored for EAE. Specifically, we hypothesize and validate that LLMs learn task-specific heuristics from demonstrations via ICL. Building upon this hypothesis, we introduce an explicit heuristic-driven demonstration construction approach, which transforms the haphazard example selection process into a methodical method that emphasizes task heuristics. Additionally, inspired by the analogical reasoning of human, we propose the link-of-analogy prompting, which enables LLMs to process new situations by drawing analogies to known situations, enhancing their performance on unseen classes beyond limited ICL examples. Experiments show that our method outperforms existing prompting methods and few-shot supervised learning methods on document-level EAE datasets. Additionally, the HD-LoA prompting shows effectiveness in diverse tasks like sentiment analysis and natural language inference, demonstrating its broad adaptability.

cs.CL↗

arXiv2024

FreeCtrl: Constructing Control Centers with Feedforward Layers for Learning-Free Controllable Text Generation

Controllable text generation (CTG) seeks to craft texts adhering to specific attributes, traditionally employing learning-based techniques such as training, fine-tuning, or prefix-tuning with attribute-specific datasets. These approaches, while effective, demand extensive computational and data resources. In contrast, some proposed learning-free alternatives circumvent learning but often yield inferior results, exemplifying the fundamental machine learning trade-off between computational expense and model efficacy. To overcome these limitations, we propose FreeCtrl, a learning-free approach that dynamically adjusts the weights of selected feedforward neural network (FFN) vectors to steer the outputs of large language models (LLMs). FreeCtrl hinges on the principle that the weights of different FFN vectors influence the likelihood of different tokens appearing in the output. By identifying and adaptively adjusting the weights of attribute-related FFN vectors, FreeCtrl can control the output likelihood of attribute keywords in the generated content. Extensive experiments on single- and multi-attribute control reveal that the learning-free FreeCtrl outperforms other learning-free and learning-based methods, successfully resolving the dilemma between learning costs and model performance.

cs.CL↗

arXiv2024

Unveiling and Manipulating Prompt Influence in Large Language Models

Prompts play a crucial role in guiding the responses of Large Language Models (LLMs). However, the intricate role of individual tokens in prompts, known as input saliency, in shaping the responses remains largely underexplored. Existing saliency methods either misalign with LLM generation objectives or rely heavily on linearity assumptions, leading to potential inaccuracies. To address this, we propose Token Distribution Dynamics (TDD), a \textcolor{black}{simple yet effective} approach to unveil and manipulate the role of prompts in generating LLM outputs. TDD leverages the robust interpreting capabilities of the language model head (LM head) to assess input saliency. It projects input tokens into the embedding space and then estimates their significance based on distribution dynamics over the vocabulary. We introduce three TDD variants: forward, backward, and bidirectional, each offering unique insights into token relevance. Extensive experiments reveal that the TDD surpasses state-of-the-art baselines with a big margin in elucidating the causal relationships between prompts and LLM outputs. Beyond mere interpretation, we apply TDD to two prompt manipulation tasks for controlled text generation: zero-shot toxic language suppression and sentiment steering. Empirical results underscore TDD's proficiency in identifying both toxic and sentimental cues in prompts, subsequently mitigating toxicity or modulating sentiment in the generated content.

cs.CL↗

arXiv2024

GAM-Depth: Self-Supervised Indoor Depth Estimation Leveraging a Gradient-Aware Mask and Semantic Constraints

Self-supervised depth estimation has evolved into an image reconstruction task that minimizes a photometric loss. While recent methods have made strides in indoor depth estimation, they often produce inconsistent depth estimation in textureless areas and unsatisfactory depth discrepancies at object boundaries. To address these issues, in this work, we propose GAM-Depth, developed upon two novel components: gradient-aware mask and semantic constraints. The gradient-aware mask enables adaptive and robust supervision for both key areas and textureless regions by allocating weights based on gradient magnitudes.The incorporation of semantic constraints for indoor self-supervised depth estimation improves depth discrepancies at object boundaries, leveraging a co-optimization network and proxy semantic labels derived from a pretrained segmentation model. Experimental studies on three indoor datasets, including NYUv2, ScanNet, and InteriorNet, show that GAM-Depth outperforms existing methods and achieves state-of-the-art performance, signifying a meaningful step forward in indoor depth estimation. Our code will be available at https://github.com/AnqiCheng1234/GAM-Depth.

cs.CV↗

arXiv2023

Particle swarm optimization with state-based adaptive velocity limit strategy

Velocity limit (VL) has been widely adopted in many variants of particle swarm optimization (PSO) to prevent particles from searching outside the solution space. Several adaptive VL strategies have been introduced with which the performance of PSO can be improved. However, the existing adaptive VL strategies simply adjust their VL based on iterations, leading to unsatisfactory optimization results because of the incompatibility between VL and the current searching state of particles. To deal with this problem, a novel PSO variant with state-based adaptive velocity limit strategy (PSO-SAVL) is proposed. In the proposed PSO-SAVL, VL is adaptively adjusted based on the evolutionary state estimation (ESE) in which a high value of VL is set for global searching state and a low value of VL is set for local searching state. Besides that, limit handling strategies have been modified and adopted to improve the capability of avoiding local optima. The good performance of PSO-SAVL has been experimentally validated on a wide range of benchmark functions with 50 dimensions. The satisfactory scalability of PSO-SAVL in high-dimension and large-scale problems is also verified. Besides, the merits of the strategies in PSO-SAVL are verified in experiments. Sensitivity analysis for the relevant hyper-parameters in state-based adaptive VL strategy is conducted, and insights in how to select these hyper-parameters are also discussed.

cs.NE↗

arXiv2023

Feature-aware conditional GAN for category text generation

Category text generation receives considerable attentions since it is beneficial for various natural language processing tasks. Recently, the generative adversarial network (GAN) has attained promising performance in text generation, attributed to its adversarial training process. However, there are several issues in text GANs, including discreteness, training instability, mode collapse, lack of diversity and controllability etc. To address these issues, this paper proposes a novel GAN framework, the feature-aware conditional GAN (FA-GAN), for controllable category text generation. In FA-GAN, the generator has a sequence-to-sequence structure for improving sentence diversity, which consists of three encoders including a special feature-aware encoder and a category-aware encoder, and one relational-memory-core-based decoder with the Gumbel SoftMax activation function. The discriminator has an additional category classification head. To generate sentences with specified categories, the multi-class classification loss is supplemented in the adversarial training. Comprehensive experiments have been conducted, and the results show that FA-GAN consistently outperforms 10 state-of-the-art text generation approaches on 6 text classification datasets. The case study demonstrates that the synthetic sentences generated by FA-GAN can match the required categories and are aware of the features of conditioned sentences, with good readability, fluency, and text authenticity.

cs.CL↗

arXiv2023

Artificial-Intelligence-Based Triple Phase Shift Modulation for Dual Active Bridge Converter with Minimized Current Stress

The dual active bridge (DAB) converter has been popular in many applications for its outstanding power density and bidirectional power transfer capacity. Up to now, triple phase shift (TPS) can be considered as one of the most advanced modulation techniques for DAB converter. It can widen zero voltage switching range and improve power efficiency significantly. Currently, current stress of the DAB converter has been an important performance indicator when TPS modulation is applied for smaller size and higher efficiency. However, to minimize the current stress when the DAB converter is under TPS modulation, two difficulties exist in analysis process and realization process, respectively. Firstly, three degrees of modulation variables in TPS modulation bring challenges to the analysis of current stress in different operating modes. This analysis and deduction process leads to heavy computational burden and also suffers from low accuracy. Secondly, to realize TPS modulation, if a lookup table is adopted after the optimization of modulation variables, modulation performance will be unsatisfactory because of the discrete nature of lookup table. Therefore, an AI-based TPS modulation (AI-TPSM) strategy is proposed in this paper. Neural network (NN) and fuzzy inference system (FIS) are utilized to deal with the two difficulties mentioned above. With the proposed AI-TPSM, the optimization of TPS modulation for minimized current stress will enjoy high degree of automation which can relieve engineers' working burden and improve accuracy. In the end of this paper, the effectiveness of the proposed AI-TPSM has been experimentally verified with a 1 kW prototype.

eess.SY↗

arXiv2023

Artificial-Intelligence-Based Hybrid Extended Phase Shift Modulation for the Dual Active Bridge Converter with Full ZVS Range and Optimal Efficiency

Dual active bridge (DAB) converter is the key enabler in many popular applications such as wireless charging, electric vehicle and renewable energy. ZVS range and efficiency are two significant performance indicators for DAB converter. To obtain the desired ZVS and efficiency performance, modulation should be carefully designed. Hybrid modulation considers several single modulation strategies to achieve good comprehensive performance. Conventionally, to design a hybrid modulation, harmonic approach or piecewise approach is used, but they suffer from time-consuming model building process and inaccuracy. Therefore, an artificial-intelligence-based hybrid extended phase shift (HEPS) modulation is proposed. Generally, the HEPS modulation is developed in an automated fashion, which alleviates cumbersome model building process while keeping high model accuracy. In HEPS modulation, two EPS strategies are considered to realize optimal efficiency with full ZVS operation over entire operating ranges. Specifically, to build data-driven models of ZVS and efficiency performance, extreme gradient boosting (XGBoost), which is a state-of-the-art ensemble learning algorithm, is adopted. Afterwards, particle swarm optimization with state-based adaptive velocity limit (PSO-SAVL) is utilized to select the best EPS strategy and optimize modulation parameters. With 1 kW hardware experiments, the feasibility of HEPS has been verified, achieving optimal efficiency with maximum of 97.1% and full-range ZVS operation.

eess.SY↗

arXiv2023

Distilling Universal and Joint Knowledge for Cross-Domain Model Compression on Time Series Data

For many real-world time series tasks, the computational complexity of prevalent deep leaning models often hinders the deployment on resource-limited environments (e.g., smartphones). Moreover, due to the inevitable domain shift between model training (source) and deploying (target) stages, compressing those deep models under cross-domain scenarios becomes more challenging. Although some of existing works have already explored cross-domain knowledge distillation for model compression, they are either biased to source data or heavily tangled between source and target data. To this end, we design a novel end-to-end framework called Universal and joint knowledge distillation (UNI-KD) for cross-domain model compression. In particular, we propose to transfer both the universal feature-level knowledge across source and target domains and the joint logit-level knowledge shared by both domains from the teacher to the student model via an adversarial learning scheme. More specifically, a feature-domain discriminator is employed to align teacher's and student's representations for universal knowledge transfer. A data-domain discriminator is utilized to prioritize the domain-shared samples for joint knowledge transfer. Extensive experimental results on four time series datasets demonstrate the superiority of our proposed method over state-of-the-art (SOTA) benchmarks.

cs.LG↗

arXiv2022

Aligning Correlation Information for Domain Adaptation in Action Recognition

Domain adaptation (DA) approaches address domain shift and enable networks to be applied to different scenarios. Although various image DA approaches have been proposed in recent years, there is limited research towards video DA. This is partly due to the complexity in adapting the different modalities of features in videos, which includes the correlation features extracted as long-term dependencies of pixels across spatiotemporal dimensions. The correlation features are highly associated with action classes and proven their effectiveness in accurate video feature extraction through the supervised action recognition task. Yet correlation features of the same action would differ across domains due to domain shift. Therefore we propose a novel Adversarial Correlation Adaptation Network (ACAN) to align action videos by aligning pixel correlations. ACAN aims to minimize the distribution of correlation information, termed as Pixel Correlation Discrepancy (PCD). Additionally, video DA research is also limited by the lack of cross-domain video datasets with larger domain shifts. We, therefore, introduce a novel HMDB-ARID dataset with a larger domain shift caused by a larger statistical difference between domains. This dataset is built in an effort to leverage current datasets for dark video classification. Empirical results demonstrate the state-of-the-art performance of our proposed ACAN for both existing and the new video DA datasets.

cs.CV↗

arXiv2022

ARID: A New Dataset for Recognizing Action in the Dark

The task of action recognition in dark videos is useful in various scenarios, e.g., night surveillance and self-driving at night. Though progress has been made in the action recognition task for videos in normal illumination, few have studied action recognition in the dark. This is partly due to the lack of sufficient datasets for such a task. In this paper, we explored the task of action recognition in dark videos. We bridge the gap of the lack of data for this task by collecting a new dataset: the Action Recognition in the Dark (ARID) dataset. It consists of over 3,780 video clips with 11 action categories. To the best of our knowledge, it is the first dataset focused on human actions in dark videos. To gain further understandings of our ARID dataset, we analyze the ARID dataset in detail and exhibited its necessity over synthetic dark videos. Additionally, we benchmarked the performance of several current action recognition models on our dataset and explored potential methods for increasing their performances. Our results show that current action recognition models and frame enhancement methods may not be effective solutions for the task of action recognition in dark videos.

cs.CV↗