Yiming Tang

Qiran Zou, Hou Hei Lam, Wenhao Zhao et al.

AI research agents accelerate ML research by automating hypothesis generation, experimentation, and empirical refinement. Existing agent strategies range from greedy hill-climbing to tree search and evolutionary optimization, yet which strategy choices drive performance remains unclear. Answering this question requires a benchmark that separates agent strategy (e.g., search topology) from execution infrastructure (e.g., code editor), so that performance differences are attributable to strategy rather than infrastructure, and that provides process-level metrics beyond final scores to analyze exploration behaviors. Existing benchmarks offer limited support. We propose FML-Bench, a benchmark of 18 fundamental ML research tasks across 10 domains that separates agent strategy from execution infrastructure and defines 12 process-level behavioral metrics. Evaluating six representative agents, we find that: (1) strategy complexity alone does not guarantee strong performance: a simple greedy hill-climber nearly matches the best-performing tree-search agent, both well above the remaining agents; (2) our analysis suggests this pattern relates to improvement opportunity structure: greedy search tends to be more effective when opportunities are dense, while tree-search and evolutionary strategies tend to be more effective when opportunities are sparse; an adaptive agent built on this insight switches to broader exploration upon detecting improvement stagnation and outperforms the other six agents, lending initial support to this observation; and (3) process-level analysis reveals that early convergence and directionally focused exploration are significantly associated with final performance, while solution diversity and compute cost are not. Our benchmark is available at: https://github.com/qrzou/FML-bench.

Gaole Dai, Chun-Kai Fan, Yiming Tang et al. · pku

Advances in Parameter-Efficient Fine-Tuning (PEFT) bridged the performance gap with Full Fine-Tuning (FFT) through sophisticated analysis of pre-trained parameter spaces. Starting from drawing insights from Neural Engrams (NE) in Biological Neural Networks (BNNs), we establish a connection between the low-rank property observed during PEFT's parameter space shifting and neurobiological mechanisms. This observation leads to our proposed method, Synapse and Neuron (SAN), which decomposes and propagates scaling components from anterior feature adjusting vectors towards posterior weight matrices. Our approach is theoretically grounded in Long-Term Potentiation/Depression (LTP/D) phenomena, which govern synapse development through neurotransmitter release modulation. Extensive experiments demonstrate its effectiveness: on \textbf{vision tasks} across VTAB, FGVC, and GIC (25 datasets) using ViT, SwinT and ConvNeXt, SAN outperforms FFT up to 8.7% and LoRA by 3.2%; on language tasks using Commonsense Reasoning (8 datasets) with LLaMA models (all generations), surpassing ChatGPT up to 8.5% and LoRA by 4.7%; on visual-language tasks using Mixed Visual Instruction (7 datasets) with LLaVA models, it exceeds FFT up to 2.4% and LoRA by 1.9%. Our code and W&B log will be released.

Yiming Tang, Yi Fan, Chenxiao Yu et al.

The integration of large language models (LLMs) into information retrieval systems introduces new attack surfaces, particularly for adversarial ranking manipulations. We present $\textbf{StealthRank}$, a novel adversarial attack method that manipulates LLM-driven ranking systems while maintaining textual fluency and stealth. Unlike existing methods that often introduce detectable anomalies, StealthRank employs an energy-based optimization framework combined with Langevin dynamics to generate StealthRank Prompts (SRPs)-adversarial text sequences embedded within item or document descriptions that subtly yet effectively influence LLM ranking mechanisms. We evaluate StealthRank across multiple LLMs, demonstrating its ability to covertly boost the ranking of target items while avoiding explicit manipulation traces. Our results show that StealthRank consistently outperforms state-of-the-art adversarial ranking baselines in both effectiveness and stealth, highlighting critical vulnerabilities in LLM-driven ranking systems. Our code is publicly available at $\href{https://github.com/Tangyiming205069/controllable-seo}{here}$.

Zhiyuan Chen, Love Jayesh Ahir, Ahmad Suleiman et al.

Privacy policies are intended to inform users about how software systems collect and handle data, yet they often remain vague or incomplete. This paper presents an empirical study of patterns in log-related statements within privacy policies and their alignment with privacy disclosures observed in Android application logs. We analyzed 1,000 Android apps across multiple categories, generating 86,836,964 log entries. Our findings reveal that while most applications (88.0%) provide privacy policies, only 28.5% explicitly mention logging practices. Among those that reference logging, most clearly describe what information is logged; however, 27.7% of log-related statements remain overly simplistic or vague, offering limited insight into actual data collection. We further observed widespread privacy leakages in application logs, with 67.6% of apps leaking sensitive information not mentioned in their policies. Alarmingly, only 4% of applications demonstrated consistent alignment between declared policy contents and actual logged data. These findings highlight that current privacy policies provide incomplete or ambiguous descriptions of logging practices, which frequently do not align with actual logging behaviors.

Yifan Luo, Yiming Tang, Chengfeng Shen et al.

Prompt Engineering (PE) has emerged as a critical technique for guiding Large Language Models (LLMs) in solving intricate tasks. Its importance is highlighted by its potential to significantly enhance the efficiency and effectiveness of human-machine interaction. As tasks grow increasingly complex, recent advanced PE methods have extended beyond the limitations of single-round interactions to embrace multi-round interactions, which allows for a deeper and more nuanced engagement with LLMs. In this paper, we propose an optimal control framework tailored for multi-round interactions with LLMs. This framework provides a unified mathematical structure that not only systematizes the existing PE methods but also sets the stage for rigorous analytical improvements. Furthermore, we extend this framework to include PE via ensemble methods and multi-agent collaboration, thereby enlarging the scope of applicability. By adopting an optimal control perspective, we offer fresh insights into existing PE methods and highlight theoretical challenges that warrant future research. Besides, our work lays a foundation for the development of more effective and interpretable PE methods.

Yiming Tang, Abrar Anwar, Jesse Thomason

Understanding social signals in multi-party conversations is important for human-robot interaction and artificial social intelligence. Social signals include body pose, head pose, speech, and context-specific activities like acquiring and taking bites of food when dining. Past work in multi-party interaction tends to build task-specific models for predicting social signals. In this work, we address the challenge of predicting multimodal social signals in multi-party settings in a single model. We introduce M3PT, a causal transformer architecture with modality and temporal blockwise attention masking to simultaneously process multiple social cues across multiple participants and their temporal interactions. We train and evaluate M3PT on the Human-Human Commensality Dataset (HHCD), and demonstrate that using multiple modalities improves bite timing and speaking status prediction. Source code: https://github.com/AbrarAnwar/masked-social-signals/.

Harshvardhan Saini, Yiming Tang, Dianbo Liu

Controlling emergent behavioral personas (e.g., sycophancy, hallucination) in Large Language Models (LLMs) is critical for AI safety, yet remains a persistent challenge. Existing solutions face a dilemma: manual prompt engineering is intuitive but unscalable and imprecise, while automatic optimization methods are effective but operate as "black boxes" with no interpretable connection to model internals. We propose a novel framework that adapts gradient ascent to LLMs, enabling targeted prompt discovery. In specific, we propose two methods, RESGA and SAEGA, that both optimize randomly initialized prompts to achieve better aligned representation with an identified persona direction. We introduce fluent gradient ascent to control the fluency of discovered persona steering prompts. We demonstrate RESGA and SAEGA's effectiveness across Llama 3.1, Qwen 2.5, and Gemma 3 for steering three different personas, sycophancy, hallucination, and myopic reward. Crucially, on sycophancy, our automatically discovered prompts achieve significant improvement (49.90% compared with 79.24%). By grounding prompt discovery in mechanistically meaningful features, our method offers a new paradigm for controllable and interpretable behavior modification.

Qiran Zou, Hou Hei Lam, Wenhao Zhao et al.

Large language models (LLMs) have sparked growing interest in automatic machine learning research agents. Among them, agents capable of autonomously proposing ideas and conducting machine learning experiments are particularly promising, as they maximize research automation and accelerate scientific progress by iteratively refining ideas based on experimental results. However, comprehensively evaluating such agents remains challenging. Existing benchmarks tend to overemphasize engineering aspects while neglecting academic rigor, creating barriers that obscure a clear assessment of an agent's scientific capabilities in machine learning research. They also suffer from limited task diversity, an overemphasis on application-oriented tasks over fundamental research problems, and limited scalability to realistic research settings. To address these limitations, we introduce FML-bench, a benchmark designed to evaluate automatic machine learning research agents on 8 diverse and fundamental machine learning research problems. It reduces coding burden, emphasizes fundamental problems rather than specific use cases, offers high task diversity, and is extensible to real-world machine learning GitHub repositories. Furthermore, we present a unified evaluation framework with five complementary metrics, designed to comprehensively assess agent performance on our benchmark. We evaluate state-of-the-art automatic research agents on FML-bench, and find that agents employing broad research exploration strategies outperform those focusing on narrow but deep exploration. These findings suggest that emphasizing the breadth of exploration may lead to more effective research outcomes than focusing solely on incremental refinement. Our benchmark is available at https://github.com/qrzou/FML-bench.

Yiming Tang, Allan Spektor, Raffi Khatchadourian et al.

Logging is a significant programming practice. Due to the highly transactional nature of modern software applications, massive amount of logs are generated every day, which may overwhelm developers. Logging information overload can be dangerous to software applications. Using log levels, developers can print the useful information while hiding the verbose logs during software runtime. As software evolves, the log levels of logging statements associated with the surrounding software feature implementation may also need to be altered. Maintaining log levels necessitates a significant amount of manual effort. In this paper, we demonstrate an automated approach that can rejuvenate feature log levels by matching the interest level of developers in the surrounding features. The approach is implemented as an open-source Eclipse plugin, using two external plug-ins (JGit and Mylyn). It was tested on 18 open-source Java projects consisting of ~3 million lines of code and ~4K log statements. Our tool successfully analyzes 99.22% of logging statements, increases log level distributions by ~20%, and increases the focus of logs in bug fix contexts ~83% of the time. For further details, interested readers can watch our demonstration video (https://www.youtube.com/watch?v=qIULoAXoDv4).

Yiming Tang, Allan Spektor, Raffi Khatchadourian et al.

Logging -- used for system events and security breaches to describe more informational yet essential aspects of software features -- is pervasive. Given the high transactionality of today's software, logging effectiveness can be reduced by information overload. Log levels help alleviate this problem by correlating a priority to logs that can be later filtered. As software evolves, however, levels of logs documenting surrounding feature implementations may also require modification as features once deemed important may have decreased in urgency and vice-versa. We present an automated approach that assists developers in evolving levels of such (feature) logs. The approach, based on mining Git histories and manipulating a degree of interest (DOI) model, transforms source code to revitalize feature log levels based on the "interestingness" of the surrounding code. Built upon JGit and Mylyn, the approach is implemented as an Eclipse IDE plug-in and evaluated on 18 Java projects with $\sim$3 million lines of code and $\sim$4K log statements. Our tool successfully analyzes 99.22% of logging statements, increases log level distributions by $\sim$20%, and increases the focus of logs in bug fix contexts $\sim$83% of the time. Moreover, pull (patch) requests were integrated into large and popular open-source projects. The results indicate that the approach is promising in assisting developers in evolving feature log levels.

Harshvardhan Saini, Samyak Jha, Yiming Tang et al.

Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence. While prior work either aggressively closes this gap or suppresses hallucinations through expensive black-box decoding strategies, none addresses the underlying geometric cause. We provide the first quantitative characterization of this over-alignment, demonstrating that linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace. Building on this insight, we propose two complementary remedies: a training-free inference strategy and a bias-aware fine-tuning paradigm, both of which explicitly project out this subspace from visual representations. Our methods significantly reduce hallucinations across POPE, CHAIR, and AMBER benchmarks, and improve CLAIR scores on long-form captioning tasks, with the training-free variant adding no computational overhead over the base model.

Yulin Luo, Ruichuan An, Bocheng Zou et al.

The distribution of subpopulations is an important property hidden within a dataset. Uncovering and analyzing the subpopulation distribution within datasets provides a comprehensive understanding of the datasets, standing as a powerful tool beneficial to various downstream tasks, including Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Despite its importance, there has been no work that systematically explores the subpopulation distribution of datasets to our knowledge. To address the limitation and solve all the mentioned tasks in a unified way, we introduce a novel concept of subpopulation structures to represent, analyze, and utilize subpopulation distributions within datasets. To characterize the structures in an interpretable manner, we propose the Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework, which employs world knowledge and instruction-following capabilities of Large Language Models (LLMs) to linguistically analyze informative image captions and summarize the structures. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery.

Zhongde An, Jinhong You, Jiyanglin Li et al.

Time series forecasting is essential in a wide range of real world applications. Recently, frequency-domain methods have attracted increasing interest for their ability to capture global dependencies. However, when applied to non-stationary time series, these methods encounter the $\textit{spectral entanglement}$ and the computational burden of complex-valued learning. The $\textit{spectral entanglement}$ refers to the overlap of trends, periodicities, and noise across the spectrum due to $\textit{spectral leakage}$ and the presence of non-stationarity. However, existing decompositions are not suited to resolving spectral entanglement. To address this, we propose the Frequency Decomposition Network (FreDN), which introduces a learnable Frequency Disentangler module to separate trend and periodic components directly in the frequency domain. Furthermore, we propose a theoretically supported ReIm Block to reduce the complexity of complex-valued operations while maintaining performance. We also re-examine the frequency-domain loss function and provide new theoretical insights into its effectiveness. Extensive experiments on seven long-term forecasting benchmarks demonstrate that FreDN outperforms state-of-the-art methods by up to 10\%. Furthermore, compared with standard complex-valued architectures, our real-imaginary shared-parameter design reduces the parameter count and computational cost by at least 50\%.

Yiming Tang, Abhijeet Sinha, Dianbo Liu

Although recent generative models are remarkably capable of producing instruction-following and realistic outputs, they remain prone to notable physical plausibility failures. Though critical in applications, these physical plausibility errors often escape detection by existing evaluation methods. Furthermore, no framework exists for automatically identifying and interpreting specific physical error patterns in natural language, preventing targeted model improvements. We introduce Matryoshka Transcoders, a novel framework for the automatic discovery and interpretation of physical plausibility features in generative models. Our approach extends the Matryoshka representation learning paradigm to transcoder architectures, enabling hierarchical sparse feature learning at multiple granularity levels. By training on intermediate representations from a physical plausibility classifier and leveraging large multimodal models for interpretation, our method identifies diverse physics-related failure modes without manual feature engineering, achieving superior feature relevance and feature accuracy compared to existing approaches. We utilize the discovered visual patterns to establish a benchmark for evaluating physical plausibility in generative models. Our analysis of eight state-of-the-art generative models provides valuable insights into how these models fail to follow physical constraints, paving the way for further model improvements.

Yiming Tang, Harshvardhan Saini, Zhaoqian Yao et al.

As AI models achieve remarkable capabilities across diverse domains, understanding what representations they learn and how they encode concepts has become increasingly important for both scientific progress and trustworthy deployment. Recent works in mechanistic interpretability have widely reported that neural networks represent meaningful concepts as linear directions in their representation spaces and often encode diverse concepts in superposition. Various sparse dictionary learning (SDL) methods, including sparse autoencoders, transcoders, and crosscoders, are utilized to address this by training auxiliary models with sparsity constraints to disentangle these superposed concepts into monosemantic features. These methods are the backbone of modern mechanistic interpretability, yet in practice they consistently produce polysemantic features, feature absorption, and dead neurons, with very limited theoretical understanding of why these phenomena occur. Existing theoretical work is limited to tied-weight sparse autoencoders, leaving the broader family of SDL methods without formal grounding. We develop the first unified theoretical framework that casts all major SDL variants as a single piecewise biconvex optimization problem, and characterize its global solution set, non-identifiability, and spurious optima. This analysis yields principled explanations for feature absorption and dead neurons. To expose these pathologies under full ground-truth access, we introduce the Linear Representation Bench. Guided by our theory, we propose feature anchoring, a novel technique that restores SDL identifiability, substantially improving feature recovery across synthetic benchmarks and real neural representations.

Di Wu, Liu Liu, Xueyu Yuan et al.

Articulated objects are ubiquitous in daily environments, and their 3D reconstruction holds great significance across various fields. However, existing articulated object reconstruction methods typically require costly inputs such as multi-stage and multi-view observations. To address the limitations, we propose a category-agnostic articulated object reconstruction framework via planar Gaussian Splatting, which only uses sparse-view RGB images from a single state. Specifically, we first introduce a Gaussian information field to perceive the optimal sparse viewpoints from candidate camera poses. Then we compress 3D Gaussians into planar Gaussians to facilitate accurate estimation of normal and depth. The planar Gaussians are optimized in a coarse-to-fine manner through depth smooth regularization and few-shot diffusion. Moreover, we introduce a part segmentation probability for each Gaussian primitive and update them by back-projecting part segmentation masks of renderings. Extensive experimental results demonstrate that our method achieves higher-fidelity part-level surface reconstruction on both synthetic and real-world data than existing methods. Codes will be made publicly available.

Yiming Tang, Wenjia Zhong, Rushi Shah et al.

Deep learning models have achieved remarkable accuracy in chest X-ray diagnosis, yet their widespread clinical adoption remains limited by the black-box nature of their predictions. Clinicians require transparent, verifiable explanations to trust automated diagnoses and identify potential failure modes. We introduce CXR-LanIC (Language-Grounded Interpretable Classifier for Chest X-rays), a novel framework that addresses this interpretability challenge through task-aligned pattern discovery. Our approach trains transcoder-based sparse autoencoders on a BiomedCLIP diagnostic classifier to decompose medical image representations into interpretable visual patterns. By training an ensemble of 100 transcoders on multimodal embeddings from the MIMIC-CXR dataset, we discover approximately 5,000 monosemantic patterns spanning cardiac, pulmonary, pleural, structural, device, and artifact categories. Each pattern exhibits consistent activation behavior across images sharing specific radiological features, enabling transparent attribution where predictions decompose into 20-50 interpretable patterns with verifiable activation galleries. CXR-LanIC achieves competitive diagnostic accuracy on five key findings while providing the foundation for natural language explanations through planned large multimodal model annotation. Our key innovation lies in extracting interpretable features from a classifier trained on specific diagnostic objectives rather than general-purpose embeddings, ensuring discovered patterns are directly relevant to clinical decision-making, demonstrating that medical AI systems can be both accurate and interpretable, supporting safer clinical deployment through transparent, clinically grounded explanations.

Yiming Tang, Arash Lagzian, Srinivas Anumasa et al.

The rapid development of generative AI has transformed content creation, communication, and human development. However, this technology raises profound concerns in high-stakes domains, demanding rigorous methods to analyze and evaluate AI-generated content. While existing analytic methods often treat images as indivisible wholes, real-world AI failures generally manifest as specific visual patterns that can evade holistic detection and suit more granular and decomposed analysis. Here we introduce a content analysis tool, Language-Grounded Sparse Encoders (LanSE), which decompose images into interpretable visual patterns with natural language descriptions. Utilizing interpretability modules and large multimodal models, LanSE can automatically identify visual patterns within data modalities. Our method discovers more than 5,000 visual patterns with 93\% human agreement, provides decomposed evaluation outperforming existing methods, establishes the first systematic evaluation of physical plausibility, and extends to medical imaging settings. Our method's capability to extract language-grounded patterns can be naturally adapted to numerous fields, including biology and geography, as well as other data modalities such as protein structures and time series, thereby advancing content analysis for generative AI.

Yiming Tang, Bin Dong

Large language models (LLMs) benefit greatly from prompt engineering, with in-context learning standing as a pivital technique. While former approaches have provided various ways to construct the demonstrations used for in-context learning, they often ignore the inherent heterogeneity within datasets, applying the same demonstrations to all reasoning questions. We observed that the effectiveness of demonstrations varies depending on the specific question. This motivates our exploration of using prompt engineering to select appropriate demonstrations. To address the challenge of automatically creating and choosing demonstrations tailored to each question, we propose a novel prompt engineering workflow built around a novel object called the "demonstration notebook." This notebook helps identify the most suitable in-context learning example for a question by gathering and reusing information from the LLM's past interactions. Our experiments show that this approach outperforms all existing methods for automatic demonstration construction and selection (as far as we know), achieving state-of-the-art results on serveral reasoning benchmarks. The method's versatility is further demonstrated by its success in text summarization and prompt compression tasks. Additionally, we contribute a rigorous analysis method to reveal the "demonstrative regime" of a demonstration, providing valuable insights into how demonstrations relate to different question types within a dataset.