arXiv subjects

Yiheng Su

Publications and source records attributed to Yiheng Su.

12 recordsLinked to original sources

arXiv2026

RELISH: LLM REgression with a Latent Iterative State Head

We present RELISH (REgression with a Latent Iterative State Head), a novel, lightweight architecture designed for text regression with large language models. Rather than decoding numeric targets as text or aggregating multiple generated outputs, RELISH predicts scalar values directly from frozen LLM representations by iteratively refining a learned latent state through cross-attention over token-level representations, and then mapping the final state to a point estimate with a linear regressor. Across six datasets, four LLM backbones, and two LLM training regimes, RELISH consistently outperforms prior baselines from all three major LLM regression families, including autoregressive decoding, regression-aware inference, and existing predictive head methods. Despite these gains, RELISH remains highly parameter-efficient, requiring only $\sim$3.4-3.7M trainable parameters across frozen LLM backbones (only 0.01-0.04$\%$ additional overhead), far less than LoRA-based alternatives that grow with model size (0.26-0.42$\%$). Our code is available at https://github.com/SamSoup/RELISH.

cs.CL↗

arXiv2026

Breaking $1/ε$ Barrier in Quantum Zero-Sum Games: Generalizing Metric Subregularity for Spectraplexes

Quantum zero-sum games provide a framework for non-local games, quantum interactive proofs, and quantum machine learning, where players optimize a bilinear payoff over quantum states. In contrast to classical bilinear games over polyhedral domains, for which gradient methods achieve linear last-iterate convergence, comparable guarantees over spectraplexes have remained open. Recent work achieved only an $O(1/\varepsilon)$ average-iterate rate and suggested that semidefinite geometry may preclude classical-style linear rates. We refute this obstruction. We prove that quantum zero-sum games admit algorithms with $O(\log(1/\varepsilon))$ last-iterate convergence to Nash equilibrium. In particular, matrix variants of Nesterov's iterative smoothing and Optimistic Gradient Descent--Ascent match the asymptotic rate of the classical polyhedral case. The key technical ingredient is a new error-bound theory for semidefinite games, establishing metric subregularity of the relevant monotone operator over spectrahedra despite the absence of polyhedral structure. We also give a geometric characterization of Nash equilibria via slack operators, classifying strategic directions as essential, neutral, or non-essential. Under strict complementarity or nondegeneracy, this reduces to a sharp classical-style dichotomy. Finally, we revisit Optimistic Matrix Multiplicative Weights Update. By extending the Quantal Response Equilibrium framework to spectraplex games, we prove an $\widetilde O(1/\varepsilon)$ last-iterate guarantee, while showing that any $O(\log(1/\varepsilon))$ speedup for this method must depend on a natural, dimension-dependent condition number. Experiments support the theoretical picture, with Optimistic Gradient Descent--Ascent outperforming Optimistic Matrix Multiplicative Weights Update in the regimes studied.

cs.GT↗

arXiv2026

No Coin Left Behind: Maximizing Strategic Surplus Against No-Regret Dynamics

We investigate the strategic surplus obtainable against a Follow-the-Regularized-Leader (FTRL) learner with constant step size $η$ in $n\times m$ two-player zero-sum games played over $T$ rounds against a clairvoyant optimizer. In contrast with prior analysis, we show that the extraction of such regret-scale surplus is an inherent feature of the FTRL family, rather than an artifact of specific instantiations. First, for a fixed max-min optimizer, we establish a sweeping law of order $Ω(N_{\mathrm{sub}}/η)$, proving that utility surplus scales with the number of the learner's suboptimal actions $N$ and vanishes in their absence. Second, for an alternating optimizer, a surplus of $Ω(ηT/\mathrm{poly}(n,m))$ can be guaranteed regardless of the equilibrium structure, with high probability, in random games. Our analysis uncovers a sharp geometric dichotomy: non-steep regularizers allow the optimizer to realize the maximal transient surplus via finite-time elimination of suboptimal actions, whereas steep regularizers introduce a vanishing tail correction that can delay surplus saturation. Finally, we discuss whether this leverage persists under bilateral payoff uncertainty and propose a susceptibility measure quantifying which regularizers are most vulnerable to learner-aware strategic steering.

cs.GT↗

arXiv2026

Rhetorical XAI: Explaining AI's Benefits as well as its Use via Rhetorical Design

We explore potential benefits of incorporating Rhetorical Design into the design of Explainable Artificial Intelligence (XAI) systems. While XAI is traditionally framed around explaining individual predictions or overall system behavior, explanations may also function as rhetorical arguments that shape how users evaluate a system's usefulness and credibility, and how they develop appropriate trust for adoption. In real-world, in-situ interactions, explanations can thus produce experiential and affective rhetorical effects that are not fully captured by traditional XAI design goals that focus primarily on how AI works. To address this gap, we propose Rhetorical XAI, which bridges two explanatory goals: how AI works and why AI merits use. Rhetorical XAI comprises three appeals in explanation design: logos, which aligns technical logic with human reasoning through visual and textual abstractions; ethos, which establishes contextual credibility based on the explanation source and its appropriateness to the decision task; and pathos, which engages user emotionally by framing explanations around their motivations, expectations, or situated needs during interaction. We conduct a narrative review synthesizing design strategies from prior XAI work aligned with these three rhetorical appeals, highlighting both opportunities and challenges of integrating rhetorical design into XAI.

cs.HC↗

arXiv2026

PIE: Performance Interval Estimation for Free-Form Generation Tasks

Confidence estimation infers a probability for whether each model output is correct or not. While predicting such binary correctness is sensible for tasks with exact answers, free-form generation tasks are often more nuanced, with output quality being both fine-grained and multi-faceted. We thus propose Performance Interval Estimation (PIE) to predict both: 1) point estimates for any arbitrary set of continuous-valued evaluation metrics; and 2) calibrated uncertainty intervals around these point estimates. We then compare two approaches: LLM-as-judge vs. classic regression with confidence estimation features. Evaluation over 11 datasets spans summarization, translation, code generation, function-calling, and question answering. Regression is seen to achieve both: i) lower error point estimates of metric scores; and ii) well-calibrated uncertainty intervals. To support reproduction and follow-on work, we share our data and code.

cs.CL↗

arXiv2025

Diverse beam search to find densest-known planar unit distance graphs

This paper addresses the problem of determining the maximum number of edges in a unit distance graph (UDG) of $n$ vertices using computer search. An unsolved problem of Paul Erdős asks the maximum number of edges $u(n)$ a UDG of $n$ vertices can have. Those UDGs that attain $u(n)$ are called "maximally dense." In this paper, we seek to demonstrate a computer algorithm to generate dense UDGs for vertex counts up to at least 100. Via beam search with an added visitation metric, our algorithm finds all known maximally dense UDGs up to isomorphism at the push of a button. In addition, for $15 < n$, where $u(n)$ is unknown, i) the algorithm finds all previously published densest UDGs up to isomorphism for $15 < n \le 30$, and ii) the rate of growth of $u(n)/n$ remains similar for $30 < n$. The code and database of over 60 million UDGs found by our algorithm can be found at https://codeberg.org/zsamboki/dbs-udg.

math.CO↗

arXiv2025

Can Large Language Models Grasp Concepts in Visual Content? A Case Study on YouTube Shorts about Depression

Large language models (LLMs) are increasingly used to assist computational social science research. While prior efforts have focused on text, the potential of leveraging multimodal LLMs (MLLMs) for online video studies remains underexplored. We conduct one of the first case studies on MLLM-assisted video content analysis, comparing AI's interpretations to human understanding of abstract concepts. We leverage LLaVA-1.6 Mistral 7B to interpret four abstract concepts regarding video-mediated self-disclosure, analyzing 725 keyframes from 142 depression-related YouTube short videos. We perform a qualitative analysis of MLLM's self-generated explanations and found that the degree of operationalization can influence MLLM's interpretations. Interestingly, greater detail does not necessarily increase human-AI alignment. We also identify other factors affecting AI alignment with human understanding, such as concept complexity and versatility of video genres. Our exploratory study highlights the need to customize prompts for specific concepts and calls for researchers to incorporate more human-centered evaluations when working with AI systems in a multimodal context.

cs.HC↗

arXiv2025

Displaying Fear, Sadness, and Joy in Public: Schizophrenia Vloggers' Video Narration of Emotion and Online Care-Seeking

Individuals with severe mental illnesses (SMI), particularly schizophrenia, experience complex and intense emotions frequently. They increasingly turn to vlogging as an authentic medium for emotional disclosure and online support-seeking. While previous research has primarily focused on text-based disclosure, little is known about how people construct narratives around emotions and emotional experiences through video blogs. Our study analyzed 401 YouTube videos created by schizophrenia vloggers, revealing that vloggers disclosed their fear, sadness, and joy through verbal narration by explicit expressions or storytelling. Visually, they employed various framing styles, including Anonymous, Talk-to-Camera, and In-the-Moment approaches, along with diverse visual narration techniques. Notably, we uncovered a concerning 'visual appeal disparity' in audience engagement, with visually appealing videos receiving significantly more views, likes, and comments. This study discusses the role of video-sharing platforms in emotional expression and offers design implications for fostering online care-seeking for emotionally vulnerable populations.

cs.HC↗

arXiv2024

Wrapper Boxes: Faithful Attribution of Model Predictions to Training Data

Can we preserve the accuracy of neural models while also providing faithful explanations of model decisions to training data? We propose a "wrapper box'' pipeline: training a neural model as usual and then using its learned feature representation in classic, interpretable models to perform prediction. Across seven language models of varying sizes, including four large language models (LLMs), two datasets at different scales, three classic models, and four evaluation metrics, we first show that the predictive performance of wrapper classic models is largely comparable to the original neural models. Because classic models are transparent, each model decision is determined by a known set of training examples that can be directly shown to users. Our pipeline thus preserves the predictive performance of neural language models while faithfully attributing classic model decisions to training data. Among other use cases, such attribution enables model decisions to be contested based on responsible training instances. Compared to prior work, our approach achieves higher coverage and correctness in identifying which training data to remove to change a model decision. To reproduce findings, our source code is online at: https://github.com/SamSoup/WrapperBox.

cs.LG↗

arXiv2024

Convergence, optimization and stability of singular eigenmaps

Eigenmaps are important in analysis, geometry, and machine learning, especially in nonlinear dimension reduction. Approximation of the eigenmaps of a Laplace operator depends crucially on the scaling parameter $ε$. If $ε$ is too small or too large, then the approximation is inaccurate or completely breaks down. However, an analytic expression for the optimal $ε$ is out of reach. In our work, we use some explicitly solvable models and Monte Carlo simulations to find the approximately optimal range of $ε$ that gives, on average, relatively accurate approximation of the eigenmaps. Numerically we can consider several model situations where eigen-coordinates can be computed analytically, including intervals with uniform and weighted measures, squares, tori, spheres, and the Sierpinski gasket. In broader terms, we intend to study eigen-coordinates on weighted Riemannian manifolds, possibly with boundary, and on some metric measure spaces, such as fractals.

math.PR↗

arXiv2024

Harnessing LLMs for Automated Video Content Analysis: An Exploratory Workflow of Short Videos on Depression

Despite the growing interest in leveraging Large Language Models (LLMs) for content analysis, current studies have primarily focused on text-based content. In the present work, we explored the potential of LLMs in assisting video content analysis by conducting a case study that followed a new workflow of LLM-assisted multimodal content analysis. The workflow encompasses codebook design, prompt engineering, LLM processing, and human evaluation. We strategically crafted annotation prompts to get LLM Annotations in structured form and explanation prompts to generate LLM Explanations for a better understanding of LLM reasoning and transparency. To test LLM's video annotation capabilities, we analyzed 203 keyframes extracted from 25 YouTube short videos about depression. We compared the LLM Annotations with those of two human coders and found that LLM has higher accuracy in object and activity Annotations than emotion and genre Annotations. Moreover, we identified the potential and limitations of LLM's capabilities in annotating videos. Based on the findings, we explore opportunities and challenges for future research and improvements to the workflow. We also discuss ethical concerns surrounding future studies based on LLM-assisted video analysis.

cs.HC↗

arXiv2020

Matching model of flow table for networked big data

Networking for big data has to be intelligent because it will adjust data transmission requirements adaptively during data splitting and merging. Software-defined networking (SDN) provides a workable and practical paradigm for designing more efficient and flexible networks. Matching strategy in the flow table of SDN switches is most crucial. In this paper, we use a classification approach to analyze the structure of packets based on the tuple-space lookup mechanism, and propose a matching model of the flow table in SDN switches by classifying packets based on a set of fields, which is called an F-OpenFlow. The experiment results show that the proposed F-OpenFlow effectively improves the utilization rate and matching efficiency of the flow table in SDN switches for networked big data.

cs.NI↗