arXiv subjects

Chongyang Gao

Publications and source records attributed to Chongyang Gao.

At least 19 recordsLinked to original sources

arXiv2026

FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents

Language model agents now execute bounded tasks reliably. Whether they can sustain effective decision-making over long horizons, where actions have cumulative consequences and the environment responds to their choices, remains largely unmeasured. FM-Bench (Football Management Benchmark) measures this. An LLM agent runs a football club for 20 in-game years through 26 tools and roughly 340 to 400 decision stops. It drafts a squad on the same budget as every rival, trades players, negotiates contracts, invests in facilities and youth, sets lineups, and answers to a board that can fire it, while a deterministic engine accumulates every year into one final score with no LLM judge or human rater. The solo track plays each of 15 frontier models against a frozen scripted world, and the Arena places the same models plus a scripted anchor in one shared 20-year world; to our knowledge, the first head-to-head evaluation at this scale. We measure six behavioral capabilities behind the score. Across three seeds, all 15 models complete every horizon while the blind scripted baselines die out in most of theirs, and claude-fable-5 tops the solo board on mean score and the Arena, where the title nonetheless rotates among ten models. Neither scale, price, nor vendor predicts the order; the order settles only late in the horizon, and the best first-play human lands only at the bottom of the model board. What separates the models is managerial behavior rather than computation. Higher-scoring models reduce slow-payoff investment near the end, keep cash invested rather than idle, and open renewals well before the deadline, while token spend predicts nothing. No model learns the market's hidden prices from hundreds of rejected bids, and self-managed memory fails in two opposite modes: an archive that only grows or a plan rewritten every season. Code is available at https://github.com/Analogy-AI/fm-bench.

cs.AI↗

arXiv2026

Equitable System-Prompt Selection via Constrained Mixed-Strategy GroupDRO

Large language models are increasingly used for information seeking, yet semantically equivalent questions phrased in different ways can receive answers of considerably different quality. System prompts are widely employed to steer response behavior, but they are typically optimized for average-case quality, so some question phrasings may still receive incomplete or low-quality answers. To address this, we formulate a constrained mixed-strategy GroupDRO framework for system-prompt selection. Instead of optimizing the system-prompt text, the framework assigns weights to system prompts in an existing pool to minimize the worst-case information-quality loss across evaluation metrics and groups, while constraining the mean loss to stay close to that of average-based selection. Because pool generation and selection are decoupled, the method applies to any system-prompt pool and can leverage an ensemble of complementary system prompts rather than a single one. Across five LLMs on two bilingual medical and consumer-finance benchmarks, the constrained method reduces the Overall Mean, Worst 25% Mean, and Worst by 13.1%, 13.2%, and 13.7% on average relative to no mitigation while keeping overall quality close to Average selection. Its multi-prompt weights reveal complementarity across metric-group pairs. Code and data are available at https://github.com/Rainxu09/equitable-system-prompt-selection.

cs.CL↗

arXiv2026

Rethinking LLM-Judged Helpfulness as a Pedagogy Signal: A Pre-Registered Audit Across Tutor Models

LLM tutoring poses a measurement problem: can a general-purpose helpfulness rubric distinguish direct answer-giving from pedagogical guidance? We audit this signal in a pre-registered study. Within each of three tutor bases, we compare conversational and pedagogical policies instantiated with the same underlying model and paired with one fixed weak simulated student. Deterministic detectors measure answer leakage and next-turn independent work. Claude Opus 4.8 is the frozen, condition-blind primary judge. After the Opus scores were fixed, GPT-5.6 Sol was prospectively specified for a post hoc robustness audit of the same 1,179 confirmatory answer-phase tutor turns under the frozen helpfulness and pedagogy rubrics. On the primary base under Opus, the policies do not differ significantly in helpfulness but are perfectly rank-separated under the pedagogy rubric (Cliff's $|\delta|{=}0.10$ vs. $1.0$). Across the two judges, pedagogy contrasts retain their direction where detected, whereas the helpfulness ordering is judge-contingent, reversing between judges on two of three bases. In an Opus-only ablation, seven primary-base policies span $2.3$ points in mean judged pedagogy within a $0.25$-point band of mean judged helpfulness. Separately, answer-revealing turns are followed by less independent student work on every base, a result that is judge-invariant by construction. In this controlled setting, general-purpose helpfulness is not a reliable pedagogy signal. Tutor evaluation should pair pedagogy-targeted rubrics with deterministic process measures.

cs.CL↗

arXiv2026

MIRA: A Bilingual Benchmark for Medical Information Response Audit

Existing safety evaluations for large language models overlook whether responses preserve comparable medical information across different user phrasings of the same question. To address this, we introduce the Medical Information Response Audit (MIRA), a bilingual, controlled benchmark that assesses whether LLMs provide comparable medical information across user-side language, register, and health literacy signals. MIRA contains 4,320 prompts built from 60 medically reviewed, low-risk health questions. Across five mainstream LLMs, models answered all medical questions, but responses to low health-literacy signals consistently omitted more key information, provided fewer concrete next steps, and offered less support for independent judgment. We term this pattern Differential Information Dilution (DID). A comparison with 300 real-world health queries provides preliminary evidence of rank-order validity. A knowledge-guided mitigation prompt reduces information dilution for most models, with the largest reductions in underinformative simplification observed for Claude (~8%) and Qwen (~6%). Code and data are available at https://github.com/Rainxu09/MIRA.

cs.AI↗

arXiv2026

Exploiting Label-Independent Regularization from Spatial Dependencies for Whole Slide Image Analysis

Whole slide images, with their gigapixel-scale panoramas of tissue samples, are pivotal for precise disease diagnosis. However, their analysis is hindered by immense data size and scarce annotations. Existing MIL methods face challenges due to the fundamental imbalance where a single bag-level label must guide the learning of numerous patch-level features. This sparse supervision makes it difficult to reliably identify discriminative patches during training, leading to unstable optimization and suboptimal solutions. We propose a spatially regularized MIL framework that leverages inherent spatial relationships among patch features as label-independent regularization signals. Our approach learns a shared representation space by jointly optimizing feature-induced spatial reconstruction and label-guided classification objectives, enforcing consistency between intrinsic structural patterns and supervisory signals. Experimental results on multiple public datasets demonstrate significant improvements over state-of-the-art methods, offering a promising direction.

cs.CV↗

arXiv2026

Classroom Final Exam: An Instructor-Tested Reasoning Benchmark

We introduce CFE-Bench (Classroom Final Exam), a multimodal benchmark for evaluating the reasoning capabilities of large language models across more than 20 STEM domains. CFE-Bench is curated from repeatedly used, authentic university homework and exam problems, paired with reference solutions provided by course instructors. CFE-Bench remains challenging for frontier models: the newly released Gemini-3.1-pro-preview achieves 59.69% overall accuracy, while the second-best model, Gemini-3-flash-preview, reaches 55.46%, leaving substantial room for improvement. Beyond aggregate scores, we conduct a diagnostic analysis by decomposing instructor reference solutions into structured reasoning flows. We find that while frontier models often answer intermediate sub-questions correctly, they struggle to reliably derive and maintain correct intermediate states throughout multi-step solutions. We further observe that model-generated solutions typically contain more reasoning steps than instructor solutions, indicating lower step efficiency and a higher risk of error accumulation. Data and code are available at https://github.com/Analogy-AI/CFE_Bench.

cs.AI↗

arXiv2026

Context and Transcripts Improve Detection of Deepfake Audios of Public Figures

Humans use context to assess the veracity of information. However, current audio deepfake detectors only analyze the audio file without considering either context or transcripts. We create and analyze a Journalist-provided Deepfake Dataset (JDD) of 255 public deepfakes which were primarily contributed by over 70 journalists since early 2024. We also generate a synthetic audio dataset (SYN) of dead public figures and propose a novel Context-based Audio Deepfake Detector (CADD) architecture. In addition, we evaluate performance on two large-scale datasets: ITW and P$^2$V. We show that sufficient context and/or the transcript can significantly improve the efficacy of audio deepfake detectors. Performance (measured via F1 score, AUC, and EER) of multiple baseline audio deepfake detectors and traditional classifiers can be improved by 5%-37.58% in F1-score, 3.77%-42.79% in AUC, and 6.17%-47.83% in EER. We additionally show that CADD, via its use of context and/or transcripts, is more robust to 5 adversarial evasion strategies, limiting performance degradation to an average of just -0.71% across all experiments. Code, models, and datasets are available at our project page: https://sites.northwestern.edu/nsail/cadd-context-based-audio-deepfake-detection (access restricted during review).

cs.AI↗

arXiv2025

Insight Miner: A Time Series Analysis Dataset for Cross-Domain Alignment with Natural Language

Time-series data is critical across many scientific and industrial domains, including environmental analysis, agriculture, transportation, and finance. However, mining insights from this data typically requires deep domain expertise, a process that is both time-consuming and labor-intensive. In this paper, we propose \textbf{Insight Miner}, a large-scale multimodal model (LMM) designed to generate high-quality, comprehensive time-series descriptions enriched with domain-specific knowledge. To facilitate this, we introduce \textbf{TS-Insights}\footnote{Available at \href{https://huggingface.co/datasets/zhykoties/time-series-language-alignment}{https://huggingface.co/datasets/zhykoties/time-series-language-alignment}.}, the first general-domain dataset for time series and language alignment. TS-Insights contains 100k time-series windows sampled from 20 forecasting datasets. We construct this dataset using a novel \textbf{agentic workflow}, where we use statistical tools to extract features from raw time series before synthesizing them into coherent trend descriptions with GPT-4. Following instruction tuning on TS-Insights, Insight Miner outperforms state-of-the-art multimodal models, such as LLaVA \citep{liu2023llava} and GPT-4, in generating time-series descriptions and insights. Our findings suggest a promising direction for leveraging LMMs in time series analysis, and serve as a foundational step toward enabling LLMs to interpret time series as a native input modality.

cs.LG↗

arXiv2025

Beyond Sharp Minima: Robust LLM Unlearning via Feedback-Guided Multi-Point Optimization

Current LLM unlearning methods face a critical security vulnerability that undermines their fundamental purpose: while they appear to successfully remove sensitive or harmful knowledge, this ``forgotten" information remains precariously recoverable through relearning attacks. We identify that the root cause is that conventional methods optimizing the forgetting loss at individual data points will drive model parameters toward sharp minima in the loss landscape. In these unstable regions, even minimal parameter perturbations can drastically alter the model's behaviors. Consequently, relearning attacks exploit this vulnerability by using just a few fine-tuning samples to navigate the steep gradients surrounding these unstable regions, thereby rapidly recovering knowledge that was supposedly erased. This exposes a critical robustness gap between apparent unlearning and actual knowledge removal. To address this issue, we propose StableUN, a bi-level feedback-guided optimization framework that explicitly seeks more stable parameter regions via neighborhood-aware optimization. It integrates forgetting feedback, which uses adversarial perturbations to probe parameter neighborhoods, with remembering feedback to preserve model utility, aligning the two objectives through gradient projection. Experiments on WMDP and MUSE benchmarks demonstrate that our method is significantly more robust against both relearning and jailbreaking attacks while maintaining competitive utility performance.

cs.LG↗

arXiv2025

Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuning

As large language models (LLMs) continue to grow in capability, so do the risks of harmful misuse through fine-tuning. While most prior studies assume that attackers rely on supervised fine-tuning (SFT) for such misuse, we systematically demonstrate that reinforcement learning (RL) enables adversaries to more effectively break safety alignment and facilitate more advanced harmful task assistance, under matched computational budgets. To counter this emerging threat, we propose TokenBuncher, the first effective defense specifically targeting RL-based harmful fine-tuning. TokenBuncher suppresses the foundation on which RL relies: model response entropy. By constraining entropy, RL-based fine-tuning can no longer exploit distinct reward signals to drive the model toward harmful behaviors. We realize this defense through entropy-as-reward RL and a Token Noiser mechanism designed to prevent the escalation of harmful capabilities. Extensive experiments across multiple models and RL algorithms show that TokenBuncher robustly mitigates harmful RL fine-tuning while preserving benign task performance and finetunability. Our results highlight that RL-based harmful fine-tuning poses a greater systemic risk than SFT, and that TokenBuncher provides an effective and general defense.

cs.LG↗

arXiv2025

Perturbed Public Voices (P$^{2}$V): A Dataset for Robust Audio Deepfake Detection

Current audio deepfake detectors cannot be trusted. While they excel on controlled benchmarks, they fail when tested in the real world. We introduce Perturbed Public Voices (P$^{2}$V), an IRB-approved dataset capturing three critical aspects of malicious deepfakes: (1) identity-consistent transcripts via LLMs, (2) environmental and adversarial noise, and (3) state-of-the-art voice cloning (2020-2025). Experiments reveal alarming vulnerabilities of 22 recent audio deepfake detectors: models trained on current datasets lose 43% performance when tested on P$^{2}$V, with performance measured as the mean of F1 score on deepfake audio, AUC, and 1-EER. Simple adversarial perturbations induce up to 16% performance degradation, while advanced cloning techniques reduce detectability by 20-30%. In contrast, P$^{2}$V-trained models maintain robustness against these attacks while generalizing to existing datasets, establishing a new benchmark for robust audio deepfake detection. P$^{2}$V will be publicly released upon acceptance by a conference/journal.

cs.SD↗

arXiv2025

Assessing and Mitigating Medical Knowledge Drift and Conflicts in Large Language Models

Large Language Models (LLMs) have great potential in the field of health care, yet they face great challenges in adapting to rapidly evolving medical knowledge. This can lead to outdated or contradictory treatment suggestions. This study investigated how LLMs respond to evolving clinical guidelines, focusing on concept drift and internal inconsistencies. We developed the DriftMedQA benchmark to simulate guideline evolution and assessed the temporal reliability of various LLMs. Our evaluation of seven state-of-the-art models across 4,290 scenarios demonstrated difficulties in rejecting outdated recommendations and frequently endorsing conflicting guidance. Additionally, we explored two mitigation strategies: Retrieval-Augmented Generation and preference fine-tuning via Direct Preference Optimization. While each method improved model performance, their combination led to the most consistent and reliable results. These findings underscore the need to improve LLM robustness to temporal shifts to ensure more dependable applications in clinical practice. The dataset is available at https://huggingface.co/datasets/RDBH/DriftMed.

cs.CL↗

arXiv2024

AlphaLoRA: Assigning LoRA Experts Based on Layer Training Quality

Parameter-efficient fine-tuning methods, such as Low-Rank Adaptation (LoRA), are known to enhance training efficiency in Large Language Models (LLMs). Due to the limited parameters of LoRA, recent studies seek to combine LoRA with Mixture-of-Experts (MoE) to boost performance across various tasks. However, inspired by the observed redundancy in traditional MoE structures, previous studies identify similar redundancy among LoRA experts within the MoE architecture, highlighting the necessity for non-uniform allocation of LoRA experts across different layers. In this paper, we leverage Heavy-Tailed Self-Regularization (HT-SR) Theory to design a fine-grained allocation strategy. Our analysis reveals that the number of experts per layer correlates with layer training quality, which exhibits significant variability across layers. Based on this, we introduce AlphaLoRA, a theoretically principled and training-free method for allocating LoRA experts to further mitigate redundancy. Experiments on three models across ten language processing and reasoning benchmarks demonstrate that AlphaLoRA achieves comparable or superior performance over all baselines. Our code is available at https://github.com/morelife2017/alphalora.

cs.CL↗

arXiv2024

On Large Language Model Continual Unlearning

While large language models have demonstrated impressive performance across various domains and tasks, their security issues have become increasingly severe. Machine unlearning has emerged as a representative approach for model safety and security by removing the influence of undesired data on the target model. However, these methods do not sufficiently consider that unlearning requests in real-world scenarios are continuously emerging, especially in the context of LLMs, which may lead to accumulated model utility loss that eventually becomes unacceptable. Moreover, existing LLM unlearning methods often ignore previous data access limitations due to privacy concerns and copyright protection. Without previous data, the utility preservation during unlearning is much harder. To overcome these challenges, we propose the OOO framework that includes an Orthogonal low-rank adapter (LoRA) for continually unlearning requested data and an Out-Of-Distribution (OOD) detector to measure the similarity between input and unlearning data. The orthogonal LoRA achieves parameter disentanglement among continual unlearning requests. The OOD detector is trained with a novel contrastive entropy loss and utilizes a glocal-aware scoring mechanism. During inference, our OOO framework can decide whether and to what extent to load the unlearning LoRA based on the OOD detector's predicted similarity between the input and the unlearned knowledge. Notably, OOO's effectiveness does not rely on any retained data. We conducted extensive experiments on OOO and state-of-the-art LLM unlearning methods across three tasks and seven datasets. The results indicate that OOO consistently achieves the best unlearning effectiveness and utility preservation, especially when facing continuous unlearning requests. The source codes can be found at https://github.com/GCYZSL/O3-LLM-UNLEARNING.

cs.LG↗

arXiv2024

Learning Spatial-Preserving Hierarchical Representations for Digital Pathology

Whole slide images (WSIs) pose fundamental computational challenges due to their gigapixel resolution and the sparse distribution of informative regions. Existing approaches often treat image patches independently or reshape them in ways that distort spatial context, thereby obscuring the hierarchical pyramid representations intrinsic to WSIs. We introduce Sparse Pyramid Attention Networks (SPAN), a hierarchical framework that preserves spatial relationships while allocating computation to informative regions. SPAN constructs multi-scale representations directly from single-scale inputs, enabling precise hierarchical modeling of WSI data. We demonstrate SPAN's versatility through two variants: SPAN-MIL for slide classification and SPAN-UNet for segmentation. Comprehensive evaluations across multiple public datasets show that SPAN effectively captures hierarchical structure and contextual relationships. Our results provide clear evidence that architectural inductive biases and hierarchical representations enhance both slide-level and patch-level performance. By addressing key computational challenges in WSI analysis, SPAN provides an effective framework for computational pathology and demonstrates important design principles for large-scale medical image analysis.

cs.CV↗

arXiv2024

Exploring the Distinctiveness and Fidelity of the Descriptions Generated by Large Vision-Language Models

Large Vision-Language Models (LVLMs) are gaining traction for their remarkable ability to process and integrate visual and textual data. Despite their popularity, the capacity of LVLMs to generate precise, fine-grained textual descriptions has not been fully explored. This study addresses this gap by focusing on \textit{distinctiveness} and \textit{fidelity}, assessing how models like Open-Flamingo, IDEFICS, and MiniGPT-4 can distinguish between similar objects and accurately describe visual features. We proposed the Textual Retrieval-Augmented Classification (TRAC) framework, which, by leveraging its generative capabilities, allows us to delve deeper into analyzing fine-grained visual description generation. This research provides valuable insights into the generation quality of LVLMs, enhancing the understanding of multimodal language models. Notably, MiniGPT-4 stands out for its better ability to generate fine-grained descriptions, outperforming the other two models in this aspect. The code is provided at \url{https://anonymous.4open.science/r/Explore_FGVDs-E277}.

cs.CV↗

arXiv2024

Higher Layers Need More LoRA Experts

Parameter-efficient tuning (PEFT) techniques like low-rank adaptation (LoRA) offer training efficiency on Large Language Models, but their impact on model performance remains limited. Recent efforts integrate LoRA and Mixture-of-Experts (MoE) to improve the performance of PEFT methods. Despite promising results, research on improving the efficiency of LoRA with MoE is still in its early stages. Recent studies have shown that experts in the MoE architecture have different strengths and also exhibit some redundancy. Does this statement also apply to parameter-efficient MoE? In this paper, we introduce a novel parameter-efficient MoE method, \textit{\textbf{M}oE-L\textbf{o}RA with \textbf{L}ayer-wise Expert \textbf{A}llocation (MoLA)} for Transformer-based models, where each model layer has the flexibility to employ a varying number of LoRA experts. We investigate several architectures with varying layer-wise expert configurations. Experiments on six well-known NLP and commonsense QA benchmarks demonstrate that MoLA achieves equal or superior performance compared to all baselines. We find that allocating more LoRA experts to higher layers further enhances the effectiveness of models with a certain number of experts in total. With much fewer parameters, this allocation strategy outperforms the setting with the same number of experts in every layer. This work can be widely used as a plug-and-play parameter-efficient tuning approach for various applications. The code is available at https://github.com/GCYZSL/MoLA.

cs.CL↗

arXiv2024

Knowledge from Large-Scale Protein Contact Prediction Models Can Be Transferred to the Data-Scarce RNA Contact Prediction Task

RNA, whose functionality is largely determined by its structure, plays an important role in many biological activities. The prediction of pairwise structural proximity between each nucleotide of an RNA sequence can characterize the structural information of the RNA. Historically, this problem has been tackled by machine learning models using expert-engineered features and trained on scarce labeled datasets. Here, we find that the knowledge learned by a protein-coevolution Transformer-based deep neural network can be transferred to the RNA contact prediction task. As protein datasets are orders of magnitude larger than those for RNA contact prediction, our findings and the subsequent framework greatly reduce the data scarcity bottleneck. Experiments confirm that RNA contact prediction through transfer learning using a publicly available protein model is greatly improved. Our findings indicate that the learned structural patterns of proteins can be transferred to RNAs, opening up potential new avenues for research.

q-bio.QM↗