Searcharxiv⌕ Search

arXiv subjects

Haoyue Bai

Publications and source records attributed to Haoyue Bai.

At least 19 recordsLinked to original sources
arXiv2026

Understanding Why Language Models Hallucinate: Testing Reasoning Against Priors

Large language models often produce hallucinated answers that violate prompt-level constraints. A key diagnostic question is whether these failures reflect missing knowledge, or whether the model has the relevant information but follows the wrong inference path. We study this phenomenon as inference misalignment: a mismatch between the answer supported by the prompt and the answer favored by statistically salient latent associations. We formalize this view with a latent key-task model, in which pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. We introduce TrapQA, a controlled diagnostic testbed with two components. ScientistQA tests disambiguation among similar scientists with supplementary factual probes, while Real-Life Constrained QA tests everyday constraint following under salient shortcuts. Our results show that hallucination can arise from biased latent inference rather than absent knowledge alone.

cs.CL↗
arXiv2026

The Long-Horizon Task Mirage? Diagnosing Where and Why Agentic Systems Break

Large language model (LLM) agents perform strongly on short- and mid-horizon tasks, but often break down on long-horizon tasks that require extended, interdependent action sequences. Despite rapid progress in agentic systems, these long-horizon failures remain poorly characterized, hindering principled diagnosis and comparison across domains. To address this gap, we introduce HORIZON, an initial cross-domain diagnostic benchmark for systematically constructing tasks and analyzing long-horizon failure behaviors in LLM-based agents. Using HORIZON, we evaluate state-of-the-art (SOTA) agents from multiple model families (GPT-5 variants and Claude models), collecting 3100+ trajectories across four representative agentic domains to study horizon-dependent degradation patterns. We further propose a trajectory-grounded LLM-as-a-Judge pipeline for scalable and reproducible failure attribution, and validate it with human annotation on trajectories, achieving strong agreement (inter-annotator κ=0.61; human-judge κ=0.84). Our findings offer an initial methodological step toward systematic, cross-domain analysis of long-horizon agent failures and offer practical guidance for building more reliable long-horizon agents. We release our project website at \href{https://xwang2775.github.io/horizon-leaderboard/}{HORIZON Leaderboard} and welcome contributions from the community.

cs.AI↗
arXiv2026

Expert-Choice Routing Enables Adaptive Computation in Diffusion Language Models

Diffusion language models (DLMs) enable parallel, non-autoregressive text generation, yet existing DLM mixture-of-experts (MoE) models inherit token-choice (TC) routing from autoregressive systems, leading to load imbalance and rigid computation allocation. We show that expert-choice (EC) routing is a better fit for DLMs: it provides deterministic load balancing by design, yielding higher throughput and faster convergence than TC. Building on the property that EC capacity is externally controllable, we introduce timestep-dependent expert capacity, which varies expert allocation according to the denoising step. We find that allocating more capacity to low-mask-ratio steps consistently achieves the best performance under matched FLOPs, and provide a mechanistic explanation: tokens in low-mask-ratio contexts exhibit an order-of-magnitude higher learning efficiency, so concentrating compute on these steps yields the largest marginal return. Finally, we show that existing pretrained TC DLMs can be retrofitted to EC by replacing only the router, achieving faster convergence and improved accuracy across diverse downstream tasks. Together, these results establish EC routing as a superior paradigm for DLM MoE models and demonstrate that computation in DLMs can be treated as an adaptive policy rather than a fixed architectural constant. Code is available at https://github.com/zhangshuibai/EC-DLM.

cs.LG↗
arXiv2026

StressWeb: A Diagnostic Benchmark for Web Agent Robustness under Realistic Interaction Variability

Large language model-based web agents have demonstrated strong performance on realistic web interaction tasks. However, existing evaluations are predominantly conducted under relatively stable and well-behaved interaction conditions, which may overestimate agent robustness. High task success in such idealized settings does not necessarily reflect performance under realistic web interaction. To address this limitation, we introduce a diagnostic stress-testing benchmark for web agents. We first construct realistic and controllable web environments that provide clean and stable interaction workflows as reference baselines. We then introduce structured and controlled perturbations that emulate interaction variability, including shifting layouts, altered interaction semantics, and execution disruptions. By comparing agent behavior between clean and perturbed settings, our framework enables systematic diagnosis of robustness under what-if interaction scenarios. Through extensive evaluation of state-of-the-art multimodal web agents, we show that stress-based evaluation exposes failure modes and substantial robustness gaps that remain hidden under clean benchmark conditions.

cs.SE↗
arXiv2026

Learning to Route: A Rule-Driven Agent Framework for Hybrid-Source Retrieval-Augmented Generation

Large Language Models (LLMs) have shown remarkable performance on general Question Answering (QA), yet they often struggle in domain-specific scenarios where accurate and up-to-date information is required. Retrieval-Augmented Generation (RAG) addresses this limitation by enriching LLMs with external knowledge, but existing systems primarily rely on unstructured documents, while largely overlooking relational databases, which provide precise, timely, and efficiently queryable factual information, serving as indispensable infrastructure in domains such as finance, healthcare, and scientific research. Motivated by this gap, we conduct a systematic analysis that reveals three central observations: (i) databases and documents offer complementary strengths across queries, (ii) naively combining both sources introduces noise and cost without consistent accuracy gains, and (iii) selecting the most suitable source for each query is crucial to balance effectiveness and efficiency. We further observe that query types show consistent regularities in their alignment with retrieval paths, suggesting that routing decisions can be effectively guided by systematic rules that capture these patterns. Building on these insights, we propose a rule-driven routing framework. A routing agent scores candidate augmentation paths based on explicit rules and selects the most suitable one; a rule-making expert agent refines the rules over time using QA feedback to maintain adaptability; and a path-level meta-cache reuses past routing decisions for semantically similar queries to reduce latency and cost. Experiments on three QA benchmarks demonstrate that our framework consistently outperforms static strategies and learned routing baselines, achieving higher accuracy while maintaining moderate computational cost.

cs.CL↗
arXiv2026

Climbing the Ladder of Reasoning: What LLMs Can-and Still Can't-Solve after SFT?

Recent supervised fine-tuning (SFT) approaches have significantly improved language models' performance on mathematical reasoning tasks, even when models are trained at a small scale. However, the specific capabilities enhanced through such fine-tuning remain poorly understood. In this paper, we conduct a detailed analysis of model performance on the AIME24 dataset to understand how reasoning capabilities evolve. We discover a ladder-like structure in problem difficulty, categorize questions into four tiers (Easy, Medium, Hard, and Extremely Hard (Exh)), and identify the specific requirements for advancing between tiers. We find that progression from Easy to Medium tier requires adopting an R1 reasoning style with minimal SFT (500-1K instances), while Hard-level questions suffer from frequent model's errors at each step of the reasoning chain, with accuracy plateauing at around 65% despite logarithmic scaling. Exh-level questions present a fundamentally different challenge; they require unconventional problem-solving skills that current models uniformly struggle with. Additional findings reveal that carefully curated small-scale datasets offer limited advantage-scaling dataset size proves far more effective. Our analysis provides a clearer roadmap for advancing language model capabilities in mathematical reasoning.

cs.AI↗
arXiv2026

Unknown Aware AI-Generated Content Attribution

The rapid advancement of photorealistic generative models has made it increasingly important to attribute the origin of synthetic content, moving beyond binary real or fake detection toward identifying the specific model that produced a given image. We study the problem of distinguishing outputs from a target generative model (e.g., OpenAI Dalle 3) from other sources, including real images and images generated by a wide range of alternative models. Using CLIP features and a simple linear classifier, shown to be effective in prior work, we establish a strong baseline for target generator attribution using only limited labeled data from the target model and a small number of known generators. However, this baseline struggles to generalize to harder, unseen, and newly released generators. To address this limitation, we propose a constrained optimization approach that leverages unlabeled wild data, consisting of images collected from the Internet that may include real images, outputs from unknown generators, or even samples from the target model itself. The proposed method encourages wild samples to be classified as non target while explicitly constraining performance on labeled data to remain high. Experimental results show that incorporating wild data substantially improves attribution performance on challenging unseen generators, demonstrating that unlabeled data from the wild can be effectively exploited to enhance AI generated content attribution in open world settings.

cs.LG↗
arXiv2025

How and Why LLMs Generalize: A Fine-Grained Analysis of LLM Reasoning from Cognitive Behaviors to Low-Level Patterns

Large Language Models (LLMs) display strikingly different generalization behaviors: supervised fine-tuning (SFT) often narrows capability, whereas reinforcement-learning (RL) tuning tends to preserve it. The reasons behind this divergence remain unclear, as prior studies have largely relied on coarse accuracy metrics. We address this gap by introducing a novel benchmark that decomposes reasoning into atomic core skills such as calculation, fact retrieval, simulation, enumeration, and diagnostic, providing a concrete framework for addressing the fundamental question of what constitutes reasoning in LLMs. By isolating and measuring these core skills, the benchmark offers a more granular view of how specific cognitive abilities emerge, transfer, and sometimes collapse during post-training. Combined with analyses of low-level statistical patterns such as distributional divergence and parameter statistics, it enables a fine-grained study of how generalization evolves under SFT and RL across mathematical, scientific reasoning, and non-reasoning tasks. Our meta-probing framework tracks model behavior at different training stages and reveals that RL-tuned models maintain more stable behavioral profiles and resist collapse in reasoning skills, whereas SFT models exhibit sharper drift and overfit to surface patterns. This work provides new insights into the nature of reasoning in LLMs and points toward principles for designing training strategies that foster broad, robust generalization.

cs.LG↗
arXiv2025

Feed Two Birds with One Scone: Exploiting Wild Data for Both Out-of-Distribution Generalization and Detection

Modern machine learning models deployed in the wild can encounter both covariate and semantic shifts, giving rise to the problems of out-of-distribution (OOD) generalization and OOD detection respectively. While both problems have received significant research attention lately, they have been pursued independently. This may not be surprising, since the two tasks have seemingly conflicting goals. This paper provides a new unified approach that is capable of simultaneously generalizing to covariate shifts while robustly detecting semantic shifts. We propose a margin-based learning framework that exploits freely available unlabeled data in the wild that captures the environmental test-time OOD distributions under both covariate and semantic shifts. We show both empirically and theoretically that the proposed margin constraint is the key to achieving both OOD generalization and detection. Extensive experiments show the superiority of our framework, outperforming competitive baselines that specialize in either OOD generalization or OOD detection. Code is publicly available at https://github.com/deeplearning-wisc/scone.

cs.LG↗
arXiv2025

Mitigating Recommendation Biases via Group-Alignment and Global-Uniformity in Representation Learning

Collaborative Filtering~(CF) plays a crucial role in modern recommender systems, leveraging historical user-item interactions to provide personalized suggestions. However, CF-based methods often encounter biases due to imbalances in training data. This phenomenon makes CF-based methods tend to prioritize recommending popular items and performing unsatisfactorily on inactive users. Existing works address this issue by rebalancing training samples, reranking recommendation results, or making the modeling process robust to the bias. Despite their effectiveness, these approaches can compromise accuracy or be sensitive to weighting strategies, making them challenging to train. In this paper, we deeply analyze the causes and effects of the biases and propose a framework to alleviate biases in recommendation from the perspective of representation distribution, namely Group-Alignment and Global-Uniformity Enhanced Representation Learning for Debiasing Recommendation (AURL). Specifically, we identify two significant problems in the representation distribution of users and items, namely group-discrepancy and global-collapse. These two problems directly lead to biases in the recommendation results. To this end, we propose two simple but effective regularizers in the representation space, respectively named group-alignment and global-uniformity. The goal of group-alignment is to bring the representation distribution of long-tail entities closer to that of popular entities, while global-uniformity aims to preserve the information of entities as much as possible by evenly distributing representations. Our method directly optimizes both the group-alignment and global-uniformity regularization terms to mitigate recommendation biases. Extensive experiments on three real datasets and various recommendation backbones verify the superiority of our proposed framework.

cs.IR↗
arXiv2025

Efficient Post-Training Refinement of Latent Reasoning in Large Language Models

Reasoning is a key component of language understanding in Large Language Models. While Chain-of-Thought prompting enhances performance via explicit intermediate steps, it suffers from sufficient token overhead and a fixed reasoning trajectory, preventing step-wise refinement. Recent advances in latent reasoning address these limitations by refining internal reasoning processes directly in the model's latent space, without producing explicit outputs. However, a key challenge remains: how to effectively update reasoning embeddings during post-training to guide the model toward more accurate solutions. To overcome this challenge, we propose a lightweight post-training framework that refines latent reasoning trajectories using two novel strategies: 1) Contrastive reasoning feedback, which compares reasoning embeddings against strong and weak baselines to infer effective update directions via embedding enhancement; 2) Residual embedding refinement, which stabilizes updates by progressively integrating current and historical gradients, enabling fast yet controlled convergence. Extensive experiments and case studies are conducted on five reasoning benchmarks to demonstrate the effectiveness of the proposed framework. Notably, a 5\% accuracy gain on MathQA without additional training.

cs.CL↗
arXiv2025

Understanding Embedding Scaling in Collaborative Filtering

Scaling recommendation models into large recommendation models has become one of the most widely discussed topics. Recent efforts focus on components beyond the scaling embedding dimension, as it is believed that scaling embedding may lead to performance degradation. Although there have been some initial observations on embedding, the root cause of their non-scalability remains unclear. Moreover, whether performance degradation occurs across different types of models and datasets is still an unexplored area. Regarding the effect of embedding dimensions on performance, we conduct large-scale experiments across 10 datasets with varying sparsity levels and scales, using 4 representative classical architectures. We surprisingly observe two novel phenomena: double-peak and logarithmic. For the former, as the embedding dimension increases, performance first improves, then declines, rises again, and eventually drops. For the latter, it exhibits a perfect logarithmic curve. Our contributions are threefold. First, we discover two novel phenomena when scaling collaborative filtering models. Second, we gain an understanding of the underlying causes of the double-peak phenomenon. Lastly, we theoretically analyze the noise robustness of collaborative filtering models, with results matching empirical observations.

cs.IR↗
arXiv2025

RL Grokking Recipe: How Does RL Unlock and Transfer New Algorithms in LLMs?

It remains an open question whether LLMs can acquire or generalize genuinely new reasoning strategies, beyond the sharpened skills encoded in their parameters during pre-training or post-training. To attempt to answer this debate, we introduce DELTA-Code -- Distributional Evaluation of Learnability and Transferrability in Algorithmic Coding -- a controlled benchmark of synthetic coding problem families designed to probe two fundamental aspects: learnability -- can LLMs, through reinforcement learning (RL), solve problem families where pretrained models exhibit failure with large enough attempts (pass@K=0)? -- and transferrability -- if learnability happens, can such skills transfer systematically to out-of-distribution (OOD) test sets? Unlike prior public coding datasets, DELTA isolates reasoning skills through templated problem generators and introduces fully OOD problem families that demand novel strategies rather than tool invocation or memorized patterns. Our experiments reveal a striking grokking phase transition: after an extended period with near-zero reward, RL-trained models abruptly climb to near-perfect accuracy. To enable learnability on previously unsolvable problem families, we explore key training ingredients such as staged warm-up with dense rewards, experience replay, curriculum training, and verification-in-the-loop. Beyond learnability, we use DELTA to evaluate transferability or generalization along exploratory, compositional, and transformative axes, as well as cross-family transfer. Results show solid gains within families and for recomposed skills, but persistent weaknesses in transformative cases. DELTA thus offers a clean testbed for probing the limits of RL-driven reasoning and for understanding how models can move beyond existing priors to acquire new algorithmic skills.

cs.LG↗
arXiv2025

DELTA: Variational Disentangled Learning for Privacy-Preserving Data Reprogramming

In real-world applications, domain data often contains identifiable or sensitive attributes, is subject to strict regulations (e.g., HIPAA, GDPR), and requires explicit data feature engineering for interpretability and transparency. Existing feature engineering primarily focuses on advancing downstream task performance, often risking privacy leakage. We generalize this learning task under such new requirements as Privacy-Preserving Data Reprogramming (PPDR): given a dataset, transforming features to maximize target attribute prediction accuracy while minimizing sensitive attribute prediction accuracy. PPDR poses challenges for existing systems: 1) generating high-utility feature transformations without being overwhelmed by a large search space, and 2) disentangling and eliminating sensitive information from utility-oriented features to reduce privacy inferability. To tackle these challenges, we propose DELTA, a two-phase variational disentangled generative learning framework. Phase I uses policy-guided reinforcement learning to discover feature transformations with downstream task utility, without any regard to privacy inferability. Phase II employs a variational LSTM seq2seq encoder-decoder with a utility-privacy disentangled latent space design and adversarial-causal disentanglement regularization to suppress privacy signals during feature generation. Experiments on eight datasets show DELTA improves predictive performance by ~9.3% and reduces privacy leakage by ~35%, demonstrating robust, privacy-aware data transformation.

cs.LG↗
arXiv2025

Data-Efficient Symbolic Regression via Foundation Model Distillation

Discovering interpretable mathematical equations from observed data (a.k.a. equation discovery or symbolic regression) is a cornerstone of scientific discovery, enabling transparent modeling of physical, biological, and economic systems. While foundation models pre-trained on large-scale equation datasets offer a promising starting point, they often suffer from negative transfer and poor generalization when applied to small, domain-specific datasets. In this paper, we introduce EQUATE (Equation Generation via QUality-Aligned Transfer Embeddings), a data-efficient fine-tuning framework that adapts foundation models for symbolic equation discovery in low-data regimes via distillation. EQUATE combines symbolic-numeric alignment with evaluator-guided embedding optimization, enabling a principled embedding-search-generation paradigm. Our approach reformulates discrete equation search as a continuous optimization task in a shared embedding space, guided by data-equation fitness and simplicity. Experiments across three standard public benchmarks (Feynman, Strogatz, and black-box datasets) demonstrate that EQUATE consistently outperforms state-of-the-art baselines in both accuracy and robustness, while preserving low complexity and fast inference. These results highlight EQUATE as a practical and generalizable solution for data-efficient symbolic regression in foundation model distillation settings.

cs.LG↗
arXiv2025

Where's the liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content

The recent proliferation of photorealistic images created by generative models has sparked both excitement and concern, as these images are increasingly indistinguishable from real ones to the human eye. While offering new creative and commercial possibilities, the potential for misuse, such as in misinformation and fraud, highlights the need for effective detection methods. Current detection approaches often rely on access to model weights or require extensive collections of real image datasets, limiting their scalability and practical application in real world scenarios. In this work, we introduce a novel black box detection framework that requires only API access, sidestepping the need for model weights or large auxiliary datasets. Our approach leverages a corrupt and recover strategy: by masking part of an image and assessing the model ability to reconstruct it, we measure the likelihood that the image was generated by the model itself. For black-box models that do not support masked image inputs, we incorporate a cost efficient surrogate model trained to align with the target model distribution, enhancing detection capability. Our framework demonstrates strong performance, outperforming baseline methods by 4.31% in mean average precision across eight diffusion model variant datasets.

cs.LG↗
arXiv2025

Causal Graph Profiling via Structural Divergence for Robust Anomaly Detection in Cyber-Physical Systems

With the growing complexity of cyberattacks targeting critical infrastructures such as water treatment networks, there is a pressing need for robust anomaly detection strategies that account for both system vulnerabilities and evolving attack patterns. Traditional methods -- statistical, density-based, and graph-based models struggle with distribution shifts and class imbalance in multivariate time series, often leading to high false positive rates. To address these challenges, we propose CGAD, a Causal Graph-based Anomaly Detection framework designed for reliable cyberattack detection in public infrastructure systems. CGAD follows a two-phase supervised framework -- causal profiling and anomaly scoring. First, it learns causal invariant graph structures representing the system's behavior under "Normal" and "Attack" states using Dynamic Bayesian Networks. Second, it employs structural divergence to detect anomalies via causal graph comparison by evaluating topological deviations in causal graphs over time. By leveraging causal structures, CGAD achieves superior adaptability and accuracy in non-stationary and imbalanced time series environments compared to conventional machine learning approaches. By uncovering causal structures beneath volatile sensor data, our framework not only detects cyberattacks with markedly higher precision but also redefines robustness in anomaly detection, proving resilience where traditional models falter under imbalance and drift. Our framework achieves substantial gains in F1 and ROC-AUC scores over best-performing baselines across four industrial datasets, demonstrating robust detection of delayed and structurally complex anomalies.

cs.LG↗
arXiv2025

Incremental Causal Graph Learning for Online Cyberattack Detection in Cyber-Physical Infrastructures

The escalating threat of cyberattacks on real-time critical infrastructures poses serious risks to public safety, demanding detection methods that effectively capture complex system interdependencies and adapt to evolving attack patterns. Traditional real-time anomaly detection techniques often suffer from excessive false positives due to their statistical sensitivity to high data variance and class imbalance. To address these limitations, recent research has explored modeling causal relationships among system components. However, prior work mainly focuses on offline causal graph-based approaches that require static historical data and fail to generalize to real-time settings. These methods are fundamentally constrained by: (1) their inability to adapt to dynamic shifts in data distribution without retraining, and (2) the risk of catastrophic forgetting when lacking timely supervision in live systems. To overcome these challenges, we propose INCADET, a novel framework for incremental causal graph learning tailored to real-time cyberattack detection. INCADET dynamically captures evolving system behavior by incrementally updating causal graphs across streaming time windows. The framework comprises three modules: 1) Early Symptom Detection: Detects transitions in system status using divergence in edge-weight distributions across sequential causal graphs. 2) Incremental Causal Graph Learning: Leverages experience replay and edge reinforcement to continually refine causal structures while preserving prior knowledge. 3) Causal Graph Classification: Employs Graph Convolutional Networks (GCNs) to classify system status using the learned causal graphs. Extensive experiments on real-world critical infrastructure datasets demonstrate that INCADET achieves superior accuracy, robustness, and adaptability compared to both static causal and deep temporal baselines in evolving attack scenarios.

cs.LG↗