arXiv subjects

Chen Ma

Publications and source records attributed to Chen Ma.

At least 73 records · Page 4Linked to original sources

arXiv2025

Spatial CAPTCHA: Generatively Benchmarking Spatial Reasoning for Human-Machine Differentiation

Online services rely on CAPTCHAs as a first line of defense against automated abuse, yet recent advances in multi-modal large language models (MLLMs) have eroded the effectiveness of conventional designs that focus on text recognition or 2D image understanding. To address this challenge, we present Spatial CAPTCHA, a novel human-verification framework that leverages fundamental differences in spatial reasoning between humans and MLLMs. Unlike existing CAPTCHAs which rely on low-level perception tasks that are vulnerable to modern AI, Spatial CAPTCHA generates dynamic questions requiring geometric reasoning, perspective-taking, occlusion handling, and mental rotation. These skills are intuitive for humans but difficult for state-of-the-art (SOTA) AI systems. The system employs a procedural generation pipeline with constraint-based difficulty control, automated correctness verification, and human-in-the-loop validation to ensure scalability, robustness, and adaptability. Evaluation on a corresponding benchmark, Spatial-CAPTCHA-Bench, demonstrates that humans vastly outperform 10 state-of-the-art MLLMs, with the best model achieving only 31.0% Pass@1 accuracy. Furthermore, we compare Spatial CAPTCHA with Google reCAPTCHA, which confirms its effectiveness as both a security mechanism and a diagnostic tool for spatial reasoning in AI.

cs.AI↗

arXiv2025

NILE: Internal Consistency Alignment in Large Language Models

As a crucial step to enhance LLMs alignment with human intentions, Instruction Fine-Tuning (IFT) has a high demand on dataset quality. However, existing IFT datasets often contain knowledge that is inconsistent with LLMs' internal knowledge learned from the pre-training phase, which can greatly affect the efficacy of IFT. To address this issue, we introduce NILE (iNternal consIstency aLignmEnt) framework, aimed at optimizing IFT datasets to unlock LLMs' capability further. NILE operates by eliciting target pre-trained LLM's internal knowledge corresponding to instruction data. The internal knowledge is leveraged to revise the answer in IFT datasets. Additionally, we propose a novel Internal Consistency Filtering (ICF) method to filter training samples, ensuring its high consistency with LLM's internal knowledge. Our experiments demonstrate that NILE-aligned IFT datasets sharply boost LLM performance across multiple LLM ability evaluation datasets, achieving up to 66.6% gain on Arena-Hard and 68.5% on Alpaca-Eval V2. Further analysis confirms that each component of the NILE}framework contributes to these substantial performance improvements, and provides compelling evidence that dataset consistency with pre-trained internal knowledge is pivotal for maximizing LLM potential.

cs.CL↗

arXiv2025

Scaling Up, Speeding Up: A Benchmark of Speculative Decoding for Efficient LLM Test-Time Scaling

Test-time scaling has emerged as a powerful paradigm for enhancing the reasoning capabilities of large language models (LLMs) by allocating additional computational resources during inference. However, this paradigm is inherently inefficient due to the generation of redundant and repetitive reasoning traces, leading to significant computational overhead. Speculative decoding offers a promising avenue for mitigating this inefficiency, yet its efficacy in the structured, repetition-rich context of test-time scaling remains largely unexplored. To bridge this gap, we introduce the first comprehensive benchmark designed to evaluate speculative decoding methods for accelerating LLM test-time scaling. Our benchmark provides consistent experimental protocols across representative test-time scaling paradigms (e.g., Best-of-N sampling and multi-round thinking), enabling a fair comparison of three major categories of speculative decoding: model-based, training-based, and n-gram-based methods. Extensive experiments reveal that simple n-gram-based methods effectively capture repetitive patterns, demonstrating unique potential in accelerating test-time scaling. This phenomenon demonstrates the value of integrating n-gram-based methods with model-based or training-based approaches to balance acceleration for both repetitive and diverse reasoning in test-time scaling. We hope this benchmark spurs further research on speculative decoding for test-time scaling, enabling faster and more practical reasoning in LLMs through better handling of repetitive and diverse reasoning paths.

cs.CL↗

arXiv2025

Average-DICE: Stationary Distribution Correction by Regression

Off-policy policy evaluation (OPE), an essential component of reinforcement learning, has long suffered from stationary state distribution mismatch, undermining both stability and accuracy of OPE estimates. While existing methods correct distribution shifts by estimating density ratios, they often rely on expensive optimization or backward Bellman-based updates and struggle to outperform simpler baselines. We introduce AVG-DICE, a computationally simple Monte Carlo estimator for the density ratio that averages discounted importance sampling ratios, providing an unbiased and consistent correction. AVG-DICE extends naturally to nonlinear function approximation using regression, which we roughly tune and test on OPE tasks based on Mujoco Gym environments and compare with state-of-the-art density-ratio estimators using their reported hyperparameters. In our experiments, AVG-DICE is at least as accurate as state-of-the-art estimators and sometimes offers orders-of-magnitude improvements. However, a sensitivity analysis shows that best-performing hyperparameters may vary substantially across different discount factors, so a re-tuning is suggested.

cs.LG↗

arXiv2025

Revisiting Adversarial Patch Defenses on Object Detectors: Unified Evaluation, Large-Scale Dataset, and New Insights

Developing reliable defenses against patch attacks on object detectors has attracted increasing interest. However, we identify that existing defense evaluations lack a unified and comprehensive framework, resulting in inconsistent and incomplete assessments of current methods. To address this issue, we revisit 11 representative defenses and present the first patch defense benchmark, involving 2 attack goals, 13 patch attacks, 11 object detectors, and 4 diverse metrics. This leads to the large-scale adversarial patch dataset with 94 types of patches and 94,000 images. Our comprehensive analyses reveal new insights: (1) The difficulty in defending against naturalistic patches lies in the data distribution, rather than the commonly believed high frequencies. Our new dataset with diverse patch distributions can be used to improve existing defenses by 15.09% AP@0.5. (2) The average precision of the attacked object, rather than the commonly pursued patch detection accuracy, shows high consistency with defense performance. (3) Adaptive attacks can substantially bypass existing defenses, and defenses with complex/stochastic models or universal patch properties are relatively robust. We hope that our analyses will serve as guidance on properly evaluating patch attacks/defenses and advancing their design. Code and dataset are available at https://github.com/Gandolfczjh/APDE, where we will keep integrating new attacks/defenses.

cs.CV↗

arXiv2025

DOGR: Towards Versatile Visual Document Grounding and Referring

With recent advances in Multimodal Large Language Models (MLLMs), grounding and referring capabilities have gained increasing attention for achieving detailed understanding and flexible user interaction. However, these capabilities still remain underdeveloped in visual document understanding due to the scarcity of fine-grained datasets and comprehensive benchmarks. To fill this gap, we propose the DOcument Grounding and Referring data engine (DOGR-Engine), which generates two types of high-quality fine-grained document data: (1) multi-granular parsing data to improve text localization and recognition, and (2) instruction-tuning data to activate MLLMs' grounding and referring capabilities in dialogue and reasoning. Using the DOGR-Engine, we construct DOGR-Bench, a benchmark covering seven grounding and referring tasks across three document types (chart, poster, and PDF document), offering a comprehensive evaluation of fine-grained document understanding. Leveraging the generated data, we further develop DOGR, a strong baseline model that excels in text localization and recognition, while precisely grounds and refers to key textual information during conversation and reasoning, thereby advancing document understanding to a finer granularity and enable flexible interaction paradigms.

cs.CV↗

arXiv2025

Boosting Ray Search Procedure of Hard-label Attacks with Transfer-based Priors

One of the most practical and challenging types of black-box adversarial attacks is the hard-label attack, where only the top-1 predicted label is available. One effective approach is to search for the optimal ray direction from the benign image that minimizes the $\ell_p$-norm distance to the adversarial region. The unique advantage of this approach is that it transforms the hard-label attack into a continuous optimization problem. The objective function value is the ray's radius, which can be obtained via binary search at a high query cost. Existing methods use a "sign trick" in gradient estimation to reduce the number of queries. In this paper, we theoretically analyze the quality of this gradient estimation and propose a novel prior-guided approach to improve ray search efficiency both theoretically and empirically. Specifically, we utilize the transfer-based priors from surrogate models, and our gradient estimators appropriately integrate them by approximating the projection of the true gradient onto the subspace spanned by these priors and random directions, in a query-efficient manner. We theoretically derive the expected cosine similarities between the obtained gradient estimators and the true gradient, and demonstrate the improvement achieved by incorporating priors. Extensive experiments on the ImageNet and CIFAR-10 datasets show that our approach significantly outperforms 11 state-of-the-art methods in terms of query efficiency.

cs.CV↗

arXiv2025

MemBench: Towards More Comprehensive Evaluation on the Memory of LLM-based Agents

Recent works have highlighted the significance of memory mechanisms in LLM-based agents, which enable them to store observed information and adapt to dynamic environments. However, evaluating their memory capabilities still remains challenges. Previous evaluations are commonly limited by the diversity of memory levels and interactive scenarios. They also lack comprehensive metrics to reflect the memory capabilities from multiple aspects. To address these problems, in this paper, we construct a more comprehensive dataset and benchmark to evaluate the memory capability of LLM-based agents. Our dataset incorporates factual memory and reflective memory as different levels, and proposes participation and observation as various interactive scenarios. Based on our dataset, we present a benchmark, named MemBench, to evaluate the memory capability of LLM-based agents from multiple aspects, including their effectiveness, efficiency, and capacity. To benefit the research community, we release our dataset and project at https://github.com/import-myself/Membench.

cs.CL↗

arXiv2025

Counterfactual Multi-player Bandits for Explainable Recommendation Diversification

Existing recommender systems tend to prioritize items closely aligned with users' historical interactions, inevitably trapping users in the dilemma of ``filter bubble''. Recent efforts are dedicated to improving the diversity of recommendations. However, they mainly suffer from two major issues: 1) a lack of explainability, making it difficult for the system designers to understand how diverse recommendations are generated, and 2) limitations to specific metrics, with difficulty in enhancing non-differentiable diversity metrics. To this end, we propose a \textbf{C}ounterfactual \textbf{M}ulti-player \textbf{B}andits (CMB) method to deliver explainable recommendation diversification across a wide range of diversity metrics. Leveraging a counterfactual framework, our method identifies the factors influencing diversity outcomes. Meanwhile, we adopt the multi-player bandits to optimize the counterfactual optimization objective, making it adaptable to both differentiable and non-differentiable diversity metrics. Extensive experiments conducted on three real-world datasets demonstrate the applicability, effectiveness, and explainability of the proposed CMB.

cs.IR↗

arXiv2025

Defurnishing with X-Ray Vision: Joint Removal of Furniture from Panoramas and Mesh

We present a pipeline for generating defurnished replicas of indoor spaces represented as textured meshes and corresponding multi-view panoramic images. To achieve this, we first segment and remove furniture from the mesh representation, extend planes, and fill holes, obtaining a simplified defurnished mesh (SDM). This SDM acts as an ``X-ray'' of the scene's underlying structure, guiding the defurnishing process. We extract Canny edges from depth and normal images rendered from the SDM. We then use these as a guide to remove the furniture from panorama images via ControlNet inpainting. This control signal ensures the availability of global geometric information that may be hidden from a particular panoramic view by the furniture being removed. The inpainted panoramas are used to texture the mesh. We show that our approach produces higher quality assets than methods that rely on neural radiance fields, which tend to produce blurry low-resolution images, or RGB-D inpainting, which is highly susceptible to hallucinations.

cs.CV↗

arXiv2025

Learning Binarized Representations with Pseudo-positive Sample Enhancement for Efficient Graph Collaborative Filtering

Learning vectorized embeddings is fundamental to many recommender systems for user-item matching. To enable efficient online inference, representation binarization, which embeds latent features into compact binary sequences, has recently shown significant promise in optimizing both memory usage and computational overhead. However, existing approaches primarily focus on numerical quantization, neglecting the associated information loss, which often results in noticeable performance degradation. To address these issues, we study the problem of graph representation binarization for efficient collaborative filtering. Our findings indicate that explicitly mitigating information loss at various stages of embedding binarization has a significant positive impact on performance. Building on these insights, we propose an enhanced framework, BiGeaR++, which specifically leverages supervisory signals from pseudo-positive samples, incorporating both real item data and latent embedding samples. Compared to its predecessor BiGeaR, BiGeaR++ introduces a fine-grained inference distillation mechanism and an effective embedding sample synthesis approach. Empirical evaluations across five real-world datasets demonstrate that the new designs in BiGeaR++ work seamlessly well with other modules, delivering substantial improvements of around 1%-10% over BiGeaR and thus achieving state-of-the-art performance compared to the competing methods. Our implementation is available at https://github.com/QueYork/BiGeaR-SS.

cs.IR↗

arXiv2025

Shapley Value-driven Data Pruning for Recommender Systems

Recommender systems often suffer from noisy interactions like accidental clicks or popularity bias. Existing denoising methods typically identify users' intent in their interactions, and filter out noisy interactions that deviate from the assumed intent. However, they ignore that interactions deemed noisy could still aid model training, while some ``clean'' interactions offer little learning value. To bridge this gap, we propose Shapley Value-driven Valuation (SVV), a framework that evaluates interactions based on their objective impact on model training rather than subjective intent assumptions. In SVV, a real-time Shapley value estimation method is devised to quantify each interaction's value based on its contribution to reducing training loss. Afterward, SVV highlights the interactions with high values while downplaying low ones to achieve effective data pruning for recommender systems. In addition, we develop a simulated noise protocol to examine the performance of various denoising approaches systematically. Experiments on four real-world datasets show that SVV outperforms existing denoising methods in both accuracy and robustness. Further analysis also demonstrates that our SVV can preserve training-critical interactions and offer interpretable noise assessment. This work shifts denoising from heuristic filtering to principled, model-driven interaction valuation.

cs.IR↗

arXiv2025

A Survey on Side Information-driven Session-based Recommendation: From a Data-centric Perspective

Session-based recommendation is gaining increasing attention due to its practical value in predicting the intents of anonymous users based on limited behaviors. Emerging efforts incorporate various side information to alleviate inherent data scarcity issues in this task, leading to impressive performance improvements. The core of side information-driven session-based recommendation is the discovery and utilization of diverse data. In this survey, we provide a comprehensive review of this task from a data-centric perspective. Specifically, this survey commences with a clear formulation of the task. This is followed by a detailed exploration of various benchmarks rich in side information that are pivotal for advancing research in this field. Afterwards, we delve into how different types of side information enhance the task, underscoring data characteristics and utility. Moreover, we discuss the usage of various side information, including data encoding, data injection, and involved techniques. A systematic review of research progress is then presented, with the taxonomy by the types of side information. Finally, we summarize the current limitations and present the future prospects of this vibrant topic.

cs.IR↗

arXiv2025

Hyperbolic Contrastive Learning with Model-augmentation for Knowledge-aware Recommendation

Benefiting from the effectiveness of graph neural networks (GNNs) and contrastive learning, GNN-based contrastive learning has become mainstream for knowledge-aware recommendation. However, most existing contrastive learning-based methods have difficulties in effectively capturing the underlying hierarchical structure within user-item bipartite graphs and knowledge graphs. Moreover, they commonly generate positive samples for contrastive learning by perturbing the graph structure, which may lead to a shift in user preference learning. To overcome these limitations, we propose hyperbolic contrastive learning with model-augmentation for knowledge-aware recommendation. To capture the intrinsic hierarchical graph structures, we first design a novel Lorentzian knowledge aggregation mechanism, which enables more effective representations of users and items. Then, we propose three model-level augmentation techniques to assist Hyperbolic contrastive learning. Different from the classical structure-level augmentation (e.g., edge dropping), the proposed model-augmentations can avoid preference shifts between the augmented positive pair. Finally, we conduct extensive experiments to demonstrate the superiority (maximum improvement of $11.03\%$) of proposed methods over existing baselines.

cs.IR↗

arXiv2025

A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?

As enthusiasm for scaling computation (data and parameters) in the pretraining era gradually diminished, test-time scaling (TTS), also referred to as ``test-time computing'' has emerged as a prominent research focus. Recent studies demonstrate that TTS can further elicit the problem-solving capabilities of large language models (LLMs), enabling significant breakthroughs not only in specialized reasoning tasks, such as mathematics and coding, but also in general tasks like open-ended Q&A. However, despite the explosion of recent efforts in this area, there remains an urgent need for a comprehensive survey offering a systemic understanding. To fill this gap, we propose a unified, multidimensional framework structured along four core dimensions of TTS research: what to scale, how to scale, where to scale, and how well to scale. Building upon this taxonomy, we conduct an extensive review of methods, application scenarios, and assessment aspects, and present an organized decomposition that highlights the unique functional roles of individual techniques within the broader TTS landscape. From this analysis, we distill the major developmental trajectories of TTS to date and offer hands-on guidelines for practical deployment. Furthermore, we identify several open challenges and offer insights into promising future directions, including further scaling, clarifying the functional essence of techniques, generalizing to more tasks, and more attributions. Our repository is available on https://github.com/testtimescaling/testtimescaling.github.io/

cs.CL↗

arXiv2025

Comprehending Knowledge Graphs with Large Language Models for Recommender Systems

In recent years, the introduction of knowledge graphs (KGs) has significantly advanced recommender systems by facilitating the discovery of potential associations between items. However, existing methods still face several limitations. First, most KGs suffer from missing facts or limited scopes. Second, existing methods convert textual information in KGs into IDs, resulting in the loss of natural semantic connections between different items. Third, existing methods struggle to capture high-order connections in the global KG. To address these limitations, we propose a novel method called CoLaKG, which leverages large language models (LLMs) to improve KG-based recommendations. The extensive knowledge and remarkable reasoning capabilities of LLMs enable our method to supplement missing facts in KGs, and their powerful text understanding abilities allow for better utilization of semantic information. Specifically, CoLaKG extracts useful information from KGs at both local and global levels. By employing the item-centered subgraph extraction and prompt engineering, it can accurately understand the local information. In addition, through the semantic-based retrieval module, each item is enriched by related items from the entire knowledge graph, effectively harnessing global information. Furthermore, the local and global information are effectively integrated into the recommendation model through a representation fusion module and a retrieval-augmented representation learning module, respectively. Extensive experiments on four real-world datasets demonstrate the superiority of our method.

cs.IR↗

arXiv2025

RevisEval: Improving LLM-as-a-Judge via Response-Adapted References

With significant efforts in recent studies, LLM-as-a-Judge has become a cost-effective alternative to human evaluation for assessing text generation quality in a wide range of tasks. However, there still remains a reliability gap between LLM-as-a-Judge and human evaluation. One important reason is the lack of guided oracles in the evaluation process. Motivated by the role of reference pervasively used in classic text evaluation, we introduce RevisEval, a novel text generation evaluation paradigm via the response-adapted references. RevisEval is driven by the key observation that an ideal reference should maintain the necessary relevance to the response to be evaluated. Specifically, RevisEval leverages the text revision capabilities of large language models (LLMs) to adaptively revise the response, then treat the revised text as the reference (response-adapted reference) for the subsequent evaluation. Extensive experiments demonstrate that RevisEval outperforms traditional reference-free and reference-based evaluation paradigms that use LLM-as-a-Judge across NLG tasks and open-ended instruction-following tasks. More importantly, our response-adapted references can further boost the classical text metrics, e.g., BLEU and BERTScore, compared to traditional references and even rival the LLM-as-a-Judge. A detailed analysis is also conducted to confirm RevisEval's effectiveness in bias reduction, the impact of inference cost, and reference relevance.

cs.CL↗

arXiv2025

Crowd Comparative Reasoning: Unlocking Comprehensive Evaluations for LLM-as-a-Judge

LLM-as-a-Judge, which generates chain-of-thought (CoT) judgments, has become a widely adopted auto-evaluation method. However, its reliability is compromised by the CoT reasoning's inability to capture comprehensive and deeper details, often leading to incomplete outcomes. Existing methods mainly rely on majority voting or criteria expansion, which is insufficient to address the limitation in CoT. We propose Crowd-based Comparative Evaluation, which introduces additional crowd responses to compare with the candidate responses, thereby exposing deeper and more comprehensive details within the candidate responses. This process effectively guides LLM-as-a-Judge to provide a more detailed CoT judgment. Extensive experiments demonstrate that our approach enhances evaluation reliability, achieving an average accuracy gain of 6.7% across five benchmarks. Moreover, our method produces higher-quality CoTs that facilitate judge distillation and exhibit superior performance in rejection sampling for supervised fine-tuning (SFT), referred to as crowd rejection sampling, thereby enabling more efficient SFT. Our analysis confirms that CoTs generated by ours are more comprehensive and of higher quality, and evaluation accuracy improves as inference scales.

cs.CL↗