arXiv subjects

Cheng Yu

Publications and source records attributed to Cheng Yu.

At least 37 records · Page 2Linked to original sources

arXiv2026

How Should AI Safety Benchmarks Benchmark Safety?

AI safety benchmarks are pivotal for safety in advanced AI systems; however, they have significant technical, epistemic, and sociotechnical shortcomings. We present a review of 210 safety benchmarks that maps out common challenges in safety benchmarking, documenting failures and limitations by drawing from engineering sciences and long-established theories of risk and safety. We argue that adhering to established risk management principles, mapping the space of what can(not) be measured, developing robust probabilistic metrics, and efficiently deploying measurement theory to connect benchmarking objectives with the world can significantly improve the validity and usefulness of AI safety benchmarks. The review provides a roadmap on how to improve AI safety benchmarking, and we illustrate the effectiveness of these recommendations through quantitative and qualitative evaluation. We also provide workflow-oriented guiding questions with illustrative benchmark that help researchers and practitioners develop robust and epistemologically sound safety benchmarks. This study advances the science of benchmarking and helps practitioners deploy AI systems more responsibly.

cs.CY↗

arXiv2026

Universality in the Low Mach number limit via a convex integration framework

We study the low Mach number limit of the compressible Euler equations through the lens of convex integration. For any prescribed $L^2$ weak solution of the incompressible Euler equations, we construct a corresponding family of weak solutions to the compressible Euler equations via a refined convex integration scheme. We then prove that, as the Mach number tends to zero, this family of solutions converges strongly to the given incompressible solution. This result demonstrates that the incompressible system acts as a universal attractor in this setting: every incompressible flow can be realized as the limit of convex integration solutions to the compressible system. Our approach highlights a new form of universality for singular limits and provides a rigorous framework for understanding the incompressible limit from the perspective of weak solution theory.

math.AP↗

arXiv2026

The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models

Diffusion Large Language Models (dLLMs) break the rigid left-to-right constraint of traditional LLMs, enabling token generation in arbitrary orders. Intuitively, this flexibility implies a solution space that strictly supersets the fixed autoregressive trajectory, theoretically unlocking superior reasoning potential. However, in this paper, we find that for general reasoning tasks (e.g., mathematics and coding), arbitrary order generation may in fact limit the reasoning potential of dLLMs. We observe that dLLMs tend to exploit this order flexibility to bypass high-uncertainty tokens that are crucial for exploration, which can lead to a premature collapse of solution coverage. This observation motivates a rethink of RL approaches for dLLMs, where considerable complexities, such as handling combinatorial trajectories and intractable likelihoods, are often devoted to preserving this flexibility. We show that effective reasoning can be elicited by simply forgoing arbitrary order and applying standard Group Relative Policy Optimization (GRPO) instead. Our approach, JustGRPO, is minimalist yet surprisingly effective (e.g., 89.1% accuracy on GSM8K) while fully retaining the parallel decoding ability of dLLMs. Project page: https://nzl-thu.github.io/the-flexibility-trap

cs.CL↗

arXiv2026

Unified Thinker: A General Reasoning Modular Core for Image Generation

Despite impressive progress in high-fidelity image synthesis, generative models still struggle with logic-intensive instruction following, exposing a persistent reasoning--execution gap. Meanwhile, closed-source systems (e.g., Nano Banana) have demonstrated strong reasoning-driven image generation, highlighting a substantial gap to current open-source models. We argue that closing this gap requires not merely better visual generators, but executable reasoning: decomposing high-level intents into grounded, verifiable plans that directly steer the generative process. To this end, we propose Unified Thinker, a task-agnostic reasoning architecture for general image generation, designed as a unified planning core that can plug into diverse generators and workflows. Unified Thinker decouples a dedicated Thinker from the image Generator, enabling modular upgrades of reasoning without retraining the entire generative model. We further introduce a two-stage training paradigm: we first build a structured planning interface for the Thinker, then apply reinforcement learning to ground its policy in pixel-level feedback, encouraging plans that optimize visual correctness over textual plausibility. Extensive experiments on text-to-image generation and image editing show that Unified Thinker substantially improves image reasoning and generation quality.

cs.CV↗

arXiv2025

Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem

Agentic crafting requires LLMs to operate in real-world environments over multiple turns by taking actions, observing outcomes, and iteratively refining artifacts. Despite its importance, the open-source community lacks a principled, end-to-end ecosystem to streamline agent development. We introduce the Agentic Learning Ecosystem (ALE), a foundational infrastructure that optimizes the production pipeline for agentic model. ALE consists of three components: ROLL, a post-training framework for weight optimization; ROCK, a sandbox environment manager for trajectory generation; and iFlow CLI, an agent framework for efficient context engineering. We release ROME, an open-source agent grounded by ALE and trained on over one million trajectories. Our approach includes data composition protocols for synthesizing complex behaviors and a novel policy optimization algorithm, Interaction-Perceptive Agentic Policy Optimization (IPA), which assigns credit over semantic interaction chunks rather than individual tokens to improve long-horizon training stability. Empirically, we evaluate ROME within a structured setting and introduce Terminal Bench Pro, a benchmark with improved scale and contamination control. ROME demonstrates strong performance across benchmarks like SWE-bench Verified and Terminal Bench, proving the effectiveness of ALE.

cs.AI↗

arXiv2025

Theoretical and experimental studies of energy modulation to demodulation in seeded free-electron lasers

Laser manipulation plays a critical role in precisely tailoring relativistic electron beams through energy modulation, enabling the generation of coherent, intense, and ultrashort radiation in accelerator-based light sources such as synchrotron radiation facilities and free-electron lasers (FELs). However, laser-induced energy modulation inevitably degrades electron beam quality by increasing the energy spread, thereby limiting high-repetition-rate operation. Here, we investigate energy modulation and demodulation in a seeded FEL using two modulators separated by a tunable phase shifter. Analytical analysis and three-dimensional simulations show that a $\pi$ phase delay can nearly reverse the laser-beam interaction and substantially suppress the residual modulation. Diagnostics based on coherent undulator radiation and time-resolved measurements are established to characterize weak residual modulation, and a dedicated demodulation undulator is designed for controlled studies. Preliminary experiments performed at the Shanghai soft X-ray FEL facility using the existing seeding beamline demonstrate laser-induced energy-modulation suppression. Together with the analytical and numerical studies, these results establish a practical framework for investigating the transition from energy modulation to demodulation in seeded FELs, with potential applications in high-repetition-rate, fully coherent X-ray sources with improved preservation of electron beam quality.

physics.acc-ph↗

arXiv2025

Understanding Diffusion Models via Code Execution

Diffusion models have achieved remarkable performance in generative modeling, yet their theoretical foundations are often intricate, and the gap between mathematical formulations in papers and practical open-source implementations can be difficult to bridge. Existing tutorials primarily focus on deriving equations, offering limited guidance on how diffusion models actually operate in code. To address this, we present a concise implementation of approximately 300 lines that explains diffusion models from a code-execution perspective. Our minimal example preserves the essential components -- including forward diffusion, reverse sampling, the noise-prediction network, and the training loop -- while removing unnecessary engineering details. This technical report aims to provide researchers with a clear, implementation-first understanding of how diffusion models work in practice and how code and theory correspond. Our code and pre-trained models are available at: https://github.com/disanda/GM/tree/main/DDPM-DDIM-ClassifierFree.

cs.CV↗

arXiv2025

Language Model Based Text-to-Audio Generation: Anti-Causally Aligned Collaborative Residual Transformers

While language models (LMs) paired with residual vector quantization (RVQ) tokenizers have shown promise in text-to-audio (T2A) generation, they still lag behind diffusion-based models by a non-trivial margin. We identify a critical dilemma underpinning this gap: incorporating more RVQ layers improves audio reconstruction fidelity but exceeds the generation capacity of conventional LMs. To address this, we first analyze RVQ dynamics and uncover two key limitations: 1) orthogonality of features across RVQ layers hinders effective LMs training, and 2) descending semantic richness in tokens from deeper RVQ layers exacerbates exposure bias during autoregressive decoding. Based on these insights, we propose Siren, a novel LM-based framework that employs multiple isolated transformers with causal conditioning and anti-causal alignment via reinforcement learning. Extensive experiments demonstrate that Siren outperforms both existing LM-based and diffusion-based T2A systems, achieving state-of-the-art results. By bridging the representational strengths of LMs with the fidelity demands of audio synthesis, our approach repositions LMs as competitive contenders against diffusion models in T2A tasks. Moreover, by aligning audio representations with linguistic structures, Siren facilitates a promising pathway toward unified multi-modal generation frameworks.

cs.SD↗

arXiv2025

Evaluating and Improving Large Language Models for Competitive Program Generation

Context: Due to the demand for strong algorithmic reasoning, complex logic implementation, and strict adherence to input/output formats and resource constraints, competitive programming generation by large language models (LLMs) is considered the most challenging problem in current LLM-based code generation. However, previous studies often evaluate LLMs using simple prompts and benchmark datasets prone to data leakage. Moreover, prior work has limited consideration of the diversity in algorithm types and difficulty levels. Objective: In this study, we aim to evaluate and improve LLMs in solving real-world competitive programming problems. Methods: We initially collect 117 problems from nine regional ICPC/CCPC contests held in 2024 and design four filtering criteria to construct a curated benchmark consisting of 80 problems. Leveraging DeepSeek-R1 as the LLM, we evaluate its competitive program generation capabilities through the online judge (OJ) platforms, guided by a carefully designed basic prompt. For incorrect submissions, we construct a fine-grained error taxonomy and then propose a targeted improvement framework by combining a multi-turn dialogue-based repair phase and an information-augmented regeneration phase. Results: Experimental results show that only 5 out of 80 problems are fully accepted when using basic prompts. For the unsolved problems, we construct the error taxonomy, including general errors (such as design, boundary, condition, data type, syntax, and input/output errors) and specialized errors (such as those in mathematical problems, greedy algorithms, and graph theories). After applying our proposed improvement strategies, we substantially increased the number of correct solutions, with 46 out of 80 problems successfully accepted.

cs.SI↗

arXiv2025

Safety Degradation in AI Agents

Despite the growing integration of retrieval-enabled AI agents into society, their safety and ethical behavior remain inadequately understood. In particular, the integration of LLMs and AI agents with external information sources and real-world environments raises critical questions about how they engage with and are influenced by these external data sources and interactive contexts. This study investigates how expanding retrieval access -- from no external sources to Wikipedia-based retrieval and open web search -- affects model reliability, bias propagation, and harmful content generation. Through extensive benchmarking of censored and uncensored LLMs and AI agents, our findings reveal a consistent degradation in refusal rates, bias sensitivity, and harmfulness safeguards as models gain broader access to external sources, culminating in a phenomenon we term safety degradation. Notably, retrieval-enabled agents built on aligned LLMs often behave more unsafely than uncensored models without retrieval. This effect persists even under strong retrieval accuracy and prompt-based mitigation, suggesting that the mere presence of retrieved content reshapes model behavior in structurally unsafe ways. These findings underscore the need for robust mitigation strategies to ensure fairness and reliability in retrieval-enabled and increasingly autonomous AI systems.

cs.CY↗

arXiv2025

First Lasing and Stable Operation of a Direct-Amplification Enabled Harmonic Generation Free-Electron laser

Seeded free-electron lasers (FELs) capable of operating at repetition rates up to the MHz level are in high demand for advanced time-resolved spectroscopies, which require both full longitudinal coherence and high average photon flux in the extreme ultraviolet (EUV) and x-ray regimes. However, conventional external-seed laser systems cannot sustain MHz operation with sufficient hundreds of megawatts peak power requirement due to their limited total power. Here, we report the first lasing and stable operation of a direct-amplification-enabled harmonic generation FEL driven by a weak seed laser with MW-level peak power. Beginning with an ultraviolet seed laser with only 0.75 {\mu}J pulse energy, we demonstrate its direct amplification to over 10 {\mu}J within an 8-meter-long modulator. We observe coherent harmonic generation up to the 12th harmonic of the seed and achieve saturation of the 7th harmonic in the radiator. These results represent a crucial milestone toward the realization of MHz-class, fully coherent EUV and x-ray light sources.

physics.acc-ph↗

arXiv2025

Enabling Continuous THz Band Coverage via Precise Electron Beam Tailoring in Free-electron Lasers

High-power, continuously tunable narrowband terahertz (THz) sources are essential for advancing nonlinear optics, THz-driven material dynamics, and ultrafast spectroscopy. Conventional techniques typically impose a trade-off between pulse energy and frequency tunability. Here, we introduce a novel free-electron laser approach that overcomes these limitations by pre-modulating a relativistic electron beam with a frequency-beating laser pulse and leveraging bunch compression along with collective effects to enhance microbunching. Experimental results demonstrate that this technique generates narrowband THz emission with continuous frequency tunability from 7.8 to 30.8THz, achieving pulse energies up to 385{\mu}J while maintaining spectral bandwidths between 7.7% and 14.7%. Moreover, the method exhibits exceptional robustness and scalability, highlighting its unique ability to bridge the long-standing THz gap and offering a promising solution for diverse cutting-edge scientific applications.

physics.acc-ph↗

arXiv2025

Synchronized Video-to-Audio Generation via Mel Quantization-Continuum Decomposition

Video-to-audio generation is essential for synthesizing realistic audio tracks that synchronize effectively with silent videos. Following the perspective of extracting essential signals from videos that can precisely control the mature text-to-audio generative diffusion models, this paper presents how to balance the representation of mel-spectrograms in terms of completeness and complexity through a new approach called Mel Quantization-Continuum Decomposition (Mel-QCD). We decompose the mel-spectrogram into three distinct types of signals, employing quantization or continuity to them, we can effectively predict them from video by a devised video-to-all (V2X) predictor. Then, the predicted signals are recomposed and fed into a ControlNet, along with a textual inversion design, to control the audio generation process. Our proposed Mel-QCD method demonstrates state-of-the-art performance across eight metrics, evaluating dimensions such as quality, synchronization, and semantic consistency. Our codes and demos will be released at \href{Website}{https://wjc2830.github.io/MelQCD/}.

cs.SD↗

arXiv2025

Tensor dynamic conditional correlation model: A new way to pursuit "Holy Grail of investing"

Style investing creates asset classes (or the so-called "styles") with low correlations, aligning well with the principle of "Holy Grail of investing" in terms of portfolio selection. The returns of styles naturally form a tensor-valued time series, which requires new tools for studying the dynamics of the conditional correlation matrix to facilitate the aforementioned principle. Towards this goal, we introduce a new tensor dynamic conditional correlation (TDCC) model, which is based on two novel treatments: trace-normalization and dimension-normalization. These two normalizations adapt to the tensor nature of the data, and they are necessary except when the tensor data reduce to vector data. Moreover, we provide an easy-to-implement estimation procedure for the TDCC model, and examine its finite sample performance by simulations. Finally, we assess the usefulness of the TDCC model in international portfolio selection across ten global markets and in large portfolio selection for 1800 stocks from the Chinese stock market.

q-fin.PM↗

arXiv2025

Large covariance matrix estimation with factor-assisted variable clustering

This paper studies the covariance matrix estimation for high-dimensional time series within a new framework that combines low-rank factor and latent variable-specific cluster structures. The popular methods based on assuming the sparse error covariance matrix after taking out common factors may be invalid for many financial applications. Our formulation postulates a latent model-based error cluster structure after removing observable factors, which not only leads to more interpretable cluster patterns but also accounts for non-sparse cross-sectional correlations among the variable-specific residuals. Our method begins with using least-squares to estimate the factor loadings, followed by identifying the latent cluster structure by thresholding the scaled covariance difference measures of residuals. A novel ratio-based criterion is introduced to determine the threshold parameter when performing the developed clustering algorithm. We then establish the cluster recovery consistency of our method and derive the convergence rates of our proposed covariance matrix estimators under different norms. Finally, we demonstrate the superior finite sample performance of our proposal over the competing methods through both extensive simulations and a real data application on minimum variance portfolio.

stat.ME↗

arXiv2024

FaceChain-FACT: Face Adapter with Decoupled Training for Identity-preserved Personalization

In the field of human-centric personalized image generation, the adapter-based method obtains the ability to customize and generate portraits by text-to-image training on facial data. This allows for identity-preserved personalization without additional fine-tuning in inference. Although there are improvements in efficiency and fidelity, there is often a significant performance decrease in test following ability, controllability, and diversity of generated faces compared to the base model. In this paper, we analyze that the performance degradation is attributed to the failure to decouple identity features from other attributes during extraction, as well as the failure to decouple the portrait generation training from the overall generation task. To address these issues, we propose the Face Adapter with deCoupled Training (FACT) framework, focusing on both model architecture and training strategy. To decouple identity features from others, we leverage a transformer-based face-export encoder and harness fine-grained identity features. To decouple the portrait generation training, we propose Face Adapting Increment Regularization~(FAIR), which effectively constrains the effect of face adapters on the facial region, preserving the generative ability of the base model. Additionally, we incorporate a face condition drop and shuffle mechanism, combined with curriculum learning, to enhance facial controllability and diversity. As a result, FACT solely learns identity preservation from training data, thereby minimizing the impact on the original text-to-image capabilities of the base model. Extensive experiments show that FACT has both controllability and fidelity in both text-to-image generation and inpainting solutions for portrait generation.

cs.CV↗

arXiv2024

Dynamic Depth Decoding: Faster Speculative Decoding for LLMs

The acceleration of Large Language Models (LLMs) with speculative decoding provides a significant runtime improvement without any loss of accuracy. Currently, EAGLE-2 is the state-of-the-art speculative decoding method, improving on EAGLE with a dynamic draft tree. We introduce Dynamic Depth Decoding (DDD), which optimises EAGLE-2's tree drafting method using a dynamic depth. This extends the average speedup that EAGLE-2 achieves over EAGLE by $44\%$, giving DDD an average speedup of $3.16$x.

cs.CL↗

arXiv2024

Two-way Matrix Autoregressive Model with Thresholds

Recently, matrix-valued time series data have attracted significant attention in the literature with the recognition of threshold nonlinearity representing a significant advance. However, given the fact that a matrix is a two-array structure, it is unfortunate, perhaps even unusual, for the threshold literature to focus on using the same threshold variable for the rows and the columns. In fact, evidence in economic, financial, environmental and other data shows advantages of allowing the possibilities of two different threshold variables (with possibly different threshold parameters for rows and columns), hence the need for a Two-way Matrix AutoRegressive model with Thresholds (2-MART). Naturally, two threshold variables pose new and perhaps even fierce challenges, which might be the reason behind the adoption of only one threshold variable in the literature up to now. In this paper, we develop a comprehensive methodology for the 2-MART model, by overcoming various challenges. Compared with existing models in the literature, the new model can achieve greater dimension reduction, much better model fitting, more accurate predictions, and more plausible interpretations.

stat.ME↗