arXiv subjects

Jiaqing Li

Publications and source records attributed to Jiaqing Li.

6 recordsLinked to original sources

arXiv2026

Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO

Evolution Strategies (ES) have recently emerged as a memory-efficient post-training paradigm for LLM reasoning. However, the optimization behavior of ES remains understudied, making it hard to define its advantage scope compared to mainstream post-training paradigms (e.g., Group Relative Policy Optimization (GRPO)). By systematically investigating ES dynamics and mechanisms, this paper first identifies a performance advantage of ES over GRPO, theoretically and empirically showing that ES can lead to broader reasoning coverage, thereby better exploiting the reasoning capabilities of pretrained LLMs. Theoretically, we show that verifier-projected Jensen-Shannon diversity across the ES population is helpful to higher Pass@K performances. Empirically, unlike GRPO, which exhibits entropy collapse, ES improves Pass@1 while attaining higher Pass@K than GRPO. We further develop a sequential GRPO-ES training strategy that combines GRPO's strength in Pass@1 with ES's gains in Pass@K. Second, we find that despite substantial whole-model parameter drift, the task-performance gains of ES are only contributed to a sparse subset of larger-magnitude updates. This functional sparsity suggests that large parameter movement need not imply widespread functional change, and held-out evaluations further show that it does not necessarily lead to catastrophic forgetting. Finally, we study how hyperparameter design affects the effectiveness of ES, demonstrating that ES requires a smaller population size in a larger LLM. These findings position ES as a distinct reasoning post-training paradigm rather than a less effective, memory-efficient alternative to GRPO.

cs.LG↗

arXiv2026

Single-Sample Black-Box Membership Inference Attack against Vision-Language Models via Cross-modal Semantic Alignment

Vision-Language Models (VLMs) have achieved remarkable success, yet their reliance on massive datasets and unintended memorization of training data raise significant data security risk. Membership Inference Attacks (MIAs) aim to assess these risks by determining whether a data sample was included in a model's training set. However, existing MIA methods against VLMs face critical bottlenecks: gray-box method relies on internal logits that are typically restricted in real-world Application Programming Interfaces (APIs), while black-box method depends on large-scale statistical distributions, which struggle in single-sample scenarios. To this end, we investigate MIAs from the perspective of cross-modal semantic alignment, and observe that member images exhibit significantly stronger image-caption alignment due to training memorization, whereas generated captions for non-members may deviate from the original visual content. Leveraging this insight, we propose a novel MIA framework designed for strict black-box and single-sample setting that quantifies such alignment within a joint embedding space, thereby bypassing these unrealistic assumptions. We conducted extensive experiments on three open-source and two closed-source VLMs. On the VL-MIA/Flicker dataset, our method achieves an AUC of 0.821 against LLaVA-1.5, significantly outperforming existing baselines. Furthermore, it remains robust under diverse image perturbations, highlighting its practicality.

cs.CV↗

arXiv2026

When Safe Models Merge into Danger: Exploiting Latent Vulnerabilities in LLM Fusion

Model merging has emerged as a powerful technique for combining specialized capabilities from multiple fine-tuned LLMs without additional training costs. However, the security implications of this widely-adopted practice remain critically underexplored. In this work, we reveal that model merging introduces a novel attack surface that can be systematically exploited to compromise safety alignment. We present TrojanMerge,, a framework that embeds latent malicious components into source models that remain individually benign but produce severely misaligned models when merged. Our key insight is formulating this attack as a constrained optimization problem: we construct perturbations that preserve source model safety through directional consistency constraints, maintain capabilities via Frobenius directional alignment constraints, yet combine during merging to form pre-computed attack vectors. Extensive experiments across 9 LLMs from 3 model families demonstrate that TrojanMerge, consistently achieves high harmful response rates in merged models while source models maintain safety scores comparable to unmodified versions. Our attack succeeds across diverse merging algorithms and remains effective under various hyperparameter configurations. These findings expose fundamental vulnerabilities in current model merging practices and highlight the urgent need for security-aware mechanisms.

cs.CR↗

arXiv2026

Towards Scalable Training for Handwritten Mathematical Expression Recognition

Large foundation models have achieved significant performance gains through scalable training on massive datasets. However, the field of \textbf{H}andwritten \textbf{M}athematical \textbf{E}xpression \textbf{R}ecognition (HMER) has been impeded by the scarcity of data, primarily due to the arduous and costly process of manual annotation. To bridge this gap, we propose a novel method integrating limited handwritten formulas with large-scale LaTeX-rendered formulas by developing a scalable data engine to generate complex and consistent LaTeX sequences. With this engine, we built the largest formula dataset to date, termed \texttt{Tex80M}, comprising over 80 million high-quality training instances. Then we propose \texttt{TexTeller}, the first HMER model trained at scale, by mix-training \texttt{Tex80M} with a relatively small HME dataset. The expansive training dataset and our refined pipeline have equipped \texttt{TexTeller} with state-of-the-art (SOTA) performance across nearly all benchmarks. To advance the field, we will openly release our complete model, entire dataset, and full codebase, enabling further research building upon our contributions.

cs.CV↗

arXiv2024

Scalable and Versatile Linear Computation with Minimalistic Photonic Matrix Processor

The advancement of artificial intelligence demands flexible multimodal data processing with high throughput and energy efficiency. Photonic integrated circuits (PIC) has demonstrated promising potentials in terms of low latency and low power consumption per operation for linear operations such as matrix-vector multiplication. However, the existing schemes face challenges in their scalability due to the use of photonic circuits that expand with the scale of the operants, despite efforts of exploiting the multiple optical parameter dimensions such as time, wavelength and spatial parallelism. They also lacked flexibility and efficiency in switching between different types of operations or tasks and adapting to multimodal data. In this article, we introduce an optical matrix processor (MP) with a minimalistic recursive structure for both multiplications and accumulations. The MP consists of an eletro-optic ring-modulator implemented as a thin-film lithium niobate PIC that allows flexible configurability and time-division multiplexed scheduling. The MP supports not only versatile linear operations including vector/matrix-vector multiplication and single/multi-kernel convolution but also ultrafast task switching and adaptability to data of different sizes, by simply adjusting the data baud rate relative to the ring delay without structural modifications. We demonstrate its capabilities in a optic-electronic convolutional neural network with a computing throughput up to 73.4 billion operations per second. The MP further supports high scalability through appropriate allocation of wavelength and space resources,extending computing parallelism to handle higher data volumes with higher energy efficiency. This novel scheme paves the way for a new class of photonic processorscapable of managing escalating data workloads with unprecedented flexibility, efficiency and scalability.

physics.optics↗

arXiv2023

Measurement of $^{19}$F($p$,$γ$)$^{20}$Ne reaction suggests CNO break-out in first stars

The origin of calcium production in the first stars (Pop III stars), which formed out of the primordial matter of the Big Bang, and their fates, remain most fascinating mysteries in astrophysics. Advanced nuclear burning and supernovae were thought to be the dominant source of the Ca production seen in all stars. Here we report on a qualitatively different path to Ca production through break-out from the "warm" carbon-nitrogen-oxygen (CNO) cycle. We extend direct measurement of the $^{19}$F($p$, $γ$)$^{20}$Ne break-out reaction down to an unprecedentedly low energy point of 186 keV and discover a key resonance at 225 keV. In the domain of astrophysical interest, at around 0.1 giga kelvin, this thermonuclear $^{19}$F($p$,$γ$)$^{20}$Ne rate is up to a factor of 7.4 larger than the previous recommended rate. Our stellar models show a stronger break-out during stellar hydrogen burning than thought before, and may reveal the nature of Ca production in Pop III stars imprinted on the oldest known ultra-iron poor star, SMSS0313-6708. This result from the China Jinping Underground Laboratory, the deepest laboratory in the world, offering an environment with extremely low cosmic-ray induced background, has far-reaching implications on our understanding of how the first stars evolve and die. Our rate showcases the impact that faint Pop III star supernovae can have on the nucleosynthesis observed in the oldest known stars and first galaxies, key mission targets of the James Webb Space Telescope.

astro-ph.SR↗