Kun Wan

Xingzhe He, Zhiwen Cao, Nicholas Kolkin et al.

Large text-to-image models have revolutionized the ability to generate imagery using natural language. However, particularly unique or personal visual concepts, such as pets and furniture, will not be captured by the original model. This has led to interest in how to personalize a text-to-image model. Despite significant progress, this task remains a formidable challenge, particularly in preserving the subject's identity. Most researchers attempt to address this issue by modifying model architectures. These methods are capable of keeping the subject structure and color but fail to preserve identity details. Towards this issue, our approach takes a data-centric perspective. We introduce a novel regularization dataset generation strategy on both the text and image level. This strategy enables the model to preserve fine details of the desired subjects, such as text and logos. Our method is architecture-agnostic and can be flexibly applied on various text-to-image models. We show on established benchmarks that our data-centric approach forms the new state of the art in terms of identity preservation and text alignment.

Yifan Zeng, Yiran Wu, Yaolun Zhang et al.

Multi-agent LLM workflows route inference through specialized roles to lift end-task accuracy, but jointly training those roles with reinforcement learning is unstable in ways that are poorly understood. We study when end-to-end RL training of multi-agent LLM workflows improves over their base models, comparing Shared-Policy training, where all roles update one policy, with Isolated-Policy training, where each role has its own parameters. Our experimental matrix spans Eval-Opt, Voting, and Orch-Workers workflows, math and code tasks, and three model scales (0.6B, 1.7B, 4B). We find that multi-agent RL usually improves over base models, but gains depend jointly on workflow, task, and scale, not on policy sharing alone. Isolated-Policy tends to reach higher peak accuracy yet more often falls off a terminal accuracy cliff, while Shared-Policy training does not eliminate failure; it redistributes failure into qualitatively different patterns. We then explain the strongest of these patterns through role-level gradient dynamics induced by workflow topology and policy routing: under Isolated-Policy, parallel same-role agents on shared prompts amplify per-role gradients and drive terminal degradation in Voting and Orch-Workers workflows; under Shared-Policy, asymmetric per-step gradient mass causes the shared policy to be captured by the dominant role, producing different failure signatures by task and workflow. Together, the empirical map and its underlying mechanisms show that policy sharing routes training pressure through different channels rather than offering uniform stability, making it a design choice with workflow- and task-conditional tradeoffs.

Zhenting Wang, Guofeng Cui, Yu-Jhe Li et al.

Recent advances in reinforcement learning (RL)-based post-training have led to notable improvements in large language models (LLMs), particularly in enhancing their reasoning capabilities to handle complex tasks. However, most existing methods treat the training data as a unified whole, overlooking the fact that modern LLM training often involves a mixture of data from diverse distributions-varying in both source and difficulty. This heterogeneity introduces a key challenge: how to adaptively schedule training across distributions to optimize learning efficiency. In this paper, we present a principled curriculum learning framework grounded in the notion of distribution-level learnability. Our core insight is that the magnitude of policy advantages reflects how much a model can still benefit from further training on a given distribution. Based on this, we propose a distribution-level curriculum learning framework for RL-based LLM post-training, which leverages the Upper Confidence Bound (UCB) principle to dynamically adjust sampling probabilities for different distrubutions. This approach prioritizes distributions with either high average advantage (exploitation) or low sample count (exploration), yielding an adaptive and theoretically grounded training schedule. We instantiate our curriculum learning framework with GRPO as the underlying RL algorithm and demonstrate its effectiveness on logic reasoning datasets with multiple difficulties and sources. Our experiments show that our framework significantly improves convergence speed and final performance, highlighting the value of distribution-aware curriculum strategies in LLM post-training. Code: https://github.com/ZhentingWang/DUMP.

Peilin Wu, Xinlu Zhang, Kun Wan et al.

Rubric-based reward shaping is an effective method for fine-tuning LLMs via RL, where structured rubrics decompose standard outcome rewards into multiple dimensions to provide richer reward signals. Recent works make the rubrics adaptive based on local signals such as the rollouts from the current step or pairwise comparisons. However, these methods discard the diagnostics produced during evaluation after immediate use and prevent the long-term accumulation and strategic reuse of evaluation knowledge. This forces the system to re-derive evaluation principles from scratch, limits its ability to detect recurring suboptimal behaviors, and forfeits the curriculum-like progression that a persistent training history would naturally support. To address these limitations, we introduce AMARIS, which grounds rubric modifications in long-term training history. At each training step, AMARIS analyzes individual rollouts, aggregates findings into step-level summaries, retrieves relevant historical context from a persistent evaluation memory through both static (recent steps) and dynamic (semantically matched) retrieval, and updates rubrics based on these accumulated analyses. This procedure runs asynchronously alongside the normal RL loop with minimal overhead. Experiments across both closed and open-ended domains show that AMARIS consistently outperforms the baselines. Ablation studies show that static and dynamic memory retrieval contributes to the performance gain and their combination provides the strongest results with moderate retrieval budgets sufficient to provide most of the gain, and that the entire pipeline adds only ~5\% time overhead through asynchronous execution. These results show that persistent evaluation memory can transform rubric-based reward shaping from a stateless, per-step heuristic into an evidence-driven loop for RL training.

Qinsi Wang, Bo Liu, Tianyi Zhou et al.

Although reinforcement learning (RL) can effectively enhance the reasoning capabilities of vision-language models (VLMs), current methods remain heavily dependent on labor-intensive datasets that require extensive manual construction and verification, leading to extremely high training costs and consequently constraining the practical deployment of VLMs. To address this challenge, we propose Vision-Zero, a domain-agnostic framework enabling VLM self-improvement through competitive visual games generated from arbitrary image pairs. Specifically, Vision-Zero encompasses three main attributes: (1) Strategic Self-Play Framework: Vision-Zero trains VLMs in "Who Is the Spy"-style games, where the models engage in strategic reasoning and actions across multiple roles. Through interactive gameplay, models autonomously generate their training data without human annotation. (2) Gameplay from Arbitrary Images: Unlike existing gamified frameworks, Vision-Zero can generate games from arbitrary images, thereby enhancing the model's reasoning ability across diverse domains and showing strong generalization to different tasks. We demonstrate this versatility using three distinct types of image datasets: CLEVR-based synthetic scenes, charts, and real-world images. (3) Sustainable Performance Gain: We introduce Iterative Self-Play Policy Optimization (Iterative-SPO), a novel training algorithm that alternates between Self-Play and reinforcement learning with verifiable rewards (RLVR), mitigating the performance plateau often seen in self-play-only training and achieving sustained long-term improvements. Despite using label-free data, Vision-Zero achieves state-of-the-art performance on reasoning, chart question answering, and vision-centric understanding tasks, surpassing other annotation-based methods. Models and code has been released at https://github.com/wangqinsi1/Vision-Zero.

Jiayi Yuan, Hao Li, Xinheng Ding et al.

Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration, such as evaluation batch size, GPU count, and GPU version, can introduce significant differences in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision - while critical for reproducibility - is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

Lu Ling, Yichen Sheng, Zhi Tu et al.

We have witnessed significant progress in deep learning-based 3D vision, ranging from neural radiance field (NeRF) based 3D representation learning to applications in novel view synthesis (NVS). However, existing scene-level datasets for deep learning-based 3D vision, limited to either synthetic environments or a narrow selection of real-world scenes, are quite insufficient. This insufficiency not only hinders a comprehensive benchmark of existing methods but also caps what could be explored in deep learning-based 3D analysis. To address this critical gap, we present DL3DV-10K, a large-scale scene dataset, featuring 51.2 million frames from 10,510 videos captured from 65 types of point-of-interest (POI) locations, covering both bounded and unbounded scenes, with different levels of reflection, transparency, and lighting. We conducted a comprehensive benchmark of recent NVS methods on DL3DV-10K, which revealed valuable insights for future research in NVS. In addition, we have obtained encouraging results in a pilot study to learn generalizable NeRF from DL3DV-10K, which manifests the necessity of a large-scale scene-level dataset to forge a path toward a foundation model for learning 3D representation. Our DL3DV-10K dataset, benchmark results, and models will be publicly accessible at https://dl3dv-10k.github.io/DL3DV-10K/.

Insu Han, Zeliang Zhang, Zhiyuan Wang et al.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance across diverse applications. However, their computational overhead during deployment remains a critical bottleneck. While Key-Value (KV) caching effectively trades memory for computation to enhance inference efficiency, the growing memory footprint from extensive KV caches significantly reduces throughput and restricts prolonged deployment on memory-constrained GPU devices. To address this challenge, we propose CalibQuant, a simple yet highly effective visual quantization strategy that drastically reduces both memory and computational overhead. Specifically, CalibQuant introduces an extreme 1-bit quantization scheme, complemented by novel post-scaling and calibration techniques tailored to the intrinsic patterns of KV caches, thereby ensuring high efficiency without compromising model performance. Leveraging Triton for runtime optimization, we achieve a 10x throughput increase on InternVL models. Our method is designed to be plug-and-play, seamlessly integrating with various existing MLLMs without requiring architectural changes. Extensive experiments confirm that our approach significantly reduces memory usage while maintaining computational efficiency and preserving multimodal capabilities. Codes are available at https://github.com/insuhan/calibquant.

Shiyu Ji, Kun Wan

In this paper we compare the performance of various homomorphic encryption methods on a private search scheme that can achieve $k$-anonymity privacy. To make our benchmarking fair, we use open sourced cryptographic libraries which are written by experts and well scrutinized. We find that Goldwasser-Micali encryption achieves good enough performance for practical use, whereas fully homomorphic encryptions are much slower than partial ones like Goldwasser-Micali and Paillier.

Shijian Deng, Wentian Zhao, Yu-Jhe Li et al.

Self-improvement in multimodal large language models (MLLMs) is crucial for enhancing their reliability and robustness. However, current methods often rely heavily on MLLMs themselves as judges, leading to high computational costs and potential pitfalls like reward hacking and model collapse. This paper introduces a novel, model-level judge-free self-improvement framework. Our approach employs a controlled feedback mechanism while eliminating the need for MLLMs in the verification loop. We generate preference learning pairs using a controllable hallucination mechanism and optimize data quality by leveraging lightweight, contrastive language-image encoders to evaluate and reverse pairs when necessary. Evaluations across public benchmarks and our newly introduced IC dataset designed to challenge hallucination control demonstrate that our model outperforms conventional techniques. We achieve superior precision and recall with significantly lower computational demands. This method offers an efficient pathway to scalable self-improvement in MLLMs, balancing performance gains with reduced resource requirements.

Yu-Jhe Li, Xinyang Zhang, Kun Wan et al.

We tackle the challenge of open-vocabulary segmentation, where we need to identify objects from a wide range of categories in different environments, using text prompts as our input. To overcome this challenge, existing methods often use multi-modal models like CLIP, which combine image and text features in a shared embedding space to bridge the gap between limited and extensive vocabulary recognition, resulting in a two-stage approach: In the first stage, a mask generator takes an input image to generate mask proposals, and the in the second stage the target mask is picked based on the query. However, the expected target mask may not exist in the generated mask proposals, which leads to an unexpected output mask. In our work, we propose a novel approach named Prompt-guided Mask Proposal (PMP) where the mask generator takes the input text prompts and generates masks guided by these prompts. Compared with mask proposals generated without input prompts, masks generated by PMP are better aligned with the input prompts. To realize PMP, we designed a cross-attention mechanism between text tokens and query tokens which is capable of generating prompt-guided mask proposals after each decoding. We combined our PMP with several existing works employing a query-based segmentation backbone and the experiments on five benchmark datasets demonstrate the effectiveness of this approach, showcasing significant improvements over the current two-stage models (1% ~ 3% absolute performance gain in terms of mIOU). The steady improvement in performance across these benchmarks indicates the effective generalization of our proposed lightweight prompt-aware method.

Peilin Wu, Mian Zhang, Kun Wan et al.

Agentic RAG is a powerful technique for incorporating external information that LLMs lack, enabling better problem solving and question answering. However, suboptimal search behaviors exist widely, such as over-search (retrieving information already known) and under-search (failing to search when necessary), which leads to unnecessary overhead and unreliable outputs. Current training methods, which typically rely on outcome-based rewards in a RL framework, lack the fine-grained control needed to address these inefficiencies. To overcome this, we introduce Hierarchical Process Rewards for Efficient agentic RAG (HiPRAG), a training methodology that incorporates a fine-grained, knowledge-grounded process reward into the RL training. Our approach evaluates the necessity of each search decision on-the-fly by decomposing the agent's reasoning trajectory into discrete, parsable steps. We then apply a hierarchical reward function that provides an additional bonus based on the proportion of optimal search and non-search steps, on top of commonly used outcome and format rewards. Experiments on the Qwen2.5 and Llama-3.2 models across seven diverse QA benchmarks show that our method achieves average accuracies of 65.4% (3B) and 67.2% (7B). This is accomplished while improving search efficiency, reducing the over-search rate to just 2.3% and concurrently lowering the under-search rate. These results demonstrate the efficacy of optimizing the reasoning process itself, not just the final outcome. Further experiments and analysis demonstrate that HiPRAG shows good generalizability across a wide range of RL algorithms, model families, sizes, and types. This work demonstrates the importance and potential of fine-grained control through RL, for improving the efficiency and optimality of reasoning for search agents.

Kaiyu He, Peilin Wu, Mian Zhang et al.

Since the advent of large language models (LLMs), research has focused on instruction following and deductive reasoning. A central question remains: can these models discover new knowledge, and how can we evaluate this ability? We address this by studying abductive reasoning-the generation of plausible hypotheses to explain observations-and introduce GEAR (General Evaluation for Abductive Reasoning), a general-purpose, fully automated, transparent, and label-free evaluation paradigm. GEAR scores hypothesis sets by three metrics: consistency (each hypothesis explains the observations), generalizability (consistent hypotheses make meaningful predictions on unseen inputs), and diversity (the set covers distinct predictions and patterns). Built this way, GEAR is scalable (no human gold answers), reliable (deterministic scoring aligned with classical abduction), and open-ended (scores improve only when models produce new plausible hypotheses, unlike static benchmarks that saturate once accuracy is high). Using GEAR, we conduct a fine-grained study of nine LLMs on four abduction benchmarks with 1,500 problems, generating over 50,000 candidate hypotheses and revealing model differences obscured by gold-answer or purely human evaluations. We further propose a momentum-based curriculum that adjusts GEAR-derived training data by learning velocity: it starts with what the model learns quickly and shifts toward harder objectives such as generating diverse hypotheses once the model is confident on foundational objectives. Without gold-label supervision, this strategy improves all GEAR objectives and these gains transfer to established abductive reasoning benchmarks. Taken together, GEAR provides a principled framework that evaluates abduction and supplies label-free, scalable training signals that help LLMs produce more diverse and reliable hypotheses.

Wujiang Xu, Wentian Zhao, Zhenting Wang et al.

Training LLM agents in multi-turn environments with sparse rewards, where completing a single task requires 30+ turns of interaction within an episode, presents a fundamental challenge for reinforcement learning. We identify a critical failure mode unique to this setting: the exploration-exploitation cascade failure. This cascade begins with early-stage policy premature convergence, where sparse feedback causes agents to commit to flawed, low-entropy strategies. Subsequently, agents enter late-stage policy collapse, where conventional entropy regularization becomes counterproductive, promoting chaotic exploration that destabilizes training. We propose Entropy-regularized Policy Optimization (EPO), a general framework that breaks this failure cycle through three synergistic mechanisms: (1) adopting entropy regularization in multi-turn settings to enhance exploration, (2) an entropy smoothing regularizer that bounds policy entropy within historical averages to prevent abrupt fluctuations, and (3) adaptive phase-based weighting that balances exploration and exploitation across training. Our analysis justifies that EPO guarantees monotonically decreasing entropy variance while maintaining convergence. EPO achieves up to 152% performance improvement on ScienceWorld and up to 19.8% on ALFWorld. Our work demonstrates that multi-turn sparse-reward settings require fundamentally different entropy control than traditional RL, with broad implications for LLM agent training.

Xiaolei Liu, Xiaosong Zhang, Kun Wan et al.

Recent studies have highlighted audio adversarial examples as a ubiquitous threat to state-of-the-art automatic speech recognition systems. Thorough studies on how to effectively generate adversarial examples are essential to prevent potential attacks. Despite many research on this, the efficiency and the robustness of existing works are not yet satisfactory. In this paper, we propose~\textit{weighted-sampling audio adversarial examples}, focusing on the numbers and the weights of distortion to reinforce the attack. Further, we apply a denoising method in the loss function to make the adversarial attack more imperceptible. Experiments show that our method is the first in the field to generate audio adversarial examples with low noise and high audio robustness at the minute time-consuming level.

Kun Wan, Boyuan Feng, Shu Yang et al.

With the increasing demand to deploy convolutional neural networks (CNNs) on mobile platforms, the sparse kernel approach was proposed, which could save more parameters than the standard convolution while maintaining accuracy. However, despite the great potential, no prior research has pointed out how to craft an sparse kernel design with such potential (i.e., effective design), and all prior works just adopt simple combinations of existing sparse kernels such as group convolution. Meanwhile due to the large design space it is also impossible to try all combinations of existing sparse kernels. In this paper, we are the first in the field to consider how to craft an effective sparse kernel design by eliminating the large design space. Specifically, we present a sparse kernel scheme to illustrate how to reduce the space from three aspects. First, in terms of composition we remove designs composed of repeated layers. Second, to remove designs with large accuracy degradation, we find an unified property named information field behind various sparse kernel designs, which could directly indicate the final accuracy. Last, we remove designs in two cases where a better parameter efficiency could be achieved. Additionally, we provide detailed efficiency analysis on the final four designs in our scheme. Experimental results validate the idea of our scheme by showing that our scheme is able to find designs which are more efficient in using parameters and computation with similar or higher accuracy.

Lingwei Xie, Song He, Shu Yang et al.

With the rapid development of high-throughput technologies, parallel acquisition of large-scale drug-informatics data provides huge opportunities to improve pharmaceutical research and development. One significant application is the purpose prediction of small molecule compounds, aiming to specify therapeutic properties of extensive purpose-unknown compounds and to repurpose novel therapeutic properties of FDA-approved drugs. Such problem is very challenging since compound attributes contain heterogeneous data with various feature patterns such as drug fingerprint, drug physicochemical property, drug perturbation gene expression. Moreover, there is complex nonlinear dependency among heterogeneous data. In this paper, we propose a novel domain-adversarial multi-task framework for integrating shared knowledge from multiple domains. The framework utilizes the adversarial strategy to effectively learn target representations and models their nonlinear dependency. Experiments on two real-world datasets illustrate that the performance of our approach obtains an obvious improvement over competitive baselines. The novel therapeutic properties of purpose-unknown compounds we predicted are mostly reported or brought to the clinics. Furthermore, our framework can integrate various attributes beyond the three domains examined here and can be applied in the industry for screening the purpose of huge amounts of as yet unidentified compounds. Source codes of this paper are available on Github.

Kun Wan, Boyuan Feng, Lingwei Xie et al.

Recently convolutional neural networks (CNNs) achieve great accuracy in visual recognition tasks. DenseNet becomes one of the most popular CNN models due to its effectiveness in feature-reuse. However, like other CNN models, DenseNets also face overfitting problem if not severer. Existing dropout method can be applied but not as effective due to the introduced nonlinear connections. In particular, the property of feature-reuse in DenseNet will be impeded, and the dropout effect will be weakened by the spatial correlation inside feature maps. To address these problems, we craft the design of a specialized dropout method from three aspects, dropout location, dropout granularity, and dropout probability. The insights attained here could potentially be applied as a general approach for boosting the accuracy of other CNN models with similar nonlinear connections. Experimental results show that DenseNets with our specialized dropout method yield better accuracy compared to vanilla DenseNet and state-of-the-art CNN models, and such accuracy boost increases with the model depth.

Boyuan Feng, Kun Wan, Shu Yang et al.

Despite that accelerating convolutional neural network (CNN) receives an increasing research focus, the save on resource consumption always comes with a decrease in accuracy. To both increase accuracy and decrease resource consumption, we explore an environment information, called class skew, which is easily available and exists widely in daily life. Since the class skew may switch as time goes, we bring up probability layer to utilize class skew without any overhead during the runtime. Further, we observe class skew dichotomy that some class skew may appear frequently in the future, called hot class skew, and others will never appear again or appear seldom, called cold class skew. Inspired by techniques from source code optimization, two modes, i.e., interpretation and compilation, are proposed. The interpretation mode pursues efficient adaption during runtime for cold class skew and the compilation mode aggressively optimize on hot ones for more efficient deployment in the future. Aggressive optimization is processed by class-specific pruning and provides extra benefit. Finally, we design a systematic framework, SECS, to dynamically detect class skew, processing interpretation and compilation, as well as select the most accurate architectures under the runtime resource budget. Extensive evaluations show that SECS can realize end-to-end classification speedups by a factor of 3x to 11x relative to state-of-the-art convolutional neural networks, at a higher accuracy.