Zhe Yuan

Zhe Yuan, Yipeng Zhou, Jinghan Li et al.

Reinforcement learning with verifiable rewards (RLVR) has become an effective paradigm for improving reasoning language models on tasks such as mathematics, coding, and scientific question answering. However, widely used group-relative objectives, such as GRPO, summarize each sampled group with scalar statistics and therefore discard fine-grained relational information among candidate responses. This weakens credit assignment under sparse outcome rewards, especially when multiple generated solutions differ only subtly in reasoning quality. We propose \textbf{LamPO}, a \textbf{Lambda-Style Policy Optimization} method that replaces scalar group advantages with a \emph{Pairwise Decomposed Advantage}. LamPO aggregates pairwise reward gaps within each response group and modulates each comparison by a confidence-aware weight computed from sequence log-probability differences, while retaining the critic-free and clipped-update structure of PPO-style optimization. When reference solutions are available, we further add a lightweight ROUGE-L-based dense auxiliary reward to reduce reward sparsity. Experiments on AIME24, AIME25, MATH-500, and GPQA-Diamond with Qwen3-1.7B, Qwen3-4B, and Phi-4-mini show that LamPO consistently improves over GRPO and recent RLVR variants, with more stable training dynamics and better sample efficiency.

Zhe Yuan, Yipeng Zhou, Jinghan Li et al.

Group Relative Policy Optimization(GRPO) has become a cornerstone of modern reinforcement learning alignment, prized for its efficacy in foregoing an explicit value-critic by leveraging reward normalization across sampled trajectory cohorts. However, the method's reliance on a monolithic statistical baseline, such as the group mean, collapses the relational topology of the trajectory space into a single scalar, thereby erasing the fine-grained preference information essential for navigating complex, rank-sensitive reward landscapes. To address this issue, we introduce a novel framework, Lambda Policy Optimization (LambdaPO), that addresses this information-theoretic bottleneck by re-conceptualizing advantage estimation from a scalar value to a decomposed, pairwise preference structure. Specifically, the advantage for any given trajectory is formulated as the integrated sum of reward differentials against all peers in its cohort, where each pairwise comparison is dynamically attenuated by the policy's own probabilistic confidence in the established preference. To further mitigate the sparsity of binary outcome supervision, we augment the objective with a semantic density reward, derived from the precision-recall alignment between generated reasoning traces and ground-truth solutions. As a result, our method can mine more fine-grained optimization signals from a group of rollouts, guiding the LLM to a better optima. Experimental results across challenging math reasoning and question-answering tasks demonstrates that LambdaPO improves performance compared to the baseline methods.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Bin Hu, Yang Liu, Xizi Liu et al.

Seated VR locomotion in constrained environments, including homes, offices, and transit settings, calls for hardware that is lightweight and deployable, steering that remains continuous enough for curved motion, and a control channel that leaves the hands free for concurrent interaction. Inspired by the steering logic of self-balancing scooters, we present Glide-in-Place, a seated foot locomotion system that maps per-foot fore-aft pressure to a differential-drive model: the two feet act as virtual wheels whose relative drive continuously determines translation and yaw. This lets users move forward, rotate in place, and follow arcs in one unified vocabulary without hand-held input or discrete mode switches. We evaluated Glide-in-Place in a counterbalanced within-subject study with 16 participants against two baselines: joystick control and a seated walking-in-place technique with discrete snap motions. Across two steering-heavy navigation tasks, zig-zag path following with multitasking and curved-path traversal, Glide-in-Place was consistently faster than Seated-WIP, reduced physical demand, and lowered fatigue-related discomfort without significantly differing from joystick control on total VRSQ. We position Glide-in-Place as a deployable hardware-control design point for constrained seated VR: thin insole sensing, continuous foot steering, and lightweight calibration packaged in one compact artifact.

Yuji Roh, Qingyun Liu, Huan Gui et al.

Fine-tuning is becoming widely used for leveraging the power of pre-trained foundation models in new downstream tasks. While there are many successes of fine-tuning on various tasks, recent studies have observed challenges in the generalization of fine-tuned models to unseen distributions (i.e., out-of-distribution; OOD). To improve OOD generalization, some previous studies identify the limitations of fine-tuning data and regulate fine-tuning to preserve the general representation learned from pre-training data. However, potential limitations in the pre-training data and models are often ignored. In this paper, we contend that overly relying on the pre-trained representation may hinder fine-tuning from learning essential representations for downstream tasks and thus hurt its OOD generalization. It can be especially catastrophic when new tasks are from different (sub)domains compared to pre-training data. To address the issues in both pre-training and fine-tuning data, we propose a novel generalizable fine-tuning method LEVI (Layer-wise Ensemble of different VIews), where the pre-trained model is adaptively ensembled layer-wise with a small task-specific model, while preserving its efficiencies. By combining two complementing models, LEVI effectively suppresses problematic features in both the fine-tuning data and pre-trained model and preserves useful features for new tasks. Broad experiments with large language and vision models show that LEVI greatly improves fine-tuning generalization via emphasizing different views from fine-tuning data and pre-trained features.

Xinwei Zhang, Hu Chen, Zhe Yuan et al.

Foundation models for medical image segmentation have achieved remarkable performance. Adaptive fine-tuning of natural image segmentation foundation models is crucial for medical image segmentation tasks. However, some limitations exist in existing fine-tuning methods: 1) insufficient representation of high-level features and 2) the fine-tuning process disrupts the structural integrity of pretrained weights. Inspired by these critical problems, we propose an intelligent communication mixture-of-experts boosted-medical image segmentation foundation model, named IC-MoE, with twofold ideas: 1) We construct basic experts, semantic experts, and adaptive experts. Moreover, we implement a pixel probability adaptive voting strategy, which enables expert selection and fusion through label consistency and load balancing. This approach preliminarily enhances the representation capability of high-level features while preserving the structural integrity of pretrained weights. 2) We propose a semantic-guided contrastive learning method to address the issue of weak supervision in contrastive learning. This method further enhances the representation capability of high-level features while preserving the structural integrity of pretrained weights. Extensive experiments across three public medical image segmentation datasets demonstrate that the IC-MoE outperforms other SOTA models. Consequently, the proposed IC-MoE effectively supplements foundational medical image segmentation models with high-level features and pretrained structural integrity. We also validate the superior generalizability of the IC-MoE across diverse medical image segmentation scenarios.

Lu Fang, Zhe Yuan, Kaifu Zhang et al.

We quantify the impact of Generative Artificial Intelligence (GenAI) on firm productivity through a series of large-scale randomized field experiments involving millions of users and products at a leading cross-border online retail platform. Over six months in 2023-2024, GenAI-based enhancements were integrated into seven consumer-facing business workflows. We find that GenAI adoption significantly increases sales, with treatment effects ranging from $0\%$ to $16.3\%$, depending on GenAI's marginal contribution relative to existing firm practices. Because inputs and prices were held constant across experimental arms, these gains map directly into total factor productivity improvements. Across the four GenAI applications with positive effects, the implied annual incremental value is approximately $\$ 5$ per consumer-an economically meaningful impact given the retailer's scale and the early stage of GenAI adoption. The primary mechanism operates through higher conversion rates, consistent with GenAI reducing frictions in the marketplace and improving consumer experience. We also document substantial heterogeneity: smaller and newer sellers, as well as less experienced consumers, exhibit disproportionately larger gains. Our findings provide novel, large-scale causal evidence on the productivity effects of GenAI in online retail, highlighting both its immediate value and broader potential.

Zhe Yuan, Jianqi Shi, Yanhong Huang

Computer-aided design (CAD) tools are utilized in the manufacturing industry for modeling everything from cups to spacecraft. These programs are complex to use and typically require years of training and experience to master. Structured and well-constrained 2D sketches and 3D constructions are crucial components of CAD modeling. A well-executed CAD model can be seamlessly integrated into the manufacturing process, thereby enhancing production efficiency. Deep generative models of 3D shapes and 3D object reconstruction models have garnered significant research interest. However, most of these models produce discrete forms of 3D objects that are not editable. Moreover, the few models based on CAD operations often have substantial input restrictions. In this work, we fine-tuned pre-trained models to create OpenECAD models (0.55B, 0.89B, 2.4B and 3.1B), leveraging the visual, logical, coding, and general capabilities of visual language models. OpenECAD models can process images of 3D designs as input and generate highly structured 2D sketches and 3D construction commands, ensuring that the designs are editable. These outputs can be directly used with existing CAD tools' APIs to generate project files. To train our network, we created a series of OpenECAD datasets. These datasets are derived from existing public CAD datasets, adjusted and augmented to meet the specific requirements of vision language model (VLM) training. Additionally, we have introduced an approach that utilizes dependency relationships to define and generate sketches, further enriching the content and functionality of the datasets.

Zhe Yuan, Yimin Wen

In many practical data mining scenarios, such as network intrusion detection, Twitter spam detection, and computer-aided diagnosis, a source domain that is different from but related to a target domain is very common. In addition, a large amount of unlabeled data is available in both source and target domains, but labeling each of them is difficult, expensive, time-consuming, and sometime unnecessary. Therefore, it is very important and worthwhile to fully explore the labeled and unlabeled data in source and target domains to settle the task in target domain. In this paper, a new semi-supervised inductive transfer learning framework, named Co-Transfer is proposed. Co-Transfer first generates three TrAdaBoost classifiers for transfer learning from the source domain to the target domain, and meanwhile another three TrAdaBoost classifiers are generated for transfer learning from the target domain to the source domain, using bootstraped samples from the original labeled data. In each round of co-transfer, each group of TrAdaBoost classifiers are refined using the carefully labeled data. Finally, the group of TrAdaBoost classifiers learned to transfer from the source domain to the target domain produce the final hypothesis. Experiments results illustrate Co-Transfer can effectively exploit and reuse the labeled and unlabeled data in source and target domains.

Zhe Yuan, Zhuoran Lyu, Jiwei Li et al.

Recently, end-to-end speech recognition with a hybrid model consisting of the connectionist temporal classification(CTC) and the attention encoder-decoder achieved state-of-the-art results. In this paper, we propose a novel CTC decoder structure based on the experiments we conducted and explore the relation between decoding performance and the depth of encoder. We also apply attention smoothing mechanism to acquire more context information for subword-based decoding. Taken together, these strategies allow us to achieve a word error rate(WER) of 4.43% without LM and 3.34% with RNN-LM on the test-clean subset of the LibriSpeech corpora, which by far are the best reported WERs for end-to-end ASR systems on this dataset.