Xing Shi

Liuji Chen, Dianxing Tang, Xing Shi et al.

Agentic reinforcement learning can induce tool abuse, where models overuse external tools even for queries solvable by internal reasoning. Existing approaches mitigate this issue with uniform tool-use penalties or hard limits, which reduce tool frequency but may also suppress useful tool-assisted exploration. We propose EAPO, an Efficient Agentic Policy Optimization framework that learns selective tool use. EAPO introduces tool-free trajectories into each rollout group, applies difficulty-aware reward shaping to penalize redundant tool calls mainly on easier queries, and uses confidence-aware token reweighting to improve policy learning. Across nine mathematical and knowledge-intensive reasoning benchmarks, EAPO consistently improves the accuracy efficiency trade-off on Qwen2.5-3B, Qwen2.5-7B, and Llama3.1-8B. Compared with GRPO, EAPO improves average performance by 10.45%, 7.27%, and 9.69%, while reducing average tool calls by 18.33%, 18.33%, and 24.59%, respectively. These results show that agents can learn when not to use tools without compromising tool-integrated reasoning.

Ziheng Wu, Jiaqi Xu, Xinyi Zou et al.

Stable Diffusion web UI (SD-WebUI) is a comprehensive project that provides a browser interface based on Gradio library for Stable Diffusion models. In this paper, We propose a novel WebUI plugin called EasyPhoto, which enables the generation of AI portraits. By training a digital doppelganger of a specific user ID using 5 to 20 relevant images, the finetuned model (according to the trained LoRA model) allows for the generation of AI photos using arbitrary templates. Our current implementation supports the modification of multiple persons and different photo styles. Furthermore, we allow users to generate fantastic template image with the strong SDXL model, enhancing EasyPhoto's capabilities to deliver more diverse and satisfactory results. The source code for EasyPhoto is available at: https://github.com/aigc-apps/sd-webui-EasyPhoto. We also support a webui-free version by using diffusers: https://github.com/aigc-apps/EasyPhoto. We are continuously enhancing our efforts to expand the EasyPhoto pipeline, making it suitable for any identification (not limited to just the face), and we enthusiastically welcome any intriguing ideas or suggestions.

Zizhuo Lin, Quanling Liu, Jinsheng Quan et al.

Large language models (LLMs) often solve a task when all instructions are given in a single prompt, but fail when the same information is revealed gradually across turns. When a clean FULL prompt and a RAW-SHARDED conversation contain the same complete user evidence, the model should still arrive at the same answer. We argue that a key reason for this gap is self-anchored drift: responses produced under partial information introduce unsupported assumptions, and those assumptions later distort the final answer. To reduce this effect, we propose Canonical-Context On-Policy Distillation (CCOPD). During training, the same base model is used in two roles: a frozen teacher conditioned on the clean FULL prompt and a trainable student that receives the same evidence incrementally through a multi-turn conversation; CCOPD aligns the student's behavior on its own trajectories with the teacher's canonical full-context behavior. Trained only on math problem conversations, CCOPD yields a 32\% average relative improvement in RAW-SHARDED performance over the original base model across math and five zero-shot out-of-domain task families, while largely preserving full-context performance. Further analyses suggest that CCOPD strengthens grounding in user evidence and reduces sensitivity to contamination from earlier assistant turns.

Yunji Li, Sujian Li, Xing Shi

In this paper, we propose the task of consecutive question generation (CQG), which generates a set of logically related question-answer pairs to understand a whole passage, with a comprehensive consideration of the aspects including accuracy, coverage, and informativeness. To achieve this, we first examine the four key elements of CQG, i.e., question, answer, rationale, and context history, and propose a novel dynamic multitask framework with one main task generating a question-answer pair, and four auxiliary tasks generating other elements. It directly helps the model generate good questions through both joint training and self-reranking. At the same time, to fully explore the worth-asking information in a given passage, we make use of the reranking losses to sample the rationales and search for the best question series globally. Finally, we measure our strategy by QA data augmentation and manual evaluation, as well as a novel application of generated question-answer pairs on DocNLI. We prove that our strategy can improve question generation significantly and benefit multiple related NLP tasks.

Yi Cheng, Renjun Hu, Haochao Ying et al.

Until recently, the question of the effective inductive bias of deep models on tabular data has remained unanswered. This paper investigates the hypothesis that arithmetic feature interaction is necessary for deep tabular learning. To test this point, we create a synthetic tabular dataset with a mild feature interaction assumption and examine a modified transformer architecture enabling arithmetical feature interactions, referred to as AMFormer. Results show that AMFormer outperforms strong counterparts in fine-grained tabular data modeling, data efficiency in training, and generalization. This is attributed to its parallel additive and multiplicative attention operators and prompt-based optimization, which facilitate the separation of tabular samples in an extended space with arithmetically-engineered features. Our extensive experiments on real-world data also validate the consistent effectiveness, efficiency, and rationale of AMFormer, suggesting it has established a strong inductive bias for deep learning on tabular data. Code is available at https://github.com/aigc-apps/AMFormer.

Jiaqi Xu, Bo Liu, Yunkuo Chen et al.

Video-and-language understanding has a variety of applications in the industry, such as video question answering, text-video retrieval, and multi-label classification. Existing video-and-language understanding methods generally adopt heavy multi-modal encoders and feature fusion modules, which consume high computational costs. Specially, they have difficulty dealing with dense video frames or long text prevalent in industrial applications. This paper proposes MuLTI, a highly accurate and efficient video-and-language understanding model that achieves efficient and effective feature fusion and rapid adaptation to downstream tasks. Specifically, we design a Text-Guided MultiWay-Sampler based on adapt-pooling residual mapping and self-attention modules to sample long sequences and fuse multi-modal features, which reduces the computational costs and addresses performance degradation caused by previous samplers. Therefore, MuLTI can handle longer sequences with limited computational costs. Then, to further enhance the model's performance and fill in the lack of pretraining tasks in the video question answering, we propose a new pretraining task named Multiple Choice Modeling. This task bridges the gap between pretraining and downstream tasks and improves the model's ability to align video and text features. Benefiting from the efficient feature fusion module and the new pretraining task, MuLTI achieves state-of-the-art performance on multiple datasets. Implementation and pretrained models will be released.

Renjun Hu, Yi Cheng, Libin Meng et al.

The rapid advancement of large language models (LLMs) has opened new possibilities for their adoption as evaluative judges. This paper introduces Themis, a fine-tuned LLM judge that delivers sophisticated context-aware evaluations. We provide a comprehensive overview of the development pipeline for Themis, highlighting its scenario-dependent evaluation prompts and two novel methods for controlled instruction generation. These designs enable Themis to effectively distill evaluative skills from teacher models, while retaining flexibility for continuous development. We introduce two human-labeled benchmarks for meta-evaluation, demonstrating that Themis can achieve high alignment with human preferences in an economical manner. Additionally, we explore insights into the LLM-as-a-judge paradigm, revealing nuances in performance and the varied effects of reference answers. Notably, we observe that pure knowledge distillation from strong LLMs, though common, does not guarantee performance improvement through scaling. We propose a mitigation strategy based on instruction-following difficulty. Furthermore, we provide practical guidelines covering data balancing, prompt customization, multi-objective training, and metric aggregation. We aim for our method and findings, along with the fine-tuning data, benchmarks, and model checkpoints, to support future research and development in this area.

Yu Feng, Xing Shi, Mengli Cheng et al.

As the task of 2D-to-3D reconstruction has gained significant attention in various real-world scenarios, it becomes crucial to be able to generate high-quality point clouds. Despite the recent success of deep learning models in generating point clouds, there are still challenges in producing high-fidelity results due to the disparities between images and point clouds. While vision transformers (ViT) and diffusion models have shown promise in various vision tasks, their benefits for reconstructing point clouds from images have not been demonstrated yet. In this paper, we first propose a neat and powerful architecture called DiffPoint that combines ViT and diffusion models for the task of point cloud reconstruction. At each diffusion step, we divide the noisy point clouds into irregular patches. Then, using a standard ViT backbone that treats all inputs as tokens (including time information, image embeddings, and noisy patches), we train our model to predict target points based on input images. We evaluate DiffPoint on both single-view and multi-view reconstruction tasks and achieve state-of-the-art results. Additionally, we introduce a unified and flexible feature fusion module for aggregating image features from single or multiple input images. Furthermore, our work demonstrates the feasibility of applying unified architectures across languages and images to improve 3D reconstruction tasks.

Longxuan Yu, Xing Shi, Xianghao Kong et al.

Mutual information (MI) is one of the most general ways to measure relationships between random variables, but estimating this quantity for complex systems is challenging. Denoising diffusion models have recently set a new bar for density estimation, so it is natural to consider whether these methods could also be used to improve MI estimation. Using the recently introduced information-theoretic formulation of denoising diffusion models, we show the diffusion models can be used in a straightforward way to estimate MI. In particular, the MI corresponds to half the gap in the Minimum Mean Square Error (MMSE) between conditional and unconditional diffusion, integrated over all Signal-to-Noise-Ratios (SNRs) in the noising process. Our approach not only passes self-consistency tests but also outperforms traditional and score-based diffusion MI estimators. Furthermore, our method leverages adaptive importance sampling to achieve scalable MI estimation, while maintaining strong performance even when the MI is high.

Yuyang Xu, Yi Cheng, Haochao Ying et al.

Test-time scaling has proven effective in further enhancing the performance of pretrained Large Language Models (LLMs). However, mainstream post-training methods (i.e., reinforcement learning (RL) with chain-of-thought (CoT) reasoning) often incur substantial computational overhead due to auxiliary models and overthinking. In this paper, we empirically reveal that the incorrect answers partially stem from verbose reasoning processes lacking correct self-fix, where errors accumulate across multiple reasoning steps. To this end, we propose Self-traced Step-wise Preference Optimization (SSPO), a pluggable RL process supervision framework that enables fine-grained optimization of each reasoning step. Specifically, SSPO requires neither auxiliary models nor stepwise manual annotations. Instead, it leverages step-wise preference signals generated by the model itself to guide the optimization process for reasoning compression. Experiments demonstrate that the generated reasoning sequences from SSPO are both accurate and succinct, effectively mitigating overthinking behaviors without compromising model performance across diverse domains and languages.

Yuyang Xu, Renjun Hu, Haochao Ying et al.

Multiple-choice question (MCQ) benchmarks are widely used for evaluating Large Language Models (LLMs), yet their reliability is undermined by benchmark contamination. In this study, we reframe contamination as an inherent aspect of learning and seek to disentangle genuine capability acquisition from superficial memorization in LLM evaluation. First, by analyzing model performance under different memorization conditions, we uncover a counterintuitive trend: LLMs perform worse on memorized MCQs than on non-memorized ones, indicating the coexistence of two distinct learning phenomena, i.e., rote memorization and genuine capability learning. To disentangle them, we propose TrinEval, a novel evaluation framework reformulating MCQs into an alternative trinity format, reducing memorization while preserving knowledge assessment. Experiments validate TrinEval's effectiveness in reformulation, and its evaluation reveals that common LLMs may memorize by rote 20.5% of knowledge points (in MMLU on average).

Xing Shi, Yijun Xiao, Kevin Knight

We investigate why neural machine translation (NMT) systems assign high probability to empty translations. We find two explanations. First, label smoothing makes correct-length translations less confident, making it easier for the empty translation to finally outscore them. Second, NMT systems use the same, high-frequency EoS word to end all target sentences, regardless of length. This creates an implicit smoothing that increases zero-length translations. Using different EoS types in target sentences of different lengths exposes and eliminates this implicit smoothing.

Tanfang Chen, Weiwei Wang, Wenyang Wei et al.

This paper describes DiDi AI Labs' submission to the WMT2020 news translation shared task. We participate in the translation direction of Chinese->English. In this direction, we use the Transformer as our baseline model, and integrate several techniques for model enhancement, including data filtering, data selection, back-translation, fine-tuning, model ensembling, and re-ranking. As a result, our submission achieves a BLEU score of $36.6$ in Chinese->English.

Arkady Arkhangorodsky, Amittai Axelrod, Christopher Chu et al.

We create a new task-oriented dialog platform (MEEP) where agents are given considerable freedom in terms of utterances and API calls, but are constrained to work within a push-button environment. We include facilities for collecting human-human dialog corpora, and for training automatic agents in an end-to-end fashion. We demonstrate MEEP with a dialog assistant that lets users specify trip destinations.

Mengli Cheng, Minghui Qiu, Xing Shi et al.

Structured information extraction from document images usually consists of three steps: text detection, text recognition, and text field labeling. While text detection and text recognition have been heavily studied and improved a lot in literature, text field labeling is less explored and still faces many challenges. Existing learning based methods for text labeling task usually require a large amount of labeled examples to train a specific model for each type of document. However, collecting large amounts of document images and labeling them is difficult and sometimes impossible due to privacy issues. Deploying separate models for each type of document also consumes a lot of resources. Facing these challenges, we explore one-shot learning for the text field labeling task. Existing one-shot learning methods for the task are mostly rule-based and have difficulty in labeling fields in crowded regions with few landmarks and fields consisting of multiple separate text regions. To alleviate these problems, we proposed a novel deep end-to-end trainable approach for one-shot text field labeling, which makes use of attention mechanism to transfer the layout information between document images. We further applied conditional random field on the transferred layout information for the refinement of field labeling. We collected and annotated a real-world one-shot field labeling dataset with a large variety of document types and conducted extensive experiments to examine the effectiveness of the proposed model. To stimulate research in this direction, the collected dataset and the one-shot model will be released1.

Xing Shi, Shizhen Xu, Kevin Knight

We present a GPU-based Locality Sensitive Hashing (LSH) algorithm to speed up beam search for sequence models. We utilize the winner-take-all (WTA) hash, which is based on relative ranking order of hidden dimensions and thus resilient to perturbations in numerical values. Our algorithm is designed by fully considering the underling architecture of CUDA-enabled GPUs (Algorithm/Architecture Co-design): 1) A parallel Cuckoo hash table is applied for LSH code lookup (guaranteed O(1) lookup time); 2) Candidate lists are shared across beams to maximize the parallelism; 3) Top frequent words are merged into candidate lists to improve performance. Experiments on 4 large-scale neural machine translation models demonstrate that our algorithm can achieve up to 4x speedup on softmax module, and 2x overall speedup without hurting BLEU on GPU.

Kuan Liu, Xing Shi, Prem Natarajan

Attributes, such as metadata and profile, carry useful information which in principle can help improve accuracy in recommender systems. However, existing approaches have difficulty in fully leveraging attribute information due to practical challenges such as heterogeneity and sparseness. These approaches also fail to combine recurrent neural networks which have recently shown effectiveness in item recommendations in applications such as video and music browsing. To overcome the challenges and to harvest the advantages of sequence models, we present a novel approach, Heterogeneous Attribute Recurrent Neural Networks (HA-RNN), which incorporates heterogeneous attributes and captures sequential dependencies in \textit{both} items and attributes. HA-RNN extends recurrent neural networks with 1) a hierarchical attribute combination input layer and 2) an output attribute embedding layer. We conduct extensive experiments on two large-scale datasets. The new approach show significant improvements over the state-of-the-art models. Our ablation experiments demonstrate the effectiveness of the two components to address heterogeneous attribute challenges including variable lengths and attribute sparseness. We further investigate why sequence modeling works well by conducting exploratory studies and show sequence models are more effective when data scale increases.

Kuan Liu, Xing Shi, Anoop Kumar et al.

We present our solution to the job recommendation task for RecSys Challenge 2016. The main contribution of our work is to combine temporal learning with sequence modeling to capture complex user-item activity patterns to improve job recommendations. First, we propose a time-based ranking model applied to historical observations and a hybrid matrix factorization over time re-weighted interactions. Second, we exploit sequence properties in user-items activities and develop a RNN-based recommendation model. Our solution achieved 5$^{th}$ place in the challenge among more than 100 participants. Notably, the strong performance of our RNN approach shows a promising new direction in employing sequence modeling for recommendation systems.