Zhihui Xie

Zhihui Xie, Jiacheng Ye, Lin Zheng et al.

We present Dream-Coder 7B, an open-source discrete diffusion language model for code generation that exhibits emergent any-order generation capabilities. Unlike traditional autoregressive (AR) models that decode strictly left-to-right, Dream-Coder 7B adaptively determines its decoding strategy based on the coding task: sketch-first generation for complex algorithms, left-to-right generation for straightforward completions, and interleaved reasoning generation for code understanding tasks. We adapt a pretrained AR checkpoint to a discrete diffusion frameworks with a continuous-time weighted cross-entropy objective. Our post-training recipe comprises (i) supervised fine-tuning, where we mitigate padding pathologies via random truncation and a padding penalty to improve sample efficiency and stabilize generation; and (ii) reinforcement learning with verifiable rewards over a curated high-quality prompt set drawn from open-source datasets, using a tailored reinforcement learning recipe for diffusion language models. The resulting Dream-Coder 7B Instruct attains 21.4\% pass@1 on LiveCodeBench (2410--2505) and demonstrates competitive performance on HumanEval, MBPP, BigCodeBench, and CRUXEval. We release Dream-Coder-7B and Dream-Coder-7B-Instruct checkpoints, training recipes, preprocessing pipelines, and inference code to facilitate reproducibility and further research.

Jiacheng Ye, Zhihui Xie, Lin Zheng et al.

We introduce Dream 7B, the most powerful open diffusion large language model to date. Unlike autoregressive (AR) models that generate tokens sequentially, Dream 7B employs discrete diffusion modeling to refine sequences in parallel through iterative denoising. Our model consistently outperforms existing diffusion language models on general, mathematical, and coding tasks. Dream 7B demonstrates superior planning abilities and inference flexibility, including arbitrary-order generation, infilling capabilities, and tunable quality-speed trade-offs. These results are achieved through simple yet effective training techniques, including AR-based LLM initialization and context-adaptive token-level noise rescheduling. We release both Dream-Base and Dream-Instruct to facilitate further research in diffusion-based language modeling.

Lei Li, Zhihui Xie, Mukai Li et al. · pku

As large vision-language models (LVLMs) evolve rapidly, the demand for high-quality and diverse data to align these models becomes increasingly crucial. However, the creation of such data with human supervision proves costly and time-intensive. In this paper, we investigate the efficacy of AI feedback to scale supervision for aligning LVLMs. We introduce VLFeedback, the first large-scale vision-language feedback dataset, comprising over 82K multi-modal instructions and comprehensive rationales generated by off-the-shelf models without human annotations. To evaluate the effectiveness of AI feedback for vision-language alignment, we train Silkie, an LVLM fine-tuned via direct preference optimization on VLFeedback. Silkie showcases exceptional performance regarding helpfulness, visual faithfulness, and safety metrics. It outperforms its base model by 6.9\% and 9.5\% in perception and cognition tasks, reduces hallucination issues on MMHal-Bench, and exhibits enhanced resilience against red-teaming attacks. Furthermore, our analysis underscores the advantage of AI feedback, particularly in fostering preference diversity to deliver more comprehensive improvements. Our dataset, training code and models are available at https://vlf-silkie.github.io.

Runze Liu, Jiakang Wang, Yuling Shi et al.

Reinforcement Learning (RL) has shown remarkable success in enhancing the reasoning capabilities of Large Language Models (LLMs). Process-Supervised RL (PSRL) has emerged as a more effective paradigm compared to outcome-based RL. However, existing PSRL approaches suffer from limited exploration efficiency, both in terms of branching positions and sampling. In this paper, we introduce a novel PSRL framework (AttnRL), which enables efficient exploration for reasoning models. Motivated by preliminary observations that steps exhibiting high attention scores correlate with reasoning behaviors, we propose to branch from positions with high values. Furthermore, we develop an adaptive sampling strategy that accounts for problem difficulty and historical batch size, ensuring that the whole training batch maintains non-zero advantage values. To further improve sampling efficiency, we design a one-step off-policy training pipeline for PSRL. Extensive experiments on multiple challenging mathematical reasoning benchmarks demonstrate that our method consistently outperforms prior approaches in terms of performance and sampling and training efficiency.

Gang Bai, Zhihui Xie, Liang Wang

Sponsored search in E-commerce platforms such as Amazon, Taobao and Tmall provides sellers an effective way to reach potential buyers with most relevant purpose. In this paper, we study the auction mechanism optimization problem in sponsored search on Alibaba's mobile E-commerce platform. Besides generating revenue, we are supposed to maintain an efficient marketplace with plenty of quality users, guarantee a reasonable return on investment (ROI) for advertisers, and meanwhile, facilitate a pleasant shopping experience for the users. These requirements essentially pose a constrained optimization problem. Directly optimizing over auction parameters yields a discontinuous, non-convex problem that denies effective solutions. One of our major contribution is a practical convex optimization formulation of the original problem. We devise a novel re-parametrization of auction mechanism with discrete sets of representative instances. To construct the optimization problem, we build an auction simulation system which estimates the resulted business indicators of the selected parameters by replaying the auctions recorded from real online requests. We summarized the experiments on real search traffics to analyze the effects of fidelity of auction simulation, the efficacy under various constraint targets and the influence of regularization. The experiment results show that with proper entropy regularization, we are able to maximize revenue while constraining other business indicators within given ranges.

Zhihui Xie, Kuang-Chih Lee, Liang Wang

The online ads trading platform plays a crucial role in connecting publishers and advertisers and generates tremendous value in facilitating the convenience of our lives. It has been evolving into a more and more complicated structure. In this paper, we consider the problem of maximizing the revenue for the seller side via utilizing proper reserve price for the auctions in a dynamical way. Predicting the optimal reserve price for each auction in the repeated auction marketplaces is a non-trivial problem. However, we were able to come up with an efficient method of improving the seller revenue by mainly focusing on adjusting the reserve price for those high-value inventories. Previously, no dedicated work has been performed from this perspective. Inspired by Paul and Michael, our model first identifies the value of the inventory by predicting the top bid price bucket using a cascade of classifiers. The cascade is essential in significantly reducing the false positive rate of a single classifier. Based on the output of the first step, we build another cluster of classifiers to predict the price separations between the top two bids. We showed that although the high-value auctions are only a small portion of all the traffic, successfully identifying them and setting correct reserve price would result in a significant revenue lift. Moreover, our optimization is compatible with all other reserve price models in the system and does not impact their performance. In other words, when combined with other models, the enhancement on exchange revenue will be aggregated. Simulations on randomly sampled Yahoo ads exchange (YAXR) data showed stable and expected lift after applying our model.