Linghui Meng

Haifeng Wang, Hua Wu, Tian Wu et al.

In this report, we introduce ERNIE 5.0, a natively autoregressive foundation model desinged for unified multimodal understanding and generation across text, image, video, and audio. All modalities are trained from scratch under a unified next-group-of-tokens prediction objective, based on an ultra-sparse mixture-of-experts (MoE) architecture with modality-agnostic expert routing. To address practical challenges in large-scale deployment under diverse resource constraints, ERNIE 5.0 adopts a novel elastic training paradigm. Within a single pre-training run, the model learns a family of sub-models with varying depths, expert capacities, and routing sparsity, enabling flexible trade-offs among performance, model size, and inference latency in memory- or time-constrained scenarios. Moreover, we systematically address the challenges of scaling reinforcement learning to unified foundation models, thereby guaranteeing efficient and stable post-training under ultra-sparse MoE architectures and diverse multimodal settings. Extensive experiments demonstrate that ERNIE 5.0 achieves strong and balanced performance across multiple modalities. To the best of our knowledge, among publicly disclosed models, ERNIE 5.0 represents the first production-scale realization of a trillion-parameter unified autoregressive model that supports both multimodal understanding and generation. To facilitate further research, we present detailed visualizations of modality-agnostic expert routing in the unified model, alongside comprehensive empirical analysis of elastic training, aiming to offer profound insights to the community.

Ziqi Wang, Boye Niu, Zipeng Gao et al.

With the increasing capabilities of Large Language Models (LLMs), parallel reasoning has emerged as a new inference paradigm that enhances reasoning robustness by concurrently exploring multiple lines of thought before converging on a final answer. It has become a significant trend to explore parallel reasoning to overcome the fragility of standard sequential methods and improve practical performance. In this paper, we aim to survey and summarize the progress and challenges of parallel reasoning. We first present a formal definition of parallel reasoning and clarify its distinction from related concepts like Chain-of-Thought. Then, we organize and discuss advanced techniques based on a novel taxonomy, including non-interactive reasoning, interactive reasoning, and efficiency-focused decoding strategies. Additionally, we explore various application scenarios, such as solving complex problems and enhancing the reliability of LLM outputs.Finally, we highlight the core challenges of parallel reasoning and suggest potential directions for future research. We hope that our work can provide a useful roadmap for beginners and encourage more research on improving parallel reasoning methods. Related source can be avaliable in https://github.com/PPPP-kaqiu/Awesome-Parallel-Reasoning.

Linghui Meng, Chun Gan, Shengsheng Niu et al.

Auto-bidding services optimize real-time bidding strategies for advertisers under key performance indicator (KPI) constraints such as target return on investment and budget. However, uncertainties such as model prediction errors and feedback latency can cause bidding strategies to deviate from ex-post optimality, leading to inefficient allocation. To address this issue, we propose JD-BP, a Joint generative Decision framework for Bidding and Pricing. Unlike prior methods, JD-BP jointly outputs a bid value and a pricing correction term that acts additively with the payment rule such as GSP. To mitigate adverse effects of historical constraint violations, we design a memory-less Return-to-Go that encourages future value maximizing of bidding actions while the cumulated bias is handled by the pricing correction. Moreover, a trajectory augmentation algorithm is proposed to generate joint bidding-pricing trajectories from a (possibly arbitrary) base bidding policy, enabling efficient plug-and-play deployment of our algorithm from existing RL/generative bidding models. Finally, we employ an Energy-Based Direct Preference Optimization method in conjunction with a cross-attention module to enhance the joint learning performance of bidding and pricing correction. Offline experiments on the AuctionNet dataset demonstrate that JD-BP achieves state-of-the-art performance. Online A/B tests at JD.com confirm its practical effectiveness, showing a 4.70% increase in ad revenue and a 6.48% improvement in target cost.

Ziqi Wang, Boye Niu, Zhongli Li et al.

Recent Large Reasoning Models have achieved significant improvements in complex task-solving capabilities by allocating more computation at the inference stage with a "thinking longer" paradigm. Even as the foundational reasoning capabilities of models advance rapidly, the persistent gap between a model's performance in a single attempt and its latent potential, often revealed only across multiple solution paths, starkly highlights the disparity between its realized and inherent capabilities. To address this, we present A2R, an Asymmetric Two-Stage Reasoning framework designed to explicitly bridge the gap between a model's potential and its actual performance. In this framework, an "explorer" model first generates potential solutions in parallel through repeated sampling. Subsequently,a "synthesizer" model integrates these references for a more refined, second stage of reasoning. This two-stage process allows computation to be scaled orthogonally to existing sequential methods. Our work makes two key innovations: First, we present A2R as a plug-and-play parallel reasoning framework that explicitly enhances a model's capabilities on complex questions. For example, using our framework, the Qwen3-8B-distill model achieves a 75% performance improvement compared to its self-consistency baseline. Second, through a systematic analysis of the explorer and synthesizer roles, we identify an effective asymmetric scaling paradigm. This insight leads to A2R-Efficient, a "small-to-big" variant that combines a Qwen3-4B explorer with a Qwen3-8B synthesizer. This configuration surpasses the average performance of a monolithic Qwen3-32B model at a nearly 30% lower cost. Collectively, these results show that A2R is not only a performance-boosting framework but also an efficient and practical solution for real-world applications.

Man Hu, Xinyi Wu, Zuofeng Suo et al.

With the rise of advanced reasoning capabilities, large language models (LLMs) are receiving increasing attention. However, although reasoning improves LLMs' performance on downstream tasks, it also introduces new security risks, as adversaries can exploit these capabilities to conduct backdoor attacks. Existing surveys on backdoor attacks and reasoning security offer comprehensive overviews but lack in-depth analysis of backdoor attacks and defenses targeting LLMs' reasoning abilities. In this paper, we take the first step toward providing a comprehensive review of reasoning-based backdoor attacks in LLMs by analyzing their underlying mechanisms, methodological frameworks, and unresolved challenges. Specifically, we introduce a new taxonomy that offers a unified perspective for summarizing existing approaches, categorizing reasoning-based backdoor attacks into associative, passive, and active. We also present defense strategies against such attacks and discuss current challenges alongside potential directions for future research. This work offers a novel perspective, paving the way for further exploration of secure and trustworthy LLM communities.

Ziyi Ni, Minglun Han, Feilong Chen et al.

Enhancing automatic speech recognition (ASR) performance by leveraging additional multimodal information has shown promising results in previous studies. However, most of these works have primarily focused on utilizing visual cues derived from human lip motions. In fact, context-dependent visual and linguistic cues can also benefit in many scenarios. In this paper, we first propose ViLaS (Vision and Language into Automatic Speech Recognition), a novel multimodal ASR model based on the continuous integrate-and-fire (CIF) mechanism, which can integrate visual and textual context simultaneously or separately, to facilitate speech recognition. Next, we introduce an effective training strategy that improves performance in modal-incomplete test scenarios. Then, to explore the effects of integrating vision and language, we create VSDial, a multimodal ASR dataset with multimodal context cues in both Chinese and English versions. Finally, empirical results are reported on the public Flickr8K and self-constructed VSDial datasets. We explore various cross-modal fusion schemes, analyze fine-grained crossmodal alignment on VSDial, and provide insights into the effects of integrating multimodal information on speech recognition.

Xiyun Li, Ziyi Ni, Jingqing Ruan et al.

Adaptive human-agent and agent-agent cooperation are becoming more and more critical in the research area of multi-agent reinforcement learning (MARL), where remarked progress has been made with the help of deep neural networks. However, many established algorithms can only perform well during the learning paradigm but exhibit poor generalization during cooperation with other unseen partners. The personality theory in cognitive psychology describes that humans can well handle the above cooperation challenge by predicting others' personalities first and then their complex actions. Inspired by this two-step psychology theory, we propose a biologically plausible mixture of personality (MoP) improved spiking actor network (SAN), whereby a determinantal point process is used to simulate the complex formation and integration of different types of personality in MoP, and dynamic and spiking neurons are incorporated into the SAN for the efficient reinforcement learning. The benchmark Overcooked task, containing a strong requirement for cooperative cooking, is selected to test the proposed MoP-SAN. The experimental results show that the MoP-SAN can achieve both high performances during not only the learning paradigm but also the generalization test (i.e., cooperation with other unseen agents) paradigm where most counterpart deep actor networks failed. Necessary ablation experiments and visualization analyses were conducted to explain why MoP and SAN are effective in multi-agent reinforcement learning scenarios while DNN performs poorly in the generalization test.

Linghui Meng, Muning Wen, Yaodong Yang et al.

Offline reinforcement learning leverages previously-collected offline datasets to learn optimal policies with no necessity to access the real environment. Such a paradigm is also desirable for multi-agent reinforcement learning (MARL) tasks, given the increased interactions among agents and with the enviroment. Yet, in MARL, the paradigm of offline pre-training with online fine-tuning has not been studied, nor datasets or benchmarks for offline MARL research are available. In this paper, we facilitate the research by providing large-scale datasets, and use them to examine the usage of the Decision Transformer in the context of MARL. We investigate the generalisation of MARL offline pre-training in the following three aspects: 1) between single agents and multiple agents, 2) from offline pretraining to the online fine-tuning, and 3) to that of multiple downstream tasks with few-shot and zero-shot capabilities. We start by introducing the first offline MARL dataset with diverse quality levels based on the StarCraftII environment, and then propose the novel architecture of multi-agent decision transformer (MADT) for effective offline learning. MADT leverages transformer's modelling ability of sequence modelling and integrates it seamlessly with both offline and online MARL tasks. A crucial benefit of MADT is that it learns generalisable policies that can transfer between different types of agents under different task scenarios. On StarCraft II offline dataset, MADT outperforms the state-of-the-art offline RL baselines. When applied to online tasks, the pre-trained MADT significantly improves sample efficiency, and enjoys strong performance both few-short and zero-shot cases. To our best knowledge, this is the first work that studies and demonstrates the effectiveness of offline pre-trained models in terms of sample efficiency and generalisability enhancements in MARL.

Jakub Grudzien Kuba, Muning Wen, Yaodong Yang et al.

Policy gradient (PG) methods are popular reinforcement learning (RL) methods where a baseline is often applied to reduce the variance of gradient estimates. In multi-agent RL (MARL), although the PG theorem can be naturally extended, the effectiveness of multi-agent PG (MAPG) methods degrades as the variance of gradient estimates increases rapidly with the number of agents. In this paper, we offer a rigorous analysis of MAPG methods by, firstly, quantifying the contributions of the number of agents and agents' explorations to the variance of MAPG estimators. Based on this analysis, we derive the optimal baseline (OB) that achieves the minimal variance. In comparison to the OB, we measure the excess variance of existing MARL algorithms such as vanilla MAPG and COMA. Considering using deep neural networks, we also propose a surrogate version of OB, which can be seamlessly plugged into any existing PG methods in MARL. On benchmarks of Multi-Agent MuJoCo and StarCraft challenges, our OB technique effectively stabilises training and improves the performance of multi-agent PPO and COMA algorithms by a significant margin.

Linghui Meng, Jin Xu, Xu Tan et al.

In this paper, we propose MixSpeech, a simple yet effective data augmentation method based on mixup for automatic speech recognition (ASR). MixSpeech trains an ASR model by taking a weighted combination of two different speech features (e.g., mel-spectrograms or MFCC) as the input, and recognizing both text sequences, where the two recognition losses use the same combination weight. We apply MixSpeech on two popular end-to-end speech recognition models including LAS (Listen, Attend and Spell) and Transformer, and conduct experiments on several low-resource datasets including TIMIT, WSJ, and HKUST. Experimental results show that MixSpeech achieves better accuracy than the baseline models without data augmentation, and outperforms a strong data augmentation method SpecAugment on these recognition tasks. Specifically, MixSpeech outperforms SpecAugment with a relative PER improvement of 10.6$\%$ on TIMIT dataset, and achieves a strong WER of 4.7$\%$ on WSJ dataset.