Xuewei Wang

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Ming Zhong, Aston Zhang, Xuewei Wang et al. · meta-ai

The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term cross capabilities. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce CrossEval, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that, in both static evaluations and attempts to enhance specific abilities, current LLMs consistently exhibit the "Law of the Weakest Link," where cross-capability performance is significantly constrained by the weakest component. Specifically, across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight the under-performance of LLMs in cross-capability tasks, making the identification and improvement of the weakest capabilities a critical priority for future research to optimize performance in complex, multi-dimensional scenarios.

Xuewei Wang, Qiang Jin, Shengyu Huang et al.

Dividing ads ranking system into retrieval, early, and final stages is a common practice in large scale ads recommendation to balance the efficiency and accuracy. The early stage ranking often uses efficient models to generate candidates out of a set of retrieved ads. The candidates are then fed into a more computationally intensive but accurate final stage ranking system to produce the final ads recommendation. As the early and final stage ranking use different features and model architectures because of system constraints, a serious ranking consistency issue arises where the early stage has a low ads recall, i.e., top ads in the final stage are ranked low in the early stage. In order to pass better ads from the early to the final stage ranking, we propose a multi-task learning framework for early stage ranking to capture multiple final stage ranking components (i.e. ads clicks and ads quality events) and their task relations. With our multi-task learning framework, we can not only achieve serving cost saving from the model consolidation, but also improve the ads recall and ranking consistency. In the online A/B testing, our framework achieves significantly higher click-through rate (CTR), conversion rate (CVR), total value and better ads-quality (e.g. reduced ads cross-out rate) in a large scale industrial ads ranking system.

Wei Zhang, Dai Li, Chen Liang et al.

Effective user representations are pivotal in personalized advertising. However, stringent constraints on training throughput, serving latency, and memory, often limit the complexity and input feature set of online ads ranking models. This challenge is magnified in extensive systems like Meta's, which encompass hundreds of models with diverse specifications, rendering the tailoring of user representation learning for each model impractical. To address these challenges, we present Scaling User Modeling (SUM), a framework widely deployed in Meta's ads ranking system, designed to facilitate efficient and scalable sharing of online user representation across hundreds of ads models. SUM leverages a few designated upstream user models to synthesize user embeddings from massive amounts of user features with advanced modeling techniques. These embeddings then serve as inputs to downstream online ads ranking models, promoting efficient representation sharing. To adapt to the dynamic nature of user features and ensure embedding freshness, we designed SUM Online Asynchronous Platform (SOAP), a latency free online serving system complemented with model freshness and embedding stabilization, which enables frequent user model updates and online inference of user embeddings upon each user request. We share our hands-on deployment experiences for the SUM framework and validate its superiority through comprehensive experiments. To date, SUM has been launched to hundreds of ads ranking models in Meta, processing hundreds of billions of user requests daily, yielding significant online metric gains and improved infrastructure efficiency.

Yichao Fu, Xuewei Wang, Yuandong Tian et al.

Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional model training or hyperparameter tuning and can be seamlessly integrated into existing serving frameworks. We evaluate DeepConf across a variety of reasoning tasks and the latest open-source models, including Qwen 3 and GPT-OSS series. Notably, on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.

Yixin Liu, Yue Yu, DiJia Su et al.

Reasoning LLMs-as-Judges, which can benefit from inference-time scaling, provide a promising path for extending the success of reasoning models to non-verifiable domains where the output correctness/quality cannot be directly checked. However, while reasoning judges have shown better performance on static evaluation benchmarks, their effectiveness in actual policy training has not been systematically examined. Therefore, we conduct a rigorous study to investigate the actual impact of non-reasoning and reasoning judges in reinforcement-learning-based LLM alignment. Our controlled synthetic setting, where a "gold-standard" judge (gpt-oss-120b) provides preference annotations to train smaller judges, reveals key differences between non-reasoning and reasoning judges: non-reasoning judges lead to reward hacking easily, while reasoning judges can lead to policies that achieve strong performance when evaluated by the gold-standard judge. Interestingly, we find that the reasoning-judge-trained policies achieve such strong performance by learning to generate highly effective adversarial outputs that can also score well on popular benchmarks such as Arena-Hard by deceiving other LLM-judges. Combined with our further analysis, our study highlights both important findings and room for improvements for applying (reasoning) LLM-judges in non-verifiable LLM post-training.

Chenglong Lu, Shen Liang, Xuewei Wang et al.

Vision Transformers (ViTs) have computational costs scaling quadratically with the number of tokens, calling for effective token pruning policies. Most existing policies are handcrafted, lacking adaptivity to varying inputs. Moreover, they fail to consider the sequential nature of token pruning across multiple layers. In this work, for the first time (as far as we know), we exploit Reinforcement Learning (RL) to data-adaptively learn a pruning policy. Formulating token pruning as a sequential decision-making problem, we model it as a Markov Game and utilize Multi-Agent Proximal Policy Optimization (MAPPO) where each agent makes an individualized pruning decision for a single token. We also develop reward functions that enable simultaneous collaboration and competition of these agents to balance efficiency and accuracy. On the well-known ImageNet-1k dataset, our method improves the inference speed by up to 44% while incurring only a negligible accuracy drop of 0.4%. The source code is available at https://github.com/daashuai/rl4evit.

Yue Yu, Zhengxing Chen, Aston Zhang et al. · allen-ai

Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy.

Weiyan Shi, Xuewei Wang, Yoo Jung Oh et al.

Intelligent conversational agents, or chatbots, can take on various identities and are increasingly engaging in more human-centered conversations with persuasive goals. However, little is known about how identities and inquiry strategies influence the conversation's effectiveness. We conducted an online study involving 790 participants to be persuaded by a chatbot for charity donation. We designed a two by four factorial experiment (two chatbot identities and four inquiry strategies) where participants were randomly assigned to different conditions. Findings showed that the perceived identity of the chatbot had significant effects on the persuasion outcome (i.e., donation) and interpersonal perceptions (i.e., competence, confidence, warmth, and sincerity). Further, we identified interaction effects among perceived identities and inquiry strategies. We discuss the findings for theoretical and practical implications for developing ethical and effective persuasive chatbots. Our published data, codes, and analyses serve as the first step towards building competent ethical persuasive chatbots.

Weiyan Shi, Kun Qian, Xuewei Wang et al.

User simulators are essential for training reinforcement learning (RL) based dialog models. The performance of the simulator directly impacts the RL policy. However, building a good user simulator that models real user behaviors is challenging. We propose a method of standardizing user simulator building that can be used by the community to compare dialog system quality using the same set of user simulators fairly. We present implementations of six user simulators trained with different dialog planning and generation methods. We then calculate a set of automatic metrics to evaluate the quality of these simulators both directly and indirectly. We also ask human users to assess the simulators directly and indirectly by rating the simulated dialogs and interacting with the trained systems. This paper presents a comprehensive evaluation framework for user simulator study and provides a better understanding of the pros and cons of different user simulators, as well as their impacts on the trained systems.

Xuewei Wang, Weiyan Shi, Richard Kim et al.

Developing intelligent persuasive conversational agents to change people's opinions and actions for social good is the frontier in advancing the ethical development of automated dialogue systems. To do so, the first step is to understand the intricate organization of strategic disclosures and appeals employed in human persuasion conversations. We designed an online persuasion task where one participant was asked to persuade the other to donate to a specific charity. We collected a large dataset with 1,017 dialogues and annotated emerging persuasion strategies from a subset. Based on the annotation, we built a baseline classifier with context information and sentence-level features to predict the 10 persuasion strategies used in the corpus. Furthermore, to develop an understanding of personalized persuasion processes, we analyzed the relationships between individuals' demographic and psychological backgrounds including personality, morality, value systems, and their willingness for donation. Then, we analyzed which types of persuasion strategies led to a greater amount of donation depending on the individuals' personal backgrounds. This work lays the ground for developing a personalized persuasive dialogue system.