The Llama 4 Herd: Architecture, Training, Evaluation, and Deployment Notes

This document consolidates publicly reported technical details about Metas Llama 4 model family. It summarizes (i) released variants (Scout and Maverick) and the broader herd context including the previewed Behemoth teacher model, (ii) architectural characteristics beyond a high-level MoE description covering routed/shared-expert structure, early-fusion multimodality, and long-context design elements reported for Scout (iRoPE and length generalization strategies), (iii) training disclosures spanning pre-training, mid-training for long-context extension, and post-training methodology (lightweight SFT, online RL, and lightweight DPO) as described in release materials, (iv) developer-reported benchmark results for both base and instruction-tuned checkpoints, and (v) practical deployment constraints observed across major serving environments, including provider-specific context limits and quantization packaging. The manuscript also summarizes licensing obligations relevant to redistribution and derivative naming, and reviews publicly described safeguards and evaluation practices. The goal is to provide a compact technical reference for researchers and practitioners who need precise, source-backed facts about Llama 4.

Q-Regularized Generative Auto-Bidding: From Suboptimal Trajectories to Optimal Policies

With the rapid development of e-commerce, auto-bidding has become a key asset in optimizing advertising performance under diverse advertiser environments. The current approaches focus on reinforcement learning (RL) and generative models. These efforts imitate offline historical behaviors by utilizing a complex structure with expensive hyperparameter tuning. The suboptimal trajectories further exacerbate the difficulty of policy learning. To address these challenges, we proposes QGA, a novel Q-value regularized Generative Auto-bidding method. In QGA, we propose to plug a Q-value regularization with double Q-learning strategy into the Decision Transformer backbone. This design enables joint optimization of policy imitation and action-value maximization, allowing the learned bidding policy to both leverage experience from the dataset and alleviate the adverse impact of the suboptimal trajectories. Furthermore, to safely explore the policy space beyond the data distribution, we propose a Q-value guided dual-exploration mechanism, in which the DT model is conditioned on multiple return-to-go targets and locally perturbed actions. This entire exploration process is dynamically guided by the aforementioned Q-value module, which provides principled evaluation for each candidate action. Experiments on public benchmarks and simulation environments demonstrate that QGA consistently achieves superior or highly competitive results compared to existing alternatives. Notably, in large-scale real-world A/B testing, QGA achieves a 3.27% increase in Ad GMV and a 2.49% improvement in Ad ROI.

Adaptive Diffusion-based Augmentation for Recommendation

Recommendation systems often rely on implicit feedback, where only positive user-item interactions can be observed. Negative sampling is therefore crucial to provide proper negative training signals. However, existing methods tend to mislabel potentially positive but unobserved items as negatives and lack precise control over negative sample selection. We aim to address these by generating controllable negative samples, rather than sampling from the existing item pool. In this context, we propose Adaptive Diffusion-based Augmentation for Recommendation (ADAR), a novel and model-agnostic module that leverages diffusion to synthesize informative negatives. Inspired by the progressive corruption process in diffusion, ADAR simulates a continuous transition from positive to negative, allowing for fine-grained control over sample hardness. To mine suitable negative samples, we theoretically identify the transition point at which a positive sample turns negative and derive a score-aware function to adaptively determine the optimal sampling timestep. By identifying this transition point, ADAR generates challenging negative samples that effectively refine the model's decision boundary. Experiments confirm that ADAR is broadly compatible and boosts the performance of existing recommendation models substantially, including collaborative filtering and sequential recommendation, without architectural modifications.

DiffKD-DCIS: Predicting Upgrade of Ductal Carcinoma In Situ with Diffusion Augmentation and Knowledge Distillation

Accurately predicting the upgrade of ductal carcinoma in situ (DCIS) to invasive ductal carcinoma (IDC) is crucial for surgical planning. However, traditional deep learning methods face challenges due to limited ultrasound data and poor generalization ability. This study proposes the DiffKD-DCIS framework, integrating conditional diffusion modeling with teacher-student knowledge distillation. The framework operates in three stages: First, a conditional diffusion model generates high-fidelity ultrasound images using multimodal conditions for data augmentation. Then, a deep teacher network extracts robust features from both original and synthetic data. Finally, a compact student network learns from the teacher via knowledge distillation, balancing generalization and computational efficiency. Evaluated on a multi-center dataset of 1,435 cases, the synthetic images were of good quality. The student network had fewer parameters and faster inference. On external test sets, it outperformed partial combinations, and its accuracy was comparable to senior radiologists and superior to junior ones, showing significant clinical potential.

Are First-Order Diffusion Samplers Really Slower? A Fast Forward-Value Approach

Higher-order ODE solvers have become a standard tool for accelerating diffusion probabilistic model (DPM) sampling, motivating the widespread view that first-order methods are inherently slower and that increasing discretization order is the primary path to faster generation. This paper challenges this belief and revisits acceleration from a complementary angle: beyond solver order, the placement of DPM evaluations along the reverse-time dynamics can substantially affect sampling accuracy in the low-neural function evaluation (NFE) regime. We propose a novel training-free, first-order sampler whose leading discretization error has the opposite sign to that of DDIM. Algorithmically, the method approximates the forward-value evaluation via a cheap one-step lookahead predictor. We provide theoretical guarantees showing that the resulting sampler provably approximates the ideal forward-value trajectory while retaining first-order convergence. Empirically, across standard image generation benchmarks (CIFAR-10, ImageNet, FFHQ, and LSUN), the proposed sampler consistently improves sample quality under the same NFE budget and can be competitive with, and sometimes outperform, state-of-the-art higher-order samplers. Overall, the results suggest that the placement of DPM evaluations provides an additional and largely independent design angle for accelerating diffusion sampling.

Max-Entropy Reinforcement Learning with Flow Matching and A Case Study on LQR

Soft actor-critic (SAC) is a popular algorithm for max-entropy reinforcement learning. In practice, the energy-based policies in SAC are often approximated using simple policy classes for efficiency, sacrificing the expressiveness and robustness. In this paper, we propose a variant of the SAC algorithm that parameterizes the policy with flow-based models, leveraging their rich expressiveness. In the algorithm, we evaluate the flow-based policy utilizing the instantaneous change-of-variable technique and update the policy with an online variant of flow matching developed in this paper. This online variant, termed importance sampling flow matching (ISFM), enables policy update with only samples from a user-specified sampling distribution rather than the unknown target distribution. We develop a theoretical analysis of ISFM, characterizing how different choices of sampling distributions affect the learning efficiency. Finally, we conduct a case study of our algorithm on the max-entropy linear quadratic regulator problems, demonstrating that the proposed algorithm learns the optimal action distribution.

Spectral Representation-based Reinforcement Learning

In real-world applications with large state and action spaces, reinforcement learning (RL) typically employs function approximations to represent core components like the policies, value functions, and dynamics models. Although powerful approximations such as neural networks offer great expressiveness, they often present theoretical ambiguities, suffer from optimization instability and exploration difficulty, and incur substantial computational costs in practice. In this paper, we introduce the perspective of spectral representations as a solution to address these difficulties in RL. Stemming from the spectral decomposition of the transition operator, this framework yields an effective abstraction of the system dynamics for subsequent policy optimization while also providing a clear theoretical characterization. We reveal how to construct spectral representations for transition operators that possess latent variable structures or energy-based structures, which implies different learning methods to extract spectral representations from data. Notably, each of these learning methods realizes an effective RL algorithm under this framework. We also provably extend this spectral view to partially observable MDPs. Finally, we validate these algorithms on over 20 challenging tasks from the DeepMind Control Suite, where they achieve performances comparable or superior to current state-of-the-art model-free and model-based baselines.

Model-Based Diffusion Sampling for Predictive Control in Offline Decision Making

Offline decision-making requires synthesizing reliable behaviors from fixed datasets without further interaction, yet existing generative approaches often yield trajectories that are dynamically infeasible. We propose Model Predictive Diffuser (MPDiffuser), a compositional model-based diffusion framework consisting of: (i) a planner that generates diverse, task-aligned trajectories; (ii) a dynamics model that enforces consistency with the underlying system dynamics; and (iii) a ranker module that selects behaviors aligned with the task objectives. MPDiffuser employs an alternating diffusion sampling scheme, where planner and dynamics updates are interleaved to progressively refine trajectories for both task alignment and feasibility during the sampling process. We also provide a theoretical rationale for this procedure, showing how it balances fidelity to data priors with dynamics consistency. Empirically, the compositional design improves sample efficiency, as it leverages even low-quality data for dynamics learning and adapts seamlessly to novel dynamics. We evaluate MPDiffuser on both unconstrained (D4RL) and constrained (DSRL) offline decision-making benchmarks, demonstrating consistent gains over existing approaches. Furthermore, we present a preliminary study extending MPDiffuser to vision-based control tasks, showing its potential to scale to high-dimensional sensory inputs. Finally, we deploy our method on a real quadrupedal robot, showcasing its practicality for real-world control.

Offline Imitation Learning upon Arbitrary Demonstrations by Pre-Training Dynamics Representations

Limited data has become a major bottleneck in scaling up offline imitation learning (IL). In this paper, we propose enhancing IL performance under limited expert data by introducing a pre-training stage that learns dynamics representations, derived from factorizations of the transition dynamics. We first theoretically justify that the optimal decision variable of offline IL lies in the representation space, significantly reducing the parameters to learn in the downstream IL. Moreover, the dynamics representations can be learned from arbitrary data collected with the same dynamics, allowing the reuse of massive non-expert data and mitigating the limited data issues. We present a tractable loss function inspired by noise contrastive estimation to learn the dynamics representations at the pre-training stage. Experiments on MuJoCo demonstrate that our proposed algorithm can mimic expert policies with as few as a single trajectory. Experiments on real quadrupeds show that we can leverage pre-trained dynamics representations from simulator data to learn to walk from a few real-world demonstrations.

One-Step Flow Policy Mirror Descent

Diffusion policies have achieved great success in online reinforcement learning (RL) due to their strong expressive capacity. However, the inference of diffusion policy models relies on a slow iterative sampling process, which limits their responsiveness. To overcome this limitation, we propose Flow Policy Mirror Descent (FPMD), an online RL algorithm that enables 1-step sampling during policy inference. Our approach exploits a theoretical connection between the distribution variance and the discretization error of single-step sampling in straight interpolation flow matching models, and requires no extra distillation or consistency training. We present two algorithm variants based on flow policy and MeanFlow policy parametrizations, respectively. Extensive empirical evaluations on MuJoCo benchmarks demonstrate that our algorithms show strong performance comparable to diffusion policy baselines while requiring hundreds of times fewer function evaluations during inference.