Large Language Models (LLMs) inference is typically deployed under a static resource assumption, where models execute a fixed computational graph regardless of the runtime environment. However, real-world cloud infrastructure is inherently dynamic, characterized by fluctuating availability (e.g., spot instance preemption) and tiered Quality-of-Service requirements. In such volatile settings, static models are inflexible: they either crash under resource constraints or waste compute on redundant operations. To bridge this gap, we propose Learning to Allocate (L2A), an end-to-end framework for resource-adaptive inference. Unlike prior methods that condition only on input difficulty, we formulate inference as a constrained allocation problem conditioned on both the input and the runtime resource budget itself. We introduce lightweight, budget-conditioned and input-aware gating networks integrated into the LLM. These gates are trained via a unified objective that jointly optimizes task performance, logical consistency, and resource costs along three axes matching how real-world dynamics manifest: layer skipping for memory and depth pressure, head pruning for throughput contention, and reasoning-token reduction for latency tightening. This lets the model learn a budget-aware policy beyond input difficulty alone: it adaptively configures its computational footprint with respect to real-time resource dynamics, maximizing reasoning depth when resources permit while enforcing strict frugality when budgets tighten. A single L2A model traces the entire compute-accuracy Pareto frontier on Llama-3-8B and Qwen-3-4B: at up to 34

Chain-of-thought (CoT) distillation in the recommendation domain is a necessary precursor to RL training, but raw teacher traces are ill-suited to this task. Large teachers approach the recommendation task with unusually high reasoning uncertainty, repeatedly rechecking their answers without revising them; supervised fine-tuning on such traces produces verbose students that never revise their initial guess. Furthermore, due to the novelty of the recommendation domain, the teacher's reasoning traces are highly out-of-distribution for the small student LLM. We propose Student-Aware CoT Optimization for Recommendation Distillation (SCOReD), a CoT optimization framework tailored to recommendation that first parses each teacher trace into typed segments and uses the student LLM's attention to score the importance of each segment. Then SCOReD dynamically selects a per-segment edit (KEEP / REWRITE / FUSE / PRUNE) based on the output length and comparative log probability lift of the answer given the edit as per the student. Therefore, SCOReD prunes redundant sections of the reasoning trace while preserving information-dense sections and adapts raw teacher traces to the student's output distribution. Training on SCOReD-optimized CoTs provides a cleaner learning signal to the student model and improves over baseline SFT by 1.56

The rapidly evolving landscape of products, surfaces, policies, and regulations poses significant challenges for deploying state-of-the-art recommendation models at industry scale, primarily due to data fragmentation across domains and escalating infrastructure costs that hinder sustained quality improvements. To address this challenge, we propose Lattice, a recommendation framework centered around model space redesign that extends Multi-Domain, Multi-Objective (MDMO) learning beyond models and learning objectives. Lattice addresses these challenges through a comprehensive model space redesign that combines cross-domain knowledge sharing, data consolidation, model unification, distillation, and system optimizations to achieve significant improvements in both quality and cost-efficiency. Our deployment of Lattice at Meta has resulted in 10% revenue-driving top-line metrics gain, 11.5% user satisfaction improvement, 6% boost in conversion rate, with 20% capacity saving.

Discrete diffusion language models (dLLMs) recover masked tokens in parallel, offering significant speedups over autoregressive (AR) generation. However, such promising frameworks face a fundamental architectural design dilemma: 182 Adopting bidirectional attention achieves strong generation quality by allowing each position to access the full context, but is inherently incompatible with KV caching, limiting inference throughput in batch-serving scenarios; 183 Conversely, causal attention enables efficient cached inference but loses all right-side context, substantially degrading generation quality. This paper introduces Bifocal dLLMs, a new paradigm that resolves this dilemma through asymmetric bidirectional context. Analogous to bifocal lenses, we instantiate the paradigm as R2LM (Right-to-Left Mamba), which combines two complementary mechanisms: a) standard causal attention providing precise left-context with full KV cache compatibility, while b) a lightweight reverse Mamba SSM sidecar supplying compressed right-side context without breaking cacheability. Comprehensive experiments on continued pretraining of Qwen3-1.7B with 60B tokens demonstrate that R2LM achieves 2.4× to 12.9× higher throughput than bidirectional dLLMs and 1.9× to 2.9× speedup over AR baselines in batch serving through parallel decoding with KV caching, while exceeding the causal baseline on most benchmarks and surpassing the bidirectional dLLM on average.

Modern recommendation models gain prediction quality by scaling feature-interaction and sequence modules, but production cost constraints cap how far systems can scale. In this work, we propose Request-Oriented Compute Sharing (ROCS), a modeling and inference paradigm that exploits a unique property of recommendation inference: each user request is evaluated against many candidates, while request-side features are shared across candidates. ROCS defers request-candidate interactions as late as possible, isolates candidate-dependent representations, and evaluates substantial portions of the model once per request rather than once per candidate, significantly improving inference efficiency while maintaining or improving prediction quality. To realize this paradigm, we develop Generalized Layer Masking (GLM) to enforce candidate isolation in feature-interaction architectures, and Deep Cross Attention (DCA) to extend request-oriented sharing to sequence architectures. To support efficient GPU deployment, we co-design In-Kernel Broadcast Optimization (IKBO) that significantly accelerates ROCS model execution. Experiments on public benchmarks show that ROCS consistently improves the quality-efficiency tradeoff across recommendation backbones. On production-scale workloads, ROCS achieves up to a 3x QPS improvement on retrieval models without quality degradation and a 0.5

Recent GPU generations deliver significantly higher FLOPs using lower-precision arithmetic, such as FP8. While successfully applied to large language models (LLMs), its adoption in large recommendation models (LRMs) has been limited. This is because LRMs are numerically sensitive, dominated by small matrix multiplications (GEMMs) followed by normalization, and trained in communication-intensive environments. Applying FP8 directly to LRMs often degrades model quality and prolongs training time. These challenges are inherent to LRM workloads and cannot be resolved merely by introducing better FP8 kernels. Instead, a system-model co-design approach is needed to successfully integrate FP8. We present LoKA (Low-precision Kernel Applications), a framework that makes FP8 practical for LRMs through three principles: profile under realistic distributions to know where low precision is safe, co-design model components with hardware to expand where it is safe, and orchestrate across kernel libraries to maximize the gains. Concretely, LoKA Probe is a statistically grounded, online benchmarking method that learns activation and weight statistics, and quantifies per-layer errors. This process pinpoints safe and unsafe, fast and slow sites for FP8 adoption. LoKA Mods is a set of reusable model adaptations that improve both numerical stability and execution efficiency with FP8. LoKA Dispatch is a runtime that leverages the statistical insights from LoKA Probe to select the fastest FP8 kernel that satisfies the accuracy requirements.

Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the evidence most relevant to a question. A promising way to improve long-context utilization is test-time training (TTT), which treats the test context as a training example for instance-specific parameter adaptation. However, applying TTT to the entire long context is prohibitively expensive, while adapting on randomly sampled spans introduces severe noise. Because most spans in a long context are irrelevant to the specific question, training on them may even degrade the base model's performance. Our preliminary study shows that TTT is highly sensitive to training-span quality: on LongBench-v2, TTT on randomly sampled spans hurts performance, whereas TTT on oracle spans substantially improves it. Motivated by this, we propose a simple method, Self-Guided TTT (S-TTT): before adaptation, the model identifies the evidence spans it should learn from, and the standard language-modeling training objective is applied only to those selected spans. On two challenging long-context reasoning benchmarks, LongBench-v2 and LongBench-Pro, S-TTT improves accuracy for both Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, achieving up to a 15

In machine learning, the goal of multi-task learning (MTL) is to optimize multiple objectives together. Recent works, for example, Multiple Gradient Descent Algorithm (MGDA) and its variants, show promising results with dynamically adjusted weights for different tasks to mitigate conflicts that may potentially degrade the performance on certain tasks. Despite the empirical success of MGDA-type methods, one major limitation of such methods is their computational inefficiency, as they require access to all task gradients. In this paper we introduce MARIGOLD, a unified algorithmic framework for efficiently solving MTL problems. Our method reveals that multi-task gradient balancing methods have a hierarchical structure, in which the model training and the gradient balancing are coupled during the whole optimization process and can be viewed as a bi-level optimization problem. Moreover, we showcase that the bi-level problem can be solved efficiently by leveraging zeroth-order method. Extensive experiments on both public datasets and industrial-scale datasets demonstrate the efficiency and superiority of our method.

Recent advances in recommendation scaling laws have led to foundation models of unprecedented complexity. While these models offer superior performance, their computational demands make real-time serving impractical, often forcing practitioners to rely on knowledge distillation-compromising serving quality for efficiency. To address this challenge, we present SOLARIS (Speculative Offloading of Latent-bAsed Representation for Inference Scaling), a novel framework inspired by speculative decoding. SOLARIS proactively precomputes user-item interaction embeddings by predicting which user-item pairs are likely to appear in future requests, and asynchronously generating their foundation model representations ahead of time. This approach decouples the costly foundation model inference from the latency-critical serving path, enabling real-time knowledge transfer from models previously considered too expensive for online use. Deployed across Meta's advertising system serving billions of daily requests, SOLARIS achieves 0.67

Large Language Models (LLMs) can reason well, yet often miss decisive evidence when it is buried in long, noisy contexts. We introduce HiLight, an Evidence Emphasis framework that decouples evidence selection from reasoning for frozen LLM solvers. HiLight avoids compressing or rewriting the input, which can discard or distort evidence, by training a lightweight Emphasis Actor to insert minimal highlight tags around pivotal spans in the unaltered context. A frozen Solver then performs downstream reasoning on the emphasized input. We cast highlighting as a weakly supervised decision-making problem and optimize the Actor with reinforcement learning using only the Solver's task reward, requiring no evidence labels and no access to or modification of the Solver. Across sequential recommendation and long-context question answering, HiLight consistently improves performance over strong prompt-based and automated prompt-optimization baselines. The learned emphasis policy transfers zero-shot to both smaller and larger unseen Solver families, including an API-based Solver, suggesting that the Actor captures genuine, reusable evidence structure rather than overfitting to a single backbone.

Generative reasoning re-rankers achieve strong recommendation accuracy by emitting a chain-of-thought before re-ordering a candidate list, but they are slow at inference: an autoregressive (AR) decoder spends one sequential forward pass per reasoning token, and the reasoning trace far exceeds the ranking it produces. To reduce this cost, block-diffusion language models decode many positions in parallel over a few denoising steps and are substantially faster, yet naively converting an AR re-ranker into one opens two accuracy gaps: (1) a structural gap: answer positions are denoised in parallel and scored independently, so the decoder emits invalid rankings (duplicated, dropped, or out-of-set identifiers) that AR avoids through left-to-right masking; and (2) a distributional gap: fine-tuning the converted model on fixed teacher trajectories is off-policy relative to its own decoding at inference, leaving a residual accuracy gap. To close both gaps while keeping the speedup, we propose Diffusion-GR2, a recipe that converts our AR reasoning re-ranker (GR2) into a block-diffusion re-ranker. First, conversion fine-tuning (CFT) adapts the AR-initialized diffusion model to denoise the answer into a valid permutation on its own, without an external constrained decoder. Next, on-policy distillation (OPD) then supervises the model on its own decoded trajectories with dense per-token targets from the AR teacher. Finally, we apply a reinforcement-learning (RL) stage against a re-ranking reward on top of OPD's on-policy policy. Experiments on Amazon Beauty demonstrate that Diffusion-GR2 recovers to near-parity with the AR re-ranker, while block-parallel decoding raises decode throughput by 2.4–3.5× at the model's reasoning output length. Ablations show that CFT recovers most of the conversion gap, and that on-policy distillation further closes it to the AR reference.

The exponential growth of online content has posed significant challenges to ID-based models in industrial recommendation systems, ranging from extremely high cardinality and dynamically growing ID space, to highly-skewed engagement distributions, to prediction instability as a result of natural id life cycles. This paper examines these challenges and introduces Semantic ID prefix-ngram, a novel token parameterization technique that significantly improves the performance of the original Semantic ID. Semantic ID prefix-ngram creates semantically meaningful collisions by hierarchically clustering items based on their content embeddings, as opposed to random assignments. Through extensive experimentation, we demonstrate that Semantic ID prefix-ngram not only addresses embedding instability but also significantly improves tail id modeling, and mitigates representation shifts. We report our experience of integrating Semantic ID into Meta's production Ads Ranking system, leading to notable performance gains.

Ads recommendation is a prominent service of online advertising systems and has been actively studied. Recent studies indicate that scaling-up and advanced design of the recommendation model can bring significant performance improvement. However, with a larger model scale, such prior studies have a significantly increasing gap from industry as they often neglect two fundamental challenges in industrial-scale applications. First, training and inference budgets are restricted for the model to be served, exceeding which may incur latency and impair user experience. Second, large-volume data arrive in a streaming mode with data distributions dynamically shifting, as new users/ads join and existing users/ads leave the system. We propose the External Large Foundation Model (ExFM) framework to address the overlooked challenges. Specifically, we develop external distillation and a data augmentation system (DAS) to control the computational cost of training/inference while maintaining high performance. We design the teacher in a way like a foundation model (FM) that can serve multiple students as vertical models (VMs) to amortize its building cost. We propose Auxiliary Head and Student Adapter to mitigate the data distribution gap between FM and VMs caused by the streaming data issue. Comprehensive experiments on internal industrial-scale applications and public datasets demonstrate significant performance gain by ExFM.

Huge embedding tables in modern Deep Learning Recommender Models (DLRM) require prohibitively large memory during training and inference. Aiming to reduce the memory footprint of training, this paper proposes FIne-grained In-Training Embedding Dimension optimization (FIITED). Given the observation that embedding vectors are not equally important, FIITED adjusts the dimension of each individual embedding vector continuously during training, assigning longer dimensions to more important embeddings while adapting to dynamic changes in data. A novel embedding storage system based on virtually hashed physically indexed hash tables is designed to efficiently implement the embedding dimension adjustment and effectively enable memory saving. Experiments on two industry models show that FIITED is able to reduce the size of embeddings by more than 65% while maintaining the trained model’s quality, saving significantly more memory than a state-of-the-art in-training embedding pruning method. On public click-through rate prediction datasets, FIITED is able to prune up to 93.75%-99.75% embeddings without significant accuracy loss. Given the same embedding size reduction, FIITED is able to achieve better model quality than the baselines.

We study online convex optimization with constraints consisting of multiple functional constraints and a relatively simple constraint set, such as a Euclidean ball. As enforcing the constraints at each time step through projections is computationally challenging in general, we allow decisions to violate the functional constraints but aim to achieve a low regret and cumulative violation of the constraints over a horizon of T time steps. First-order methods achieve an O(sqrt{T}) regret and an O(1) constraint violation, which is the best-known bound under the Slater's condition, but do not take into account the structural information of the problem. Furthermore, the existing algorithms and analysis are limited to Euclidean space. In this paper, we provide an instance-dependent bound for online convex optimization with complex constraints obtained by a novel online primal-dual mirror-prox algorithm. Our instance-dependent regret is quantified by the total gradient variation V_*(T) in the sequence of loss functions. The proposed algorithm works in general normed spaces and simultaneously achieves an O(sqrt{V_*(T)}) regret and an O(1) constraint violation, which is never worse than the best-known (O(sqrt{T}), O(1)) result and improves over previous works that applied mirror-prox-type algorithms for this problem achieving O(T^{2/3}) regret and constraint violation. Finally, our algorithm is computationally efficient, as it only performs mirror descent steps in each iteration instead of solving a general Lagrangian minimization problem.

We examine online safe multi-agent reinforcement learning using constrained Markov games in which agents compete by maximizing their expected total rewards under a constraint on expected total utilities. Our focus is confined to an episodic two-player zero-sum constrained Markov game with independent transition functions that are unknown to agents, adversarial reward functions, and stochastic utility functions. For such a Markov game, we employ an approach based on the occupancy measure to formulate it as an online constrained saddle-point problem with an explicit constraint. We extend the Lagrange multiplier method in constrained optimization to handle the constraint by creating a generalized Lagrangian with minimax decision primal variables and a dual variable. Next, we develop an upper confidence reinforcement learning algorithm to solve this Lagrangian problem while balancing exploration and exploitation. Our algorithm updates the minimax decision primal variables via online mirror descent and the dual variable via projected gradient step and we prove that it enjoys sublinear rate $ O((|X|+|Y|) L \sqrt{T(|A|+|B|)}))$ for both regret and constraint violation after playing $T$ episodes of the game. Here, $L$ is the horizon of each episode, $(|X|,|A|)$ and $(|Y|,|B|)$ are the state/action space sizes of the min-player and the max-player, respectively. To the best of our knowledge, we provide the first provably efficient online safe reinforcement learning algorithm in constrained Markov games.

Deep learning recommendation models (DLRMs) are using increasingly larger embedding tables to represent categorical sparse features such as video genres. Each embedding row of the table represents the trainable weight vector for a specific instance of that feature. While increasing the number of embedding rows typically improves model accuracy by considering more feature instances, it can lead to larger deployment costs and slower model execution. Unlike existing efforts that primarily focus on optimizing DLRMs for the given embedding, we present a complementary system, AdaEmbed, to reduce the size of embeddings needed for the same DLRM accuracy via in-training embedding pruning. Our key insight is that the access patterns and weights of different embeddings are heterogeneous across embedding rows, and dynamically change over the training process, implying varying embedding importance with respect to model accuracy. However, identifying important embeddings and then enforcing pruning for modern DLRMs with up to billions of embeddings (terabytes) is challenging. Given the total embedding size, AdaEmbed considers embeddings with higher runtime access frequencies and larger training gradients to be more important, and it dynamically prunes less important embeddings at scale to automatically determine per-feature embeddings. Our evaluations in industrial settings show that AdaEmbed saves 35-60% embedding size needed in deployment and improves model execution speed by 11-34%, while achieving noticeable accuracy gains.

Learning feature interactions is important to the model performance of online advertising services. As a result, extensive efforts have been devoted to designing effective architectures to learn feature interactions. However, we observe that the practical performance of those designs can vary from dataset to dataset, even when the order of interactions claimed to be captured is the same. That indicates different designs may have different advantages and the interactions captured by them have non-overlapping information. Motivated by this observation, we propose DHEN - a deep and hierarchical ensemble architecture that can leverage strengths of heterogeneous interaction modules and learn a hierarchy of the interactions under different orders. To overcome the challenge brought by DHEN's deeper and multi-layer structure in training, we propose a novel co-designed training system that can further improve the training efficiency of DHEN. Experiments of DHEN on large-scale dataset from CTR prediction tasks attained 0.27\% improvement on the Normalized Entropy (NE) of prediction and 1.2x better training throughput than state-of-the-art baseline, demonstrating their effectiveness in practice.

Embedding learning has found widespread applications in recommendation systems and natural language modeling, among other domains. To learn quality embeddings efficiently, adaptive learning rate algorithms have demonstrated superior empirical performance over SGD, largely accredited to their token-dependent learning rate. However, the underlying mechanism for the efficiency of token-dependent learning rate remains underexplored. We show that incorporating frequency information of tokens in the embedding learning problems leads to provably efficient algorithms, and demonstrate that common adaptive algorithms implicitly exploit the frequency information to a large extent. Specifically, we propose (Counter-based) Frequency-aware Stochastic Gradient Descent, which applies a frequency-dependent learning rate for each token, and exhibits provable speed-up compared to SGD when the token distribution is imbalanced. Empirically, we show the proposed algorithms are able to improve or match adaptive algorithms on benchmark recommendation tasks and a large-scale industrial recommendation system, closing the performance gap between SGD and adaptive algorithms. Our results are the first to show token-dependent learning rate provably improves convergence for non-convex embedding learning problems.

Tremendous success of machine learning (ML) and the unabated growth in model complexity motivated many ML-specific designs in hardware architectures to speed up the model inference. While these architectures are diverse, highly optimized low-precision arithmetic is a component shared by most. Nevertheless, recommender systems important to Facebook’s personalization services are demanding and complex: They must serve billions of users per month responsively with low latency while maintaining high prediction accuracy. Do these low-precision architectures work well with our production recommendation systems? They do. But not without significant effort. In this article, we share our search strategies to adapt reference recommendation models to low-precision hardware, our optimization of low-precision compute kernels, and the tool chain to maintain our models’ accuracy throughout their lifespan. We believe our lessons from the trenches can promote better codesign between hardware architecture and software engineering, and advance the state of the art of ML in industry.