The training of large language models (LLMs) is expensive. In this paper, we study data-efficient approaches for pre-training LLMs, i.e., techniques that aim to optimize the Pareto frontier of model quality and training resource/data consumption. We seek to understand the tradeoffs associated with data selection routines based on (i) expensive-to-compute data-quality estimates, and (ii) maximization of coverage and diversity-based measures in the feature space. Our first technique, Ask-LLM, leverages the zero-shot reasoning capabilities of instruction-tuned LLMs to directly assess the quality of a training example. To target coverage, we propose Density sampling, which models the data distribution to select a diverse sample. In our comparison of 19 samplers, involving hundreds of evaluation tasks and pre-training runs, we find that Ask-LLM and Density are the best methods in their respective categories. Coverage sampling can recover the performance of the full data, while models trained on Ask-LLM data consistently outperform full-data training -- even when we reject 90% of the original dataset, while converging up to 70% faster.

We study the problem of building space-efficient, in-memory indexes for massive key-value datasets with highly skewed value distributions. This challenge arises in many data-intensive domains and is particularly acute in computational genomics, where k-mer count tables can contain billions of entries dominated by a single frequent value. While recent work has proposed to address this problem by augmenting compressed static functions (CSFs) with pre-filters, existing approaches rely on complex heuristics and lack formal guarantees. In this paper, we introduce a principled algorithm, called AutoCSF, for combining CSFs with pre-filtering to provably handle skewed distributions with near-optimal space usage. We improve upon prior CSF pre-filtering constructions by (1) deriving a mathematically rigorous decision criterion for when filter augmentation is beneficial; (2) presenting a general algorithmic framework for integrating CSFs with modern set membership data structures beyond the classic Bloom filter; and (3) establishing theoretical guarantees on the overall space usage of the resulting indexes. Our open-source implementation of AutoCSF demonstrates space savings over baseline methods while maintaining low query latency.

High-quality representations are a core requirement for effective recommendation. In this work, we study the problem of LLM-based descriptor generation, i.e., keyphrase-like natural language item representation generation frameworks with minimal constraints on downstream applications. We propose AgenticTagger, a framework that queries LLMs for representing items with sequences of text descriptors. However, open-ended generation provides little control over the generation space, leading to high cardinality, low-performance descriptors that renders downstream modeling challenging. To this end, AgenticTagger features two core stages: (1) a vocabulary building stage where a set of hierarchical, low-cardinality, and high-quality descriptors is identified, and (2) a vocabulary assignment stage where LLMs assign in-vocabulary descriptors to items. To effectively and efficiently ground vocabulary in the item corpus of interest, we design a multi-agent reflection mechanism where an architect LLM iteratively refines the vocabulary guided by parallelized feedback from annotator LLMs that validates the vocabulary against item data. Experiments on public and private data show AgenticTagger brings consistent improvements across diverse recommendation scenarios, including generative and term-based retrieval, ranking, and controllability-oriented, critique-based recommendation.

Large language models have become drivers of evolutionary search, but most systems rely on a fixed, prompt-elicited policy to sample next candidates. This limits adaptation in practical engineering and research tasks, where evaluations are expensive, and progress depends on learning task-specific search dynamics. We introduce PACEvolve++, an advisor-model reinforcement learning framework for test-time policy adaptation in evolutionary search agents. PACEvolve++ decouples strategic search decisions from implementation: a trainable advisor generates, assesses, and selects hypotheses, while a stronger frontier model translates selected hypotheses into executable candidates. To train the advisor under non-stationary feedback, we propose a phase-adaptive approach that adapts its optimization strategy to different phases of the evolutionary process. Early in evolution, it uses group-relative feedback to learn broad search preferences; later, as reward gaps compress, it emphasizes best-of-k frontier contribution to support stable refinement. Across expert-parallel load balancing, sequential recommendation, and protein fitness extrapolation, PACEvolve++ outperforms the state-of-the-art evolutionary search framework with frontier models, achieving faster convergence and stabilizing test-time training during evolutionary search.

Large Language Models (LLMs) have emerged as powerful operators for evolutionary search, yet the design of efficient search scaffolds remains ad hoc. While promising, current LLM-in-the-loop systems lack a systematic approach to managing the evolutionary process. We identify three distinct failure modes: Context Pollution, where experiment history biases future candidate generation; Mode Collapse, where agents stagnate in local minima due to poor exploration-exploitation balance; and Weak Collaboration, where rigid crossover strategies fail to leverage parallel search trajectories effectively. We introduce Progress-Aware Consistent Evolution (PACEvolve), a framework designed to robustly govern the agent's context and search dynamics, to address these challenges. PACEvolve combines hierarchical context management (HCM) with pruning to address context pollution; momentum-based backtracking (MBB) to escape local minima; and a self-adaptive sampling policy that unifies backtracking and crossover for dynamic search coordination (CE), allowing agents to balance internal refinement with cross-trajectory collaboration. We demonstrate that PACEvolve provides a systematic path to consistent, long-horizon self-improvement, achieving state-of-the-art results on LLM-SR and KernelBench, while discovering solutions surpassing the record on Modded NanoGPT.

Driven by recent breakthrough advances in neural representation learning, approximate near-neighbor (ANN) search over vector embeddings has emerged as a critical computational workload. With the introduction of the seminal Hierarchical Navigable Small World (HNSW) algorithm, graph-based indexes have established themselves as the overwhelmingly dominant paradigm for efficient and scalable ANN search. As the name suggests, HNSW searches a layered hierarchical graph to quickly identify neighborhoods of similar points to a given query vector. But is this hierarchy even necessary? A rigorous experimental analysis to answer this question would provide valuable insights into the nature of algorithm design for ANN search and motivate directions for future work in this increasingly crucial domain. To that end, we conduct an extensive reproducibility study covering more large-scale datasets than prior investigations of this question. We ultimately find that a flat navigable small world graph retains all of the benefits of HNSW on high-dimensional datasets, with latency and recall performance essentially identical to the original algorithm but with less memory overhead. Furthermore, we go a step beyond prior works and study why the hierarchy of HNSW provides no benefit in high dimensions, hypothesizing that navigable small world graphs contain a well-connected, frequently traversed “highway” of hub nodes that maintain the same purported function as the hierarchical layers. We present compelling empirical evidence that the Hub Highway Hypothesis holds for real datasets and investigate the mechanisms by which the highway forms. The implications of this hypothesis may also provide future research directions in developing enhancements to graph-based ANN search.

We introduce Gemma 4, a new generation of open-weight, natively multimodal language models in the Gemma model family. Designed to advance compute efficiency and reasoning, the Gemma 4 model suite features dense and Mixture-of-Experts architectures, ranging from 2.3B to 31B parameters. Alongside improved vision and audio encoders for all model sizes, we propose a unified, encoder-free architecture for our 12B model, which ingests raw audio and image patches. Furthermore, we integrate a thinking mode, enabling Gemma models to generate reasoning traces prior to responding. We improve inference speed, memory, and compute efficiency, as well as long-context abilities through critical design choices. Gemma 4 establishes a leap in performance across STEM, multimodal, and long-context benchmarks, and rivals larger, frontier open models in human-rated tasks.

Modern vector databases enable efficient retrieval over high-dimensional neural embeddings, powering applications from web search to retrieval-augmented generation. However, classical theory predicts such tasks should suffer from the curse of dimensionality, where distances between points become nearly indistinguishable, thereby crippling efficient nearest-neighbor search. We revisit this paradox through the lens of stability, the property that small perturbations to a query do not radically alter its nearest neighbors. Building on foundational results, we extend stability theory to three key retrieval settings widely used in practice: (i) multi-vector search, where we prove that the popular Chamfer distance metric preserves single-vector stability, while average pooling aggregation may destroy it; (ii) filtered vector search, where we show that sufficiently large penalties for mismatched filters can induce stability even when the underlying search is unstable; and (iii) sparse vector search, where we formalize and prove novel sufficient stability conditions. Across synthetic and real datasets, our experimental results match our theoretical predictions, offering concrete guidance for model and system design to avoid the curse of dimensionality.

Statefulness is essential for large language model (LLM) agents to perform long-term planning and problem-solving. This makes memory a critical component, yet its management and evolution remain largely underexplored. Existing evaluations mostly focus on static conversational settings, where memory is passively retrieved from dialogue to answer queries, overlooking the dynamic ability to accumulate and reuse experience across evolving task streams. In real-world environments such as interactive problem assistants or embodied agents, LLMs are required to handle continuous task streams, yet often fail to learn from accumulated interactions, losing valuable contextual insights, a limitation that calls for test-time evolution, where LLMs retrieve, integrate, and update memory continuously during deployment. To bridge this gap, we introduce Evo-Memory, a comprehensive streaming benchmark and framework for evaluating self-evolving memory in LLM agents. Evo-Memory structures datasets into sequential task streams, requiring LLMs to search, adapt, and evolve memory after each interaction. We unify and implement over ten representative memory modules and evaluate them across 10 diverse multi-turn goal-oriented and single-turn reasoning and QA datasets. To better benchmark experience reuse, we provide a baseline method, ExpRAG, for retrieving and utilizing prior experience, and further propose ReMem, an action-think-memory refine pipeline that tightly integrates reasoning, task actions, and memory updates to achieve continual improvement.

We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community.

In recent years, massive transformer-based architectures have driven breakthrough performance in practical applications like autoregressive text-generation (LLMs) and click-prediction (recommenders). A common recipe for success is to train large models on massive webscale datasets [3, 15], e.g., modern recommenders are trained on billions of user-item click events, and LLMs are trained on trillions of tokens extracted from the public internet. We are close to hitting the computational and economical limits of scaling up the size of these models, and we expect the next frontier of gains to come from improving the: (i) data quality of the training dataset, and (ii) data efficiency of the extremely expensive training procedure. Inspired by this shift, we present a set of "data-centric" techniques for recommendation and language models that summarizes a dataset into a terse data summary, which is both (i) high-quality, i.e., trains better quality models, and (ii) improves the data-efficiency of the overall training procedure. We propose techniques from two disparate data frameworks: (i) data selection (a.k.a., coreset construction) methods that sample portions of the dataset using grounded heuristics, and (ii) data distillation techniques that generate synthetic examples which are optimized to retain the signals needed for training high-quality models. Overall, this work sheds light on the challenges and opportunities offered by data optimization in web-scale systems, a particularly relevant focus as the recommendation community grapples with the grand challenge of leveraging LLMs.

Efficient large-scale neural network training and inference on commodity CPU hardware is of immense practical significance in democratizing deep learning (DL) capabilities. Presently, the process of training massive models consisting of hundreds of millions to billions of parameters requires the extensive use of specialized hardware accelerators, such as GPUs, which are only accessible to a limited number of institutions with considerable financial resources. Moreover, there is often an alarming carbon footprint associated with training and deploying these models. In this paper, we take a step towards addressing these challenges by introducing BOLT, a sparse deep learning library for training large-scale search and recommendation models on standard CPU hardware. BOLT provides a flexible, high-level API for constructing models that will be familiar to users of existing popular DL frameworks. By automatically tuning specialized hyperparameters, BOLT also abstracts away the algorithmic details of sparse network training. We evaluate BOLT on a number of information retrieval tasks including product recommendations, text classification, graph neural networks, and personalization. We find that our proposed system achieves competitive performance with state-of-the-art techniques at a fraction of the cost and energy consumption and an order-of-magnitude faster inference time. BOLT has also been successfully deployed by multiple businesses to address critical problems, and we highlight one customer case study in the field of e-commerce.

It is an ongoing endeavor in chemistry and materials science to regulate coacervate droplets on a physiologically relevant spatiotemporal scale to ultimately match or even surpass living cells' precision, complexity, and functionality. Herein, we develop a magnetic strategy orthogonal to the thermal, pH, light, or chemical counterparts that are commonly employed by biotic or artificial systems; its successful implementation thus adds a missing piece to the current arsenal of manipulative methodologies. Specifically, we paramagnetize the otherwise diamagnetic coacervate droplets by cooperatively combining paramagnetic ingredients (including organic radicals, metal ions, and Fe3O4 nanoparticles) and coacervate ingredients to obtain "MagCoa" droplets. A simple model is derived theoretically to account for migration and division of MagCoa droplets in an uneven magnetic field. Experimentally, we produce an array of compartmentalized and monodispersed droplets using microfluidics and magnetically steer them with uniformity and synchronicity. We design and fabricate spatial magnetic modulators to engineer the landscape of a magnetic field that, in turn, directs the MagCoa droplets into predesigned patterns in a reconfigurable fashion. These programmable liquid patterns can be potentially extended to dynamic assembly and information encryption. We envision that the toolbox established here is of generality and multitudes to serve as a practical guide to control droplets magnetically.

Lookup tables are a fundamental structure in many data processing and systems applications. Examples include tokenized text in NLP, quantized embedding collections in recommendation systems, integer sketches for streaming data, and hash-based string representations in genomics. With the increasing size of web-scale data, such applications often require compression techniques that support fast random $O(1)$ lookup of individual parameters directly on the compressed data (i.e. without blockwise decompression in RAM). While the community has proposd a number of succinct data structures that support queries over compressed representations, these approaches do not fully leverage the low-entropy structure prevalent in real-world workloads to reduce space. Inspired by recent advances in static function construction techniques, we propose a space-efficient representation of immutable key-value data, called CARAMEL, specifically designed for the case where the values are multi-sets. By carefully combining multiple compressed static functions, CARAMEL occupies space proportional to the data entropy with low memory overheads and minimal lookup costs. We demonstrate 1.25-16x compression on practical lookup tasks drawn from real-world systems, improving upon established techniques, including a production-grade read-only database widely used for development within Amazon.com.

"Garbage in, Garbage out" is a familiar maxim to ML practitioners and researchers, because the quality of a learned data representation is highly crucial to the quality of any ML model that consumes it as an input. To handle systems that serve billions of users at millions of queries per second (QPS), we need representation learning algorithms with significantly improved efficiency. At Google, we have dedicated thousands of iterations to develop a set of powerful techniques that efficiently learn high quality data representations. We have thoroughly validated these methods through offline evaluation, online A/B testing, and deployed these in over 50 models across major Google products. In this paper, we consider a generalized data representation learning problem that allows us to identify feature embeddings and crosses as common challenges. We propose two solutions, including: 1. Multi-size Unified Embedding to learn high-quality embeddings; and 2. Deep Cross Network V2 for learning effective feature crosses. We discuss the practical challenges we encountered and solutions we developed during deployment to production systems, compare with SOTA methods, and report offline and online experimental results. This work sheds light on the challenges and opportunities for developing next-gen algorithms for web-scale systems.

With the surging popularity of approximate near-neighbor search (ANNS), driven by advances in neural representation learning, the ability to serve queries accompanied by a set of constraints has become an area of intense interest. While the community has recently proposed several algorithms for constrained ANNS, almost all of these methods focus on integration with graph-based indexes, the predominant class of algorithms achieving state-of-the-art performance in latency-recall tradeoffs. In this work, we take a different approach and focus on developing a constrained ANNS algorithm via space partitioning as opposed to graphs. To that end, we introduce Constrained Approximate Partitioned Search (CAPS), an index for ANNS with filters via space partitions that not only retains the benefits of a partition-based algorithm but also outperforms state-of-the-art graph-based constrained search techniques in recall-latency tradeoffs, with only 10% of the index size.

We study the problem of $\textit{vector set search}$ with $\textit{vector set queries}$. This task is analogous to traditional near-neighbor search, with the exception that both the query and each element in the collection are $\textit{sets}$ of vectors. We identify this problem as a core subroutine for semantic search applications and find that existing solutions are unacceptably slow. Towards this end, we present a new approximate search algorithm, DESSERT (${\bf D}$ESSERT ${\bf E}$ffeciently ${\bf S}$earches ${\bf S}$ets of ${\bf E}$mbeddings via ${\bf R}$etrieval ${\bf T}$ables). DESSERT is a general tool with strong theoretical guarantees and excellent empirical performance. When we integrate DESSERT into ColBERT, a state-of-the-art semantic search model, we find a 2-5x speedup on the MS MARCO and LoTTE retrieval benchmarks with minimal loss in recall, underscoring the effectiveness and practical applicability of our proposal.

There has been an exponential increase in the data generated worldwide. Insights into this data led by machine learning (ML) have given rise to exciting applications such as recommendation engines, conversational agents, and so on. Often, data for these applications is generated at a rate faster than ML pipelines can consume it. In this paper, we propose Density Sketches(DS) - a cheap and practical approach to reducing data redundancy in a streaming fashion. DS creates a succinct online summary of data distribution. While DS does not store the samples from the stream, we can sample unseen data on the fly from DS to use for downstream learning tasks. In this sense, DS can replace actual data in many machine learning pipelines analogous to generative models. Importantly, unlike generative models, which do not have statistical guarantees, the sampling distribution of DS asymptotically converges to underlying unknown density distribution.

Data sampling is an effective method to improve the training speed of neural networks, with recent results demonstrating that it can even break the neural scaling laws. These results critically rely on high-quality scores to estimate the importance of an input to the network. We observe that there are two dominant strategies: static sampling, where the scores are determined before training, and dynamic sampling, where the scores can depend on the model weights. Static algorithms are computationally inexpensive but less effective than their dynamic counterparts, which can cause end-to-end slowdown due to their need to explicitly compute losses. To address this problem, we propose a novel sampling distribution based on nonparametric kernel regression that learns an effective importance score as the neural network trains. However, nonparametric regression models are too computationally expensive to accelerate end-to-end training. Therefore, we develop an efficient sketch-based approximation to the Nadaraya-Watson estimator. Using recent techniques from high-dimensional statistics and randomized algorithms, we prove that our Nadaraya-Watson sketch approximates the estimator with exponential convergence guarantees. Our sampling algorithm outperforms the baseline in terms of wall-clock time and accuracy on four datasets.

Federated learning (FL) is a machine learning paradigm where multiple client devices train models collaboratively without data exchange. Data heterogeneity problem is naturally inherited in FL since data in different clients follow diverse distributions. To mitigate the negative influence of data heterogeneity, we need to start by measuring it across clients. However, the efficient measurement between distributions is a challenging problem, especially in high dimensionality. In this paper, we propose a one-pass distribution sketch to represent the client data distribution. Our sketching algorithm only requires a single pass of the client data, which is efficient in terms of time and memory. Moreover, we show in both theory and practice that the distance between two distribution sketches represents the divergence between their corresponding distributions. Furthermore, we demonstrate with extensive experiments that our distribution sketch improves the client selection in the FL training. We also showcase that our distribution sketch is an efficient solution to the cold start problem in FL for new clients with unlabeled data.