Sparton: Fast and Memory-Efficient Triton Kernel for Learned Sparse Retrieval

State-of-the-art Learned Sparse Retrieval (LSR) models, such as Splade, typically employ a Language Modeling (LM) head to project latent hidden states into a lexically-anchored logit matrix. This intermediate matrix is subsequently transformed into a sparse lexical representation through element-wise operations (ReLU, Log1P) and max-pooling over the sequence dimension. Despite its effectiveness, the LM head creates a massive memory bottleneck due to the sheer size of the vocabulary (V), which can range from 30,000 to over 250,000 tokens in recent models. Materializing this matrix creates a significant memory bottleneck, limiting model scaling. The resulting I/O overhead between operators further throttles throughput and runtime performance. In this paper, we propose Sparton, a fast memory-efficient Triton kernel tailored for the LM head in LSR models. Sparton utilizes a fused approach that integrates the tiled matrix multiplication, ReLU, Log1P, and max-reduction into a single GPU kernel. By performing an early online reduction directly on raw logit tiles, Sparton avoids materializing the full logit matrix in memory. Our experiments demonstrate that the Sparton kernel, in isolation, achieves up to a 4.8x speedup and an order-of-magnitude reduction in peak memory usage compared to PyTorch baselines. Integrated into Splade (|V| ~ 30k), Sparton enables a 33% larger batch size and 14% faster training with no effectiveness loss. On a multilingual backbone (|V| ~ 250k), these gains jump to a 26x larger batch size and 2.5x faster training.

Forward Index Compression for Learned Sparse Retrieval

Text retrieval using learned sparse representations of queries and documents has, over the years, evolved into a highly effective approach to search. It is thanks to recent advances in approximate nearest neighbor search-with the emergence of highly efficient algorithms such as the inverted index-based Seismic and the graph-based Hnsw-that retrieval with sparse representations became viable in practice. In this work, we scrutinize the efficiency of sparse retrieval algorithms and focus particularly on the size of a data structure that is common to all algorithmic flavors and that constitutes a substantial fraction of the overall index size: the forward index. In particular, we seek compression techniques to reduce the storage footprint of the forward index without compromising search quality or inner product computation latency. In our examination with various integer compression techniques, we report that StreamVByte achieves the best trade-off between memory footprint, retrieval accuracy, and latency. We then improve StreamVByte by introducing DotVByte, a new algorithm tailored to inner product computation. Experiments on MsMarco show that our improvements lead to significant space savings while maintaining retrieval efficiency.

Multivector Reranking in the Era of Strong First-Stage Retrievers

Learned multivector representations power modern search systems with strong retrieval effectiveness, but their real-world use is limited by the high cost of exhaustive token-level retrieval. Therefore, most systems adopt a \emph{gather-and-refine} strategy, where a lightweight gather phase selects candidates for full scoring. However, this approach requires expensive searches over large token-level indexes and often misses the documents that would rank highest under full similarity. In this paper, we reproduce several state-of-the-art multivector retrieval methods on two publicly available datasets, providing a clear picture of the current multivector retrieval field and observing the inefficiency of token-level gathering. Building on top of that, we show that replacing the token-level gather phase with a single-vector document retriever -- specifically, a learned sparse retriever (LSR) -- produces a smaller and more semantically coherent candidate set. This recasts the gather-and-refine pipeline into the well-established two-stage retrieval architecture. As retrieval latency decreases, query encoding with two neural encoders becomes the dominant computational bottleneck. To mitigate this, we integrate recent inference-free LSR methods, demonstrating that they preserve the retrieval effectiveness of the dual-encoder pipeline while substantially reducing query encoding time. Finally, we investigate multiple reranking configurations that balance efficiency, memory, and effectiveness, and we introduce two optimization techniques that prune low-quality candidates early. Empirical results show that these techniques improve retrieval efficiency by up to 1.8$\times$ with no loss in quality. Overall, our two-stage approach achieves over $24\times$ speedup over the state-of-the-art multivector retrieval systems, while maintaining comparable or superior retrieval quality.

Early-Exit Graph Neural Networks

Early-exit mechanisms allow deep neural networks to halt inference as soon as classification confidence is high enough, adaptively trading depth for confidence, and thereby cutting latency and energy on easy inputs while retaining full-depth accuracy for harder ones. Similarly, adding early exit mechanisms to Graph Neural Networks (GNNs), the go-to models for graph-structured data, allows for dynamic trading depth for confidence on simple graphs while maintaining full-depth accuracy on harder and more complex graphs to capture intricate relationships. Although early exits have proven effective across various deep learning domains, their potential within GNNs in scenarios that require deep architectures while resisting over-smoothing and over-squashing remains largely unexplored. We unlock that potential by first introducing Symmetric-Anti-Symmetric Graph Neural Networks (SAS-GNN), whose symmetry-based inductive biases mitigate these issues and yield stable intermediate representations that can be useful to allow early exiting in GNNs. Building on this backbone, we present Early-Exit Graph Neural Networks (EEGNNs), which append confidence-aware exit heads that allow on-the-fly termination of propagation based on each node or the entire graph. Experiments show that EEGNNs preserve robust performance as depth grows and deliver competitive accuracy on heterophilic and long-range benchmarks, matching attention-based and asynchronous message-passing models while substantially reducing computation and latency. We plan to release the code to reproduce our experiments.

Efficient Conversational Search via Topical Locality in Dense Retrieval

Pre-trained language models have been widely exploited to learn dense representations of documents and queries for information retrieval. While previous efforts have primarily focused on improving effectiveness and user satisfaction, response time remains a critical bottleneck of conversational search systems. To address this, we exploit the topical locality inherent in conversational queries, i.e., the tendency of queries within a conversation to focus on related topics. By leveraging query embedding similarities, we dynamically restrict the search space to semantically relevant document clusters, reducing computational complexity without compromising retrieval quality. We evaluate our approach on the TREC CAsT 2019 and 2020 datasets using multiple embedding models and vector indexes, achieving improvements in processing speed of up to 10.4X with little loss in performance (4.4X without any loss). Our results show that the proposed system effectively handles complex, multiturn queries with high precision and efficiency, offering a practical solution for real-time conversational search.

Effective Inference-Free Retrieval for Learned Sparse Representations

Learned Sparse Retrieval (LSR) is an effective IR approach that exploits pre-trained language models for encoding text into a learned bag of words. Several efforts in the literature have shown that sparsity is key to enabling a good trade-off between the efficiency and effectiveness of the query processor. To induce the right degree of sparsity, researchers typically use regularization techniques when training LSR models. Recently, new efficient -- inverted index-based -- retrieval engines have been proposed, leading to a natural question: has the role of regularization changed in training LSR models? In this paper, we conduct an extended evaluation of regularization approaches for LSR where we discuss their effectiveness, efficiency, and out-of-domain generalization capabilities. We first show that regularization can be relaxed to produce more effective LSR encoders. We also show that query encoding is now the bottleneck limiting the overall query processor performance. To remove this bottleneck, we advance the state-of-the-art of inference-free LSR by proposing Learned Inference-free Retrieval (Li-LSR). At training time, Li-LSR learns a score for each token, casting the query encoding step into a seamless table lookup. Our approach yields state-of-the-art effectiveness for both in-domain and out-of-domain evaluation, surpassing Splade-v3-Doc by 1 point of mRR@10 on MS MARCO and 1.8 points of nDCG@10 on BEIR.

Investigating the Scalability of Approximate Sparse Retrieval Algorithms to Massive Datasets

Learned sparse text embeddings have gained popularity due to their effectiveness in top-k retrieval and inherent interpretability. Their distributional idiosyncrasies, however, have long hindered their use in real-world retrieval systems. That changed with the recent development of approximate algorithms that leverage the distributional properties of sparse embeddings to speed up retrieval. Nonetheless, in much of the existing literature, evaluation has been limited to datasets with only a few million documents such as MSMARCO. It remains unclear how these systems behave on much larger datasets and what challenges lurk in larger scales. To bridge that gap, we investigate the behavior of state-of-the-art retrieval algorithms on massive datasets. We compare and contrast the recently-proposed Seismic and graph-based solutions adapted from dense retrieval. We extensively evaluate Splade embeddings of 138M passages from MsMarco-v2 and report indexing time and other efficiency and effectiveness metrics.

kANNolo: Sweet and Smooth Approximate k-Nearest Neighbors Search

Approximate Nearest Neighbors (ANN) search is a crucial task in several applications like recommender systems and information retrieval. Current state-of-the-art ANN libraries, although being performance-oriented, often lack modularity and ease of use. This translates into them not being fully suitable for easy prototyping and testing of research ideas, an important feature to enable. We address these limitations by introducing kANNolo, a novel research-oriented ANN library written in Rust and explicitly designed to combine usability with performance effectively. kANNolo is the first ANN library that supports dense and sparse vector representations made available on top of different similarity measures, e.g., euclidean distance and inner product. Moreover, it also supports vector quantization techniques, e.g., Product Quantization, on top of the indexing strategies implemented. These functionalities are managed through Rust traits, allowing shared behaviors to be handled abstractly. This abstraction ensures flexibility and facilitates an easy integration of new components. In this work, we detail the architecture of kANNolo and demonstrate that its flexibility does not compromise performance. The experimental analysis shows that kANNolo achieves state-of-the-art performance in terms of speed-accuracy trade-off while allowing fast and easy prototyping, thus making kANNolo a valuable tool for advancing ANN research. Source code available on GitHub: https://github.com/TusKANNy/kannolo.

Power- and Fragmentation-aware Online Scheduling for GPU Datacenters

The rise of Artificial Intelligence and Large Language Models is driving increased GPU usage in data centers for complex training and inference tasks, impacting operational costs, energy demands, and the environmental footprint of large-scale computing infrastructures. This work addresses the online scheduling problem in GPU datacenters, which involves scheduling tasks without knowledge of their future arrivals. We focus on two objectives: minimizing GPU fragmentation and reducing power consumption. GPU fragmentation occurs when partial GPU allocations hinder the efficient use of remaining resources, especially as the datacenter nears full capacity. A recent scheduling policy, Fragmentation Gradient Descent (FGD), leverages a fragmentation metric to address this issue. Reducing power consumption is also crucial due to the significant power demands of GPUs. To this end, we propose PWR, a novel scheduling policy to minimize power usage by selecting power-efficient GPU and CPU combinations. This involves a simplified model for measuring power consumption integrated into a Kubernetes score plugin. Through an extensive experimental evaluation in a simulated cluster, we show how PWR, when combined with FGD, achieves a balanced trade-off between reducing power consumption and minimizing GPU fragmentation.

Rewriting Conversational Utterances with Instructed Large Language Models

Many recent studies have shown the ability of large language models (LLMs) to achieve state-of-the-art performance on many NLP tasks, such as question answering, text summarization, coding, and translation. In some cases, the results provided by LLMs are on par with those of human experts. These models' most disruptive innovation is their ability to perform tasks via zero-shot or few-shot prompting. This capability has been successfully exploited to train instructed LLMs, where reinforcement learning with human feedback is used to guide the model to follow the user's requests directly. In this paper, we investigate the ability of instructed LLMs to improve conversational search effectiveness by rewriting user questions in a conversational setting. We study which prompts provide the most informative rewritten utterances that lead to the best retrieval performance. Reproducible experiments are conducted on publicly-available TREC CAST datasets. The results show that rewriting conversational utterances with instructed LLMs achieves significant improvements of up to 25.2% in MRR, 31.7% in Precision@1, 27% in NDCG@3, and 11.5% in Recall@500 over state-of-the-art techniques.