arXiv subjects

Oshri Naparstek

Publications and source records attributed to Oshri Naparstek.

16 recordsLinked to original sources

arXiv2026

FLASH-MAXSIM: IO-Aware Fused Kernels for Late-Interaction Retrieval

Late-interaction retrieval (ColBERT, ColPali) scores a query against a document via the MaxSim operator. The standard PyTorch implementation materialises the full query-token $\times$ document-token similarity tensor only to reduce it away. At ColPali scale this is the single largest tensor in the pipeline (e.g. 21 GB in FP16 for 10K documents) and limits both candidate set size at inference and batch size during contrastive training. We present FLASH-MAXSIM (FM), an IO-aware fused GPU kernel that computes the same MaxSim scores without ever materialising the tensor, and extends the same principle to the training backward. At ColPali scale on A100, FM reduces inference peak memory by 1.4-2.6$\times$ relative to the deployed chunked baseline (4.9-8.9$\times$ relative to unchunked eager execution) and reduces MaxSim-operator training memory by two orders of magnitude, enabling exact reranking over larger resident candidate pools and contrastive batch sizes that vanilla autograd cannot fit on a single GPU. The kernel is a drop-in replacement, exact up to floating-point evaluation order under its stated FP32-accumulation protocol: nDCG@10 differs from the FP32 reference by at most $5\times10^{-4}$ on BEIR and REAL-MM-RAG. A separate INT8 path trades exactness for halved index storage at high fidelity. Code, benchmark scripts, and raw results: https://github.com/roipony/flash-maxsim

cs.IR↗

arXiv2026

Evolution or Illusion? Rethinking Evaluation in LLM Evolutionary Search

LLM-driven evolutionary search finds programs by launching seeds and iterating each one. Papers report a single budget setting, usually one seed run for a fixed number of iterations, and rank methods from that one point. We show this is not enough. We evaluate three evolutionary search strategies on five optimization tasks, commonly used by papers in the genre to report results. We run the analysis over a full grid of seeds and iterations. Our findings suggest that the best way to split a fixed budget between more seeds (width) and more iterations (depth) changes with the strategy, the task, and the total budget. Furthermore, we observe that the ranking of strategies also changes with the budget. On one task the strategy that looks worst at one seed is best at forty seeds. On another the best number of iterations is well below the value common in practice, so extra depth wastes budget that more seeds would turn into score. We provide a measurement protocol that reports the seeds-by-iterations frontier and practical guidance for using it.

cs.CL↗

arXiv2026

Spotlights: Discovering Improvement Opportunities in Software Repositories

Coding agents and evolutionary code-search systems can improve implementations once a target and evaluation criterion have been specified. Applying these methods to an existing software repository raises an earlier question: which implementation choices are worth investigating for a high-level engineering objective? We introduce \emph{optimization-opportunity discovery}, the repository-level task of identifying candidate source regions, explaining how they relate to the objective, and proposing possible changes. The task takes as input a repository, an engineering objective, and optional runtime evidence such as offline telemetry observations or profiles. It does not require the user to specify a defect, bottleneck, or code location. We present \emph{Spotlights}, a system that performs this task through logical repository mapping, successive agent reviews, and optional research linking candidates to relevant techniques. We evaluate Spotlights across model serving, document retrieval, blockchain ordering, and document processing. Across three cases, it recovers seven of nine expert-selected targets. In the reliability study, 70\% of the top ten candidates meet the stated correctness and severity thresholds. Across five repeated retrieval runs, 73.6\% of candidate occurrences have a matching source region in all five runs. Spotlights also rediscovers the target of a withheld retrieval optimization and connects it to a relevant tiling technique. In an implementation study, a discovered change reduces end-to-end page-processing runtime by 10.6\% while preserving measured output quality. These results establish optimization-opportunity discovery as a distinct and empirically evaluable step between a broad engineering objective and subsequent implementation and validation.

cs.SE↗

arXiv2026

Optimize Cheap, Deploy Strong: Cost-Aware Cross-Tier Transfer for Evolutionary Optimization

Evolutionary optimization of LLM prompts and agentic programs (e.g., GEPA) is dominated by fitness evaluation: scoring each candidate runs an answering LLM over a validation set, so the evaluator's price tier dictates total search cost. We restructure that search by decoupling the three roles an LLM plays, running the high-volume answering role on the cheapest tier, reserving a strong model for the rare reflection/variation operator, then exploiting upward cross-tier transfer to deploy the cheaply evolved prompt on a stronger target. We contribute a cost-controlled characterization of when cheap-tier search substitutes for target-tier search, and where it fails. Across four tasks (HotpotQA, IFBench, LiveBench-Math, HoVer) and eleven models in four model families, the resulting prompt matches or exceeds same-tier optimization while placing over 96% of search tokens on the cheapest tier, at 5.6-14x lower search cost, rising to 25-54x where reasoning tiers emit long chains of thought on every fitness call.

cs.LG↗

arXiv2026

Col-Bandit: Query-Time Top-$K$ Estimation for Late-Interaction Retrieval

Multi-vector late-interaction retrievers such as ColBERT achieve state-of-the-art quality, but their query-time cost is dominated by exhaustively computing token-level MaxSim interactions for every candidate document. The MaxSim scores of $N$ candidates against $T$ query tokens form an $N\times T$ matrix whose row-sums are the late-interaction scores, and identifying the top-$K$ rarely requires every entry. We introduce Col-Bandit, a query-time estimator of the exhaustive-MaxSim top-$K$: it reveals matrix entries in batches, maintains a finite-population Bernstein-Serfling confidence interval on each candidate's score, and permanently drops any document whose upper bound falls below the $K$-th largest lower bound, computing only the cells needed to separate the top-$K$. A single relaxation knob $α_{\mathrm{ef}}\in(0,1]$ tunes the compute-fidelity trade-off. We deploy $α_{\mathrm{ef}}{=}0.2$, while $α_{\mathrm{ef}}{=}1$ admits a $δ$-PAC guarantee under a simplified radius. On BEIR and REAL-MM-RAG, Col-Bandit preserves $\geq 90\%$ fidelity to the exhaustive top-$5$ on every corpus while cutting MaxSim FLOPs by up to ${\sim}8\times$, for up to ${\sim}13\times$ single-thread CPU speedups across x86 and ARM. A drop-in reranking layer, it needs no retraining or index changes.

cs.IR↗

arXiv2026

Is the Modality Gap a Bug or a Feature? A Robustness Perspective

Many modern multi-modal models (e.g. CLIP) seek an embedding space in which the two modalities are aligned. Somewhat surprisingly, almost all existing models show a strong modality gap: the distribution of images is well-separated from the distribution of texts in the shared embedding space. Despite a series of recent papers on this topic, it is still not clear why this gap exists nor whether closing the gap in post-processing will lead to better performance on downstream tasks. In this paper we show that under certain conditions, minimizing the contrastive loss yields a representation in which the two modalities are separated by a global gap vector that is orthogonal to their embeddings. We also show that under these conditions the modality gap is monotonically related to robustness: decreasing the gap does not change the clean accuracy of the models but makes it less likely that a model will change its output when the embeddings are perturbed. Our experiments show that for many real-world VLMs we can significantly increase robustness by a simple post-processing step that moves one modality towards the mean of the other modality, without any loss of clean accuracy.

cs.CV↗

arXiv2026

Projected Autoregression: Autoregressive Language Generation in Continuous State Space

Standard autoregressive language models generate text by repeatedly selecting a discrete next token, coupling prediction with irreversible commitment at every step. We show that token selection is not the only viable autoregressive interface. \textbf{Projected Autoregression} replaces token selection with continuous prediction in embedding space followed by discrete projection at commitment time. The model predicts next-token vectors via regression and contrastive objectives, while discrete tokens arise only by nearest-neighbor projection. An optional mutable suffix (``liquid tail'') enables iterative refinement before commitment, but the central change is more basic: next-step prediction is continuous, and discrete tokens are produced only as a downstream interface. Projected Autoregression establishes a concrete alternative to token-selection autoregression: language generation can be organized around continuous-state prediction with delayed discrete commitment. Refinement remains local to a short causal suffix within a left-to-right causal process, rather than a sequence-wide denoising process. This separation has two consequences. First, it induces a \emph{distinct generation regime}: even with immediate projection ($K{=}1$), continuous prediction yields text structure and dynamics that differ from tested token-space AR baselines, including a compute-matched best-of-16 reranking baseline. Second, it exposes a \emph{continuous control surface} inside autoregressive generation: direction rate, history noise, delayed commitment, state-space guidance, and embedding geometry act directly on the evolving generative state before token commitment. Taken together, these results place repeated token selection within a larger family of autoregressive interfaces and expose continuous state space as a broader algorithmic design space for language generation.

cs.CL↗

arXiv2025

Complexity as Advantage: A Regret-Based Perspective on Emergent Structure

We introduce Complexity as Advantage (CAA), a framework that defines the complexity of a system relative to a family of observers. Instead of measuring complexity as an intrinsic property, we evaluate how much predictive regret a system induces for different observers attempting to model it. A system is complex when it is easy for some observers and hard for others, creating an information advantage. We show that this formulation unifies several notions of emergent behavior, including multiscale entropy, predictive information, and observer-dependent structure. The framework suggests that "interesting" systems are those positioned to create differentiated regret across observers, providing a quantitative grounding for why complexity can be functionally valuable. We demonstrate the idea through simple dynamical models and discuss implications for learning, evolution, and artificial agents.

cs.LG↗

arXiv2025

REAL-MM-RAG: A Real-World Multi-Modal Retrieval Benchmark

Accurate multi-modal document retrieval is crucial for Retrieval-Augmented Generation (RAG), yet existing benchmarks do not fully capture real-world challenges with their current design. We introduce REAL-MM-RAG, an automatically generated benchmark designed to address four key properties essential for real-world retrieval: (i) multi-modal documents, (ii) enhanced difficulty, (iii) Realistic-RAG queries and (iv) accurate labeling. Additionally, we propose a multi-difficulty-level scheme based on query rephrasing to evaluate models' semantic understanding beyond keyword matching. Our benchmark reveals significant model weaknesses, particularly in handling table-heavy documents and robustness to query rephrasing. To mitigate these shortcomings, we curate a rephrased training set and introduce a new finance-focused, table-heavy dataset. Fine-tuning on these datasets enables models to achieve state-of-the-art retrieval performance on REAL-MM-RAG benchmark. Our work offers a better way to evaluate and improve retrieval in multi-modal RAG systems while also providing training data and models that address current limitations.

cs.IR↗

arXiv2025

Granite Vision: a lightweight, open-source multimodal model for enterprise Intelligence

We introduce Granite Vision, a lightweight large language model with vision capabilities, specifically designed to excel in enterprise use cases, particularly in visual document understanding. Our model is trained on a comprehensive instruction-following dataset, including document-related tasks, such as content extraction from tables, charts, diagrams, sketches, and infographics, as well as general image tasks. The architecture of Granite Vision is centered around visual modality alignment with a decoder-only, 2 billion parameter Granite large language model. Additionally, we introduce a dedicated safety classification approach in test-time that leverages a sparse set of attention vectors to identify potential harmful inputs. Despite its lightweight architecture, Granite Vision achieves strong results in standard benchmarks related to visual document understanding, as well as on the LiveXiv benchmark, which is designed to avoid test set contamination by using a constantly updated corpus of recently published Arxiv papers. We are releasing the model under the Apache-2 license, allowing for both research and commercial use, while offering complete visibility into the training data and other relevant details. See https://huggingface.co/ibm-granite/ for model weights.

cs.CV↗

arXiv2024

KVP10k : A Comprehensive Dataset for Key-Value Pair Extraction in Business Documents

In recent years, the challenge of extracting information from business documents has emerged as a critical task, finding applications across numerous domains. This effort has attracted substantial interest from both industry and academy, highlighting its significance in the current technological landscape. Most datasets in this area are primarily focused on Key Information Extraction (KIE), where the extraction process revolves around extracting information using a specific, predefined set of keys. Unlike most existing datasets and benchmarks, our focus is on discovering key-value pairs (KVPs) without relying on predefined keys, navigating through an array of diverse templates and complex layouts. This task presents unique challenges, primarily due to the absence of comprehensive datasets and benchmarks tailored for non-predetermined KVP extraction. To address this gap, we introduce KVP10k , a new dataset and benchmark specifically designed for KVP extraction. The dataset contains 10707 richly annotated images. In our benchmark, we also introduce a new challenging task that combines elements of KIE as well as KVP in a single task. KVP10k sets itself apart with its extensive diversity in data and richly detailed annotations, paving the way for advancements in the field of information extraction from complex business documents.

cs.IR↗

arXiv2022

BusiNet -- a Light and Fast Text Detection Network for Business Documents

For digitizing or indexing physical documents, Optical Character Recognition (OCR), the process of extracting textual information from scanned documents, is a vital technology. When a document is visually damaged or contains non-textual elements, existing technologies can yield poor results, as erroneous detection results can greatly affect the quality of OCR. In this paper we present a detection network dubbed BusiNet aimed at OCR of business documents. Business documents often include sensitive information and as such they cannot be uploaded to a cloud service for OCR. BusiNet was designed to be fast and light so it could run locally preventing privacy issues. Furthermore, BusiNet is built to handle scanned document corruption and noise using a specialized synthetic dataset. The model is made robust to unseen noise by employing adversarial training strategies. We perform an evaluation on publicly available datasets demonstrating the usefulness and broad applicability of our model.

cs.CV↗

arXiv2018

Deep Multi-User Reinforcement Learning for Distributed Dynamic Spectrum Access

We consider the problem of dynamic spectrum access for network utility maximization in multichannel wireless networks. The shared bandwidth is divided into K orthogonal channels. In the beginning of each time slot, each user selects a channel and transmits a packet with a certain transmission probability. After each time slot, each user that has transmitted a packet receives a local observation indicating whether its packet was successfully delivered or not (i.e., ACK signal). The objective is a multi-user strategy for accessing the spectrum that maximizes a certain network utility in a distributed manner without online coordination or message exchanges between users. Obtaining an optimal solution for the spectrum access problem is computationally expensive in general due to the large state space and partial observability of the states. To tackle this problem, we develop a novel distributed dynamic spectrum access algorithm based on deep multi-user reinforcement leaning. Specifically, at each time slot, each user maps its current state to spectrum access actions based on a trained deep-Q network used to maximize the objective function. Game theoretic analysis of the system dynamics is developed for establishing design principles for the implementation of the algorithm. Experimental results demonstrate strong performance of the algorithm.

cs.NI↗

arXiv2018

Distributed Energy Efficient Channel Allocation

Design of energy efficient protocols for modern wireless systems has become an important area of research. In this paper, we propose a distributed optimization algorithm for the channel assignment problem for multiple interfering transceiver pairs that cannot communicate with each other. We first modify the auction algorithm for maximal energy efficiency and show that the problem can be solved without explicit message passing using the carrier sense multiple access (CSMA) protocols. We then develop a novel scheme by converting the channel assignment problem into perfect matchings on bipartite graphs. The proposed scheme improves the energy efficiency and does not require any explicit message passing or a shared memory between the users. We derive bounds on the convergence rate and show that the proposed algorithm converges faster than the distributed auction algorithm and achieves near-optimal performance under Rayleigh fading channels. We also present an asymptotic performance analysis of the fast matching algorithm for energy efficient resource allocation and prove the optimality for large enough number of users and number of channels. Finally, we provide numerical assessments that confirm the energy efficiency gains compared to the state of the art.

cs.NI↗

arXiv2013

Expected time complexity of the auction algorithm and the push relabel algorithm for maximal bipartite matching on random graphs

In this paper we analyze the expected time complexity of the auction algorithm for the matching problem on random bipartite graphs. We prove that the expected time complexity of the auction algorithm for bipartite matching is $O\left(\frac{N\log^2(N)}{\log\left(Np\right)}\right)$ on sequential machines. This is equivalent to other augmenting path algorithms such as the HK algorithm. Furthermore, we show that the algorithm can be implemented on parallel machines with $O(\log(N))$ processors and shared memory with an expected time complexity of $O(N\log(N))$.

cs.DS↗

arXiv2013

Fully distributed optimal channel assignment for open spectrum access

In this paper we address the problem of fully distributed assignment of users to sub-bands such that the sum-rate of the system is maximized. We introduce a modified auction algorithm that can be applied in a fully distributed way using an opportunistic CSMA assignment scheme and is $ε$ optimal. We analyze the expected time complexity of the algorithm and suggest a variant to the algorithm that has lower expected complexity. We then show that in the case of i.i.d Rayleigh channels a simple greedy scheme is asymptotically optimal as $\SNR$ increases or as the number of users is increased to infinity. We conclude by providing simulated results of the suggested algorithms.

cs.DC↗