[{"type":"conference","department":[{"_id":"DaAl"}],"year":"2025","alternative_title":["Advances in Neural Information Processing Systems"],"publication_status":"published","language":[{"iso":"eng"}],"user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","publisher":"Neural Information Processing Systems Foundation","doi":"10.52202/085713-1551","citation":{"mla":"Rodionov, Gleb, et al. “Hogwild! Inference: Parallel LLM Generation via Concurrent Attention.” <i>39th Conference on Neural Information Processing Systems</i>, vol. 38, Neural Information Processing Systems Foundation, 2025, pp. 52014–55, doi:<a href=\"https://doi.org/10.52202/085713-1551\">10.52202/085713-1551</a>.","ieee":"G. Rodionov <i>et al.</i>, “Hogwild! Inference: Parallel LLM generation via concurrent attention,” in <i>39th Conference on Neural Information Processing Systems</i>, San Diego, CA, United States, 2025, vol. 38, pp. 52014–52055.","ama":"Rodionov G, Garipov R, Shutova A, et al. Hogwild! Inference: Parallel LLM generation via concurrent attention. In: <i>39th Conference on Neural Information Processing Systems</i>. Vol 38. Neural Information Processing Systems Foundation; 2025:52014-52055. doi:<a href=\"https://doi.org/10.52202/085713-1551\">10.52202/085713-1551</a>","chicago":"Rodionov, Gleb, Roman Garipov, Alina Shutova, George Yakushev, Erik Schultheis, Vage Egiazarian, Anton Sinitsin, Denis Kuznedelev, and Dan-Adrian Alistarh. “Hogwild! Inference: Parallel LLM Generation via Concurrent Attention.” In <i>39th Conference on Neural Information Processing Systems</i>, 38:52014–55. Neural Information Processing Systems Foundation, 2025. <a href=\"https://doi.org/10.52202/085713-1551\">https://doi.org/10.52202/085713-1551</a>.","apa":"Rodionov, G., Garipov, R., Shutova, A., Yakushev, G., Schultheis, E., Egiazarian, V., … Alistarh, D.-A. (2025). Hogwild! Inference: Parallel LLM generation via concurrent attention. In <i>39th Conference on Neural Information Processing Systems</i> (Vol. 38, pp. 52014–52055). San Diego, CA, United States: Neural Information Processing Systems Foundation. <a href=\"https://doi.org/10.52202/085713-1551\">https://doi.org/10.52202/085713-1551</a>","short":"G. Rodionov, R. Garipov, A. Shutova, G. Yakushev, E. Schultheis, V. Egiazarian, A. Sinitsin, D. Kuznedelev, D.-A. Alistarh, in:, 39th Conference on Neural Information Processing Systems, Neural Information Processing Systems Foundation, 2025, pp. 52014–52055.","ista":"Rodionov G, Garipov R, Shutova A, Yakushev G, Schultheis E, Egiazarian V, Sinitsin A, Kuznedelev D, Alistarh D-A. 2025. Hogwild! Inference: Parallel LLM generation via concurrent attention. 39th Conference on Neural Information Processing Systems. NeurIPS: Neural Information Processing Systems, Advances in Neural Information Processing Systems, vol. 38, 52014–52055."},"conference":{"location":"San Diego, CA, United States","end_date":"2025-12-07","name":"NeurIPS: Neural Information Processing Systems","start_date":"2025-12-02"},"has_accepted_license":"1","publication":"39th Conference on Neural Information Processing Systems","article_processing_charge":"No","intvolume":"        38","oa_version":"Published Version","day":"02","scopus_import":"1","date_published":"2025-12-02T00:00:00Z","_id":"22826","quality_controlled":"1","date_updated":"2026-09-10T07:06:31Z","status":"public","volume":38,"title":"Hogwild! Inference: Parallel LLM generation via concurrent attention","oa":1,"page":"52014-52055","month":"12","researchdata_availability":"no","main_file_link":[{"url":"https://doi.org/10.52202/085713-1551","open_access":"1"}],"ddc":["000"],"supplementarymaterial":"yes","author":[{"first_name":"Gleb","full_name":"Rodionov, Gleb","last_name":"Rodionov"},{"last_name":"Garipov","full_name":"Garipov, Roman","first_name":"Roman"},{"full_name":"Shutova, Alina","last_name":"Shutova","first_name":"Alina"},{"first_name":"George","last_name":"Yakushev","full_name":"Yakushev, George"},{"first_name":"Erik","full_name":"Schultheis, Erik","id":"2786b299-e6b0-11f0-91da-9243fe3ef96b","last_name":"Schultheis"},{"first_name":"Vage","full_name":"Egiazarian, Vage","id":"77451e76-92b2-11ef-a4d1-8dbaa06e16ad","last_name":"Egiazarian"},{"full_name":"Sinitsin, Anton","last_name":"Sinitsin","first_name":"Anton"},{"last_name":"Kuznedelev","full_name":"Kuznedelev, Denis","first_name":"Denis"},{"full_name":"Alistarh, Dan-Adrian","last_name":"Alistarh","id":"4A899BFC-F248-11E8-B48F-1D18A9856A87","orcid":"0000-0003-3650-940X","first_name":"Dan-Adrian"}],"OA_type":"gold","OA_place":"publisher","date_created":"2026-09-06T22:01:59Z","publication_identifier":{"isbn":["9798331338275"],"issn":["1049-5258"]},"abstract":[{"text":"Large Language Models (LLMs) have demonstrated the ability to tackle increasingly complex tasks through advanced reasoning, long-form content generation,\r\nand tool use. Solving these tasks often involves long inference-time computations.\r\nIn human problem solving, a common strategy to expedite work is collaboration: by\r\ndividing the problem into sub-tasks, exploring different strategies concurrently, etc.\r\nRecent research has shown that LLMs can also operate in parallel by implementing\r\nexplicit cooperation frameworks, such as voting mechanisms or the explicit creation of independent sub-tasks that can be executed in parallel. However, each of\r\nthese frameworks may not be suitable for all types of tasks, which can hinder their\r\napplicability. In this work, we propose a different design approach: we run LLM\r\n“workers” in parallel , allowing them to synchronize via a concurrently-updated\r\nattention cache and prompt these workers to decide how best to collaborate. Our\r\napproach allows the LLM instances to come up with their own collaboration strategy for the problem at hand, all the while “seeing” each other’s memory in the\r\nconcurrent KV cache. We implement this approach via Hogwild! Inference: a\r\nparallel LLM inference engine where multiple instances of the same LLM run in\r\nparallel with the same attention cache, with “instant” access to each other’s memory.1 Hogwild! Inference takes advantage of Rotary Position Embeddings (RoPE)\r\nto avoid recomputation while improving parallel hardware utilization. We find that\r\nmodern reasoning-capable LLMs can perform inference with shared Key-Value\r\ncache out of the box, without additional fine-tuning.","lang":"eng"}],"fulldoi":"https://doi.org/10.52202/085713-1551","acknowledgement":"We thank Vladimir Malinovskii for his help with brainstorming, helpful\r\nfeedback and suggesting future work directions. We also thank Philip Zmushko for proofreading.","das_tickbox":"0"},{"external_id":{"arxiv":["2505.15239"]},"oa_version":"Published Version","day":"02","intvolume":"        38","doi":"10.52202/085713-1450","publisher":"Neural Information Processing Systems Foundation","user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","publication_status":"published","language":[{"iso":"eng"}],"alternative_title":["Advances in Neural Information Processing Systems"],"acknowledged_ssus":[{"_id":"ScienComp"}],"year":"2025","department":[{"_id":"MaMo"},{"_id":"GradSch"},{"_id":"ChLa"}],"type":"conference","article_processing_charge":"No","publication":"39th Conference on Neural Information Processing Systems","conference":{"start_date":"2025-12-02","end_date":"2025-12-07","name":"NeurIPS: Neural Information Processing Systems","location":"San Diego, CA, United States"},"citation":{"ista":"Súkeník P, Lampert C, Mondelli M. 2025. Neural collapse is globally optimal in deep regularized ResNets and transformers. 39th Conference on Neural Information Processing Systems. NeurIPS: Neural Information Processing Systems, Advances in Neural Information Processing Systems, vol. 38, 48646–48677.","ieee":"P. Súkeník, C. Lampert, and M. Mondelli, “Neural collapse is globally optimal in deep regularized ResNets and transformers,” in <i>39th Conference on Neural Information Processing Systems</i>, San Diego, CA, United States, 2025, vol. 38, pp. 48646–48677.","ama":"Súkeník P, Lampert C, Mondelli M. Neural collapse is globally optimal in deep regularized ResNets and transformers. In: <i>39th Conference on Neural Information Processing Systems</i>. Vol 38. Neural Information Processing Systems Foundation; 2025:48646-48677. doi:<a href=\"https://doi.org/10.52202/085713-1450\">10.52202/085713-1450</a>","apa":"Súkeník, P., Lampert, C., &#38; Mondelli, M. (2025). Neural collapse is globally optimal in deep regularized ResNets and transformers. In <i>39th Conference on Neural Information Processing Systems</i> (Vol. 38, pp. 48646–48677). San Diego, CA, United States: Neural Information Processing Systems Foundation. <a href=\"https://doi.org/10.52202/085713-1450\">https://doi.org/10.52202/085713-1450</a>","short":"P. Súkeník, C. Lampert, M. Mondelli, in:, 39th Conference on Neural Information Processing Systems, Neural Information Processing Systems Foundation, 2025, pp. 48646–48677.","chicago":"Súkeník, Peter, Christoph Lampert, and Marco Mondelli. “Neural Collapse Is Globally Optimal in Deep Regularized ResNets and Transformers.” In <i>39th Conference on Neural Information Processing Systems</i>, 38:48646–77. Neural Information Processing Systems Foundation, 2025. <a href=\"https://doi.org/10.52202/085713-1450\">https://doi.org/10.52202/085713-1450</a>.","mla":"Súkeník, Peter, et al. “Neural Collapse Is Globally Optimal in Deep Regularized ResNets and Transformers.” <i>39th Conference on Neural Information Processing Systems</i>, vol. 38, Neural Information Processing Systems Foundation, 2025, pp. 48646–77, doi:<a href=\"https://doi.org/10.52202/085713-1450\">10.52202/085713-1450</a>."},"author":[{"id":"d64d6a8d-eb8e-11eb-b029-96fd216dec3c","last_name":"Súkeník","full_name":"Súkeník, Peter","first_name":"Peter"},{"first_name":"Christoph","orcid":"0000-0001-8622-7887","id":"40C20FD2-F248-11E8-B48F-1D18A9856A87","last_name":"Lampert","full_name":"Lampert, Christoph"},{"id":"27EB676C-8706-11E9-9510-7717E6697425","last_name":"Mondelli","full_name":"Mondelli, Marco","first_name":"Marco","orcid":"0000-0002-3242-7020"}],"OA_type":"free access","supplementarymaterial":"no","ddc":["000"],"main_file_link":[{"url":"https://doi.org/10.52202/085713-1450","open_access":"1"}],"researchdata_availability":"no","das_tickbox":"0","acknowledgement":"M. M. and P. S. are funded by the European Union (ERC, INF2\r\n, project number 101161364). Views\r\nand opinions expressed are however those of the author(s) only and do not necessarily reflect those\r\nof the European Union or the European Research Council Executive Agency. Neither the European\r\nUnion nor the granting authority can be held responsible for them. This research was supported\r\nby the Scientific Service Units (SSU) of ISTA through resources provided by Scientific Computing (SciComp).","fulldoi":"https://doi.org/10.52202/085713-1450","abstract":[{"text":"The empirical emergence of neural collapse—a surprising symmetry in the feature representations of the training data in the penultimate layer of deep neural\r\nnetworks—has spurred a line of theoretical research aimed at its understanding.\r\nHowever, existing work focuses on data-agnostic models or, when data structure is\r\ntaken into account, it remains limited to multi-layer perceptrons. Our paper fills\r\nboth these gaps by analyzing modern architectures in a data-aware regime: we\r\nprove that global optima of deep regularized transformers and residual networks\r\n(ResNets) with LayerNorm trained with cross entropy or mean squared error loss\r\nare approximately collapsed, and the approximation gets tighter as the depth grows.\r\nMore generally, we formally reduce any end-to-end large-depth ResNet or transformer training into an equivalent unconstrained features model, thus justifying its\r\nwide use in the literature even beyond data-agnostic settings. Our theoretical results\r\nare supported by experiments on computer vision and language datasets showing\r\nthat, as the depth grows, neural collapse indeed becomes more prominent.","lang":"eng"}],"publication_identifier":{"issn":["1049-5258"],"isbn":["9798331338275"]},"OA_place":"publisher","date_created":"2026-09-06T22:01:59Z","date_updated":"2026-09-10T08:11:46Z","corr_author":"1","quality_controlled":"1","_id":"22825","date_published":"2025-12-02T00:00:00Z","scopus_import":"1","arxiv":1,"month":"12","project":[{"name":"Inference in High Dimensions: Light-speed Algorithms and Information Limits","grant_number":"101161364","_id":"911e6d1f-16d5-11f0-9cad-c5c68c6a1cdf"}],"page":"48646-48677","oa":1,"title":"Neural collapse is globally optimal in deep regularized ResNets and transformers","status":"public","volume":38},{"publication":"39th Conference on Neural Information Processing Systems","article_processing_charge":"No","citation":{"apa":"Zakerinia, H., &#38; Lampert, C. (2025). Fast rate bounds for multi-task and meta-learning with different sample sizes. In <i>39th Conference on Neural Information Processing Systems</i> (Vol. 38, pp. 9062–9093). San Diego, CA, United States: Neural Information Processing Systems Foundation. <a href=\"https://doi.org/10.52202/085713-0278\">https://doi.org/10.52202/085713-0278</a>","short":"H. Zakerinia, C. Lampert, in:, 39th Conference on Neural Information Processing Systems, Neural Information Processing Systems Foundation, 2025, pp. 9062–9093.","chicago":"Zakerinia, Hossein, and Christoph Lampert. “Fast Rate Bounds for Multi-Task and Meta-Learning with Different Sample Sizes.” In <i>39th Conference on Neural Information Processing Systems</i>, 38:9062–93. Neural Information Processing Systems Foundation, 2025. <a href=\"https://doi.org/10.52202/085713-0278\">https://doi.org/10.52202/085713-0278</a>.","ieee":"H. Zakerinia and C. Lampert, “Fast rate bounds for multi-task and meta-learning with different sample sizes,” in <i>39th Conference on Neural Information Processing Systems</i>, San Diego, CA, United States, 2025, vol. 38, pp. 9062–9093.","ama":"Zakerinia H, Lampert C. Fast rate bounds for multi-task and meta-learning with different sample sizes. In: <i>39th Conference on Neural Information Processing Systems</i>. Vol 38. Neural Information Processing Systems Foundation; 2025:9062-9093. doi:<a href=\"https://doi.org/10.52202/085713-0278\">10.52202/085713-0278</a>","mla":"Zakerinia, Hossein, and Christoph Lampert. “Fast Rate Bounds for Multi-Task and Meta-Learning with Different Sample Sizes.” <i>39th Conference on Neural Information Processing Systems</i>, vol. 38, Neural Information Processing Systems Foundation, 2025, pp. 9062–93, doi:<a href=\"https://doi.org/10.52202/085713-0278\">10.52202/085713-0278</a>.","ista":"Zakerinia H, Lampert C. 2025. Fast rate bounds for multi-task and meta-learning with different sample sizes. 39th Conference on Neural Information Processing Systems. NeurIPS: Neural Information Processing Systems, Advances in Neural Information Processing Systems, vol. 38, 9062–9093."},"conference":{"location":"San Diego, CA, United States","start_date":"2025-12-02","name":"NeurIPS: Neural Information Processing Systems","end_date":"2025-12-07"},"publisher":"Neural Information Processing Systems Foundation","doi":"10.52202/085713-0278","acknowledged_ssus":[{"_id":"ScienComp"}],"alternative_title":["Advances in Neural Information Processing Systems"],"language":[{"iso":"eng"}],"publication_status":"published","user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","department":[{"_id":"GradSch"},{"_id":"ChLa"}],"year":"2025","type":"conference","oa_version":"Published Version","day":"02","intvolume":"        38","month":"12","oa":1,"page":"9062-9093","volume":38,"status":"public","title":"Fast rate bounds for multi-task and meta-learning with different sample sizes","corr_author":"1","quality_controlled":"1","date_updated":"2026-09-10T08:37:33Z","_id":"22824","scopus_import":"1","date_published":"2025-12-02T00:00:00Z","fulldoi":"https://doi.org/10.52202/085713-0278","das_tickbox":"0","acknowledgement":"This research was supported by the Scientific Service Units (SSU) of ISTA through resources provided by Scientific Computing (SciComp).","publication_identifier":{"issn":["1049-5258"],"isbn":["9798331338275"]},"abstract":[{"lang":"eng","text":"We present new fast-rate PAC-Bayesian generalization bounds for multi-task and\r\nmeta-learning in the unbalanced setting, i.e. when the tasks have training sets of\r\ndifferent sizes, as is typically the case in real-world scenarios. Previously, only\r\nstandard-rate bounds were known for this situation, while fast-rate bounds were\r\nlimited to the setting where all training sets are of equal size. Our new bounds\r\nare numerically computable as well as interpretable, and we demonstrate their\r\nflexibility in handling a number of cases where they give stronger guarantees\r\nthan previous bounds. Besides the bounds themselves, we also make conceptual\r\ncontributions: we demonstrate that the unbalanced multi-task setting has different\r\nstatistical properties than the balanced situation, specifically that proofs from\r\nthe balanced situation do not carry over to the unbalanced setting. Additionally,\r\nwe shed light on the fact that the unbalanced situation allows two meaningful\r\ndefinitions of multi-task risk, depending on whether all tasks should be considered\r\nequally important or if sample-rich tasks should receive more weight than samplepoor ones."}],"OA_place":"publisher","date_created":"2026-09-06T22:01:59Z","OA_type":"free access","author":[{"orcid":"0009-0007-3977-6462","first_name":"Hossein","full_name":"Zakerinia, Hossein","last_name":"Zakerinia","id":"653bd8b6-f394-11eb-9cf6-c0bbf6cd78d4"},{"first_name":"Christoph","orcid":"0000-0001-8622-7887","last_name":"Lampert","id":"40C20FD2-F248-11E8-B48F-1D18A9856A87","full_name":"Lampert, Christoph"}],"supplementarymaterial":"yes","researchdata_availability":"no","main_file_link":[{"open_access":"1","url":"https://doi.org/10.52202/085713-0278"}]},{"oa_version":"Published Version","day":"02","intvolume":"        38","article_processing_charge":"No","publication":"39th Conference on Neural Information Processing Systems","conference":{"end_date":"2025-12-07","name":"NeurIPS: Neural Information Processing Systems","start_date":"2025-12-02","location":"San Diego, CA, United States"},"citation":{"mla":"Cyffers, Edwige Audrey Lucienne. “Setting ε Is Not the Issue in Differential Privacy.” <i>39th Conference on Neural Information Processing Systems</i>, vol. 38, Neural Information Processing Systems Foundation, 2025, pp. 159016–28, doi:<a href=\"https://doi.org/10.52202/085713-4798\">10.52202/085713-4798</a>.","short":"E.A.L. Cyffers, in:, 39th Conference on Neural Information Processing Systems, Neural Information Processing Systems Foundation, 2025, pp. 159016–159028.","chicago":"Cyffers, Edwige Audrey Lucienne. “Setting ε Is Not the Issue in Differential Privacy.” In <i>39th Conference on Neural Information Processing Systems</i>, 38:159016–28. Neural Information Processing Systems Foundation, 2025. <a href=\"https://doi.org/10.52202/085713-4798\">https://doi.org/10.52202/085713-4798</a>.","apa":"Cyffers, E. A. L. (2025). Setting ε is not the issue in differential privacy. In <i>39th Conference on Neural Information Processing Systems</i> (Vol. 38, pp. 159016–159028). San Diego, CA, United States: Neural Information Processing Systems Foundation. <a href=\"https://doi.org/10.52202/085713-4798\">https://doi.org/10.52202/085713-4798</a>","ama":"Cyffers EAL. Setting ε is not the issue in differential privacy. In: <i>39th Conference on Neural Information Processing Systems</i>. Vol 38. Neural Information Processing Systems Foundation; 2025:159016-159028. doi:<a href=\"https://doi.org/10.52202/085713-4798\">10.52202/085713-4798</a>","ieee":"E. A. L. Cyffers, “Setting ε is not the issue in differential privacy,” in <i>39th Conference on Neural Information Processing Systems</i>, San Diego, CA, United States, 2025, vol. 38, pp. 159016–159028.","ista":"Cyffers EAL. 2025. Setting ε is not the issue in differential privacy. 39th Conference on Neural Information Processing Systems. NeurIPS: Neural Information Processing Systems, Advances in Neural Information Processing Systems, vol. 38, 159016–159028."},"user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","language":[{"iso":"eng"}],"publication_status":"published","alternative_title":["Advances in Neural Information Processing Systems"],"doi":"10.52202/085713-4798","publisher":"Neural Information Processing Systems Foundation","type":"conference","year":"2025","department":[{"_id":"ChLa"}],"abstract":[{"lang":"eng","text":"This position paper argues that setting the privacy budget in differential privacy should not be viewed as an important limitation of differential privacy compared to alternative methods for privacy-preserving machine learning. The so-called problem of interpreting the privacy budget is often presented as a major hindrance to the wider adoption of differential privacy in real-world deployments and is sometimes used to promote alternative mitigation techniques for data protection. We believe this misleads decision-makers into choosing unsafe methods. We argue that the difficulty in interpreting privacy budgets does not stem from the definition of differential privacy itself, but from the intrinsic difficulty of estimating privacy risks in context, a challenge that any rigorous method for privacy risk assessment face. Moreover, we claim that any sound method for estimating privacy risks should, given the current state of research, be expressible within the differential privacy framework or justify why it cannot."}],"publication_identifier":{"isbn":["9798331338275"],"issn":["1049-5258"]},"das_tickbox":"0","acknowledgement":"This research was funded in whole by the Austrian Science Fund (FWF) 10.55776/COE12. I thank\r\nAlberto Naibo for the discussion and his invitation to \"Logique, droit, IA : penser les algorithmes\"\r\ncolloquium. I thank Simone Bombari for the discussions. I thank EurIPS for the active poster session.","fulldoi":"https://doi.org/10.52202/085713-4798","OA_place":"publisher","date_created":"2026-09-06T22:01:59Z","supplementarymaterial":"yes","OA_type":"gold","author":[{"id":"20d4c299-977a-11ef-ae55-98b15ac64a57","last_name":"Cyffers","full_name":"Cyffers, Edwige Audrey Lucienne","first_name":"Edwige Audrey Lucienne"}],"researchdata_availability":"no","main_file_link":[{"url":"https://doi.org/10.52202/085713-4798","open_access":"1"}],"month":"12","title":"Setting ε is not the issue in differential privacy","status":"public","volume":38,"page":"159016-159028","oa":1,"date_updated":"2026-09-10T08:44:14Z","quality_controlled":"1","corr_author":"1","date_published":"2025-12-02T00:00:00Z","scopus_import":"1","_id":"22823"},{"researchdata_availability":"no","main_file_link":[{"url":"https://doi.org/10.52202/085713-3163","open_access":"1"}],"supplementarymaterial":"yes","OA_type":"free access","author":[{"first_name":"Roman","last_name":"Garipov","full_name":"Garipov, Roman"},{"first_name":"Fedor","last_name":"Velikonivtsev","full_name":"Velikonivtsev, Fedor"},{"full_name":"Ermakov, Ivan","last_name":"Ermakov","first_name":"Ivan"},{"full_name":"Svirschevski, Ruslan","last_name":"Svirschevski","first_name":"Ruslan"},{"id":"77451e76-92b2-11ef-a4d1-8dbaa06e16ad","last_name":"Egiazarian","full_name":"Egiazarian, Vage","first_name":"Vage"},{"last_name":"Ryabinin","full_name":"Ryabinin, Max","first_name":"Max"}],"OA_place":"publisher","date_created":"2026-09-06T22:02:00Z","publication_identifier":{"isbn":["9798331338275"],"issn":["1049-5258"]},"abstract":[{"text":"We introduce AutoJudge, a method that accelerates large language model (LLM) inference with task-specific lossy speculative decoding. Instead of matching the original model output distribution token-by-token, we identify the generated tokens that affect the downstream quality of the response, relaxing the distribution match guarantee so that the \"unimportant\" tokens can be generated faster. Our approach relies on a semi‑greedy search algorithm to test which of the mismatches between target and draft models should be corrected to preserve quality and which ones may be skipped. We then train a lightweight classifier based on existing LLM embeddings to predict, at inference time, which mismatching tokens can be safely accepted without compromising the final answer quality. We evaluate AutoJudge with multiple draft/target model pairs on mathematical reasoning and programming benchmarks, achieving significant speedups at the cost of a minor accuracy reduction. Notably, on GSM8K with the Llama 3.1 70B target model, our approach achieves up to \r\n≈\r\n2\r\n×\r\n speedup \\textit{over speculative decoding} at the cost of a \r\n≤\r\n1\r\n%\r\n drop in accuracy. When applied to the LiveCodeBench benchmark, AutoJudge automatically detects programming-specific important tokens, accepting \r\n≥\r\n25\r\n tokens per speculation cycle at a\r\n \r\n2\r\n%\r\n drop in Pass@1. Our approach requires no human annotation and is easy to integrate with modern LLM inference frameworks.","lang":"eng"}],"fulldoi":"https://doi.org/10.52202/085713-3163","acknowledgement":"We would like to express our sincere gratitude to Denis Mazur for his valuable contributions to the\r\nimplementation of API calls used in Algorithm 1 and for supporting inference with the Llama 405B\r\nmodel. We are also thankful for his positive influence on the overall atmosphere and team morale\r\nthroughout the course of this project.","das_tickbox":"0","scopus_import":"1","date_published":"2025-12-02T00:00:00Z","_id":"22830","quality_controlled":"1","date_updated":"2026-09-10T11:08:33Z","status":"public","volume":38,"title":"AutoJudge: Judge decoding without manual annotation","oa":1,"page":"104904-104941","month":"12","arxiv":1,"external_id":{"arxiv":["2504.20039"]},"intvolume":"        38","oa_version":"Published Version","day":"02","type":"conference","year":"2025","department":[{"_id":"DaAl"}],"alternative_title":["Advances in Neural Information Processing Systems"],"publication_status":"published","user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","language":[{"iso":"eng"}],"publisher":"Neural Information Processing Systems Foundation","doi":"10.52202/085713-3163","citation":{"ieee":"R. Garipov, F. Velikonivtsev, I. Ermakov, R. Svirschevski, V. Egiazarian, and M. Ryabinin, “AutoJudge: Judge decoding without manual annotation,” in <i>39th Annual Conference on Neural Information Processing Systems</i>, San Diego, CA, United States, 2025, vol. 38, pp. 104904–104941.","ama":"Garipov R, Velikonivtsev F, Ermakov I, Svirschevski R, Egiazarian V, Ryabinin M. AutoJudge: Judge decoding without manual annotation. In: <i>39th Annual Conference on Neural Information Processing Systems</i>. Vol 38. Neural Information Processing Systems Foundation; 2025:104904-104941. doi:<a href=\"https://doi.org/10.52202/085713-3163\">10.52202/085713-3163</a>","chicago":"Garipov, Roman, Fedor Velikonivtsev, Ivan Ermakov, Ruslan Svirschevski, Vage Egiazarian, and Max Ryabinin. “AutoJudge: Judge Decoding without Manual Annotation.” In <i>39th Annual Conference on Neural Information Processing Systems</i>, 38:104904–41. Neural Information Processing Systems Foundation, 2025. <a href=\"https://doi.org/10.52202/085713-3163\">https://doi.org/10.52202/085713-3163</a>.","apa":"Garipov, R., Velikonivtsev, F., Ermakov, I., Svirschevski, R., Egiazarian, V., &#38; Ryabinin, M. (2025). AutoJudge: Judge decoding without manual annotation. In <i>39th Annual Conference on Neural Information Processing Systems</i> (Vol. 38, pp. 104904–104941). San Diego, CA, United States: Neural Information Processing Systems Foundation. <a href=\"https://doi.org/10.52202/085713-3163\">https://doi.org/10.52202/085713-3163</a>","short":"R. Garipov, F. Velikonivtsev, I. Ermakov, R. Svirschevski, V. Egiazarian, M. Ryabinin, in:, 39th Annual Conference on Neural Information Processing Systems, Neural Information Processing Systems Foundation, 2025, pp. 104904–104941.","mla":"Garipov, Roman, et al. “AutoJudge: Judge Decoding without Manual Annotation.” <i>39th Annual Conference on Neural Information Processing Systems</i>, vol. 38, Neural Information Processing Systems Foundation, 2025, pp. 104904–41, doi:<a href=\"https://doi.org/10.52202/085713-3163\">10.52202/085713-3163</a>.","ista":"Garipov R, Velikonivtsev F, Ermakov I, Svirschevski R, Egiazarian V, Ryabinin M. 2025. AutoJudge: Judge decoding without manual annotation. 39th Annual Conference on Neural Information Processing Systems. NeurIPS: Neural Information Processing Systems, Advances in Neural Information Processing Systems, vol. 38, 104904–104941."},"conference":{"location":"San Diego, CA, United States","name":"NeurIPS: Neural Information Processing Systems","end_date":"2025-12-07","start_date":"2025-12-02"},"publication":"39th Annual Conference on Neural Information Processing Systems","article_processing_charge":"No"},{"conference":{"location":"San Diego, CA, United States","name":"NeurIPS: Neural Information Processing Systems"},"citation":{"ama":"Kalinin N, Upadhyay J, Lampert C. Continual release moment estimation with differential privacy. In: <i>Advances in Neural Information Processing Systems</i>. Vol 38. Neural Information Processing Systems Foundation; 2025:64002-64047. doi:<a href=\"https://doi.org/10.52202/085713-1924\">10.52202/085713-1924</a>","ieee":"N. Kalinin, J. Upadhyay, and C. Lampert, “Continual release moment estimation with differential privacy,” in <i>Advances in Neural Information Processing Systems</i>, San Diego, CA, United States, 2025, vol. 38, pp. 64002–64047.","apa":"Kalinin, N., Upadhyay, J., &#38; Lampert, C. (2025). Continual release moment estimation with differential privacy. In <i>Advances in Neural Information Processing Systems</i> (Vol. 38, pp. 64002–64047). San Diego, CA, United States: Neural Information Processing Systems Foundation. <a href=\"https://doi.org/10.52202/085713-1924\">https://doi.org/10.52202/085713-1924</a>","short":"N. Kalinin, J. Upadhyay, C. Lampert, in:, Advances in Neural Information Processing Systems, Neural Information Processing Systems Foundation, 2025, pp. 64002–64047.","chicago":"Kalinin, Nikita, Jalaj Upadhyay, and Christoph Lampert. “Continual Release Moment Estimation with Differential Privacy.” In <i>Advances in Neural Information Processing Systems</i>, 38:64002–47. Neural Information Processing Systems Foundation, 2025. <a href=\"https://doi.org/10.52202/085713-1924\">https://doi.org/10.52202/085713-1924</a>.","mla":"Kalinin, Nikita, et al. “Continual Release Moment Estimation with Differential Privacy.” <i>Advances in Neural Information Processing Systems</i>, vol. 38, Neural Information Processing Systems Foundation, 2025, pp. 64002–47, doi:<a href=\"https://doi.org/10.52202/085713-1924\">10.52202/085713-1924</a>.","ista":"Kalinin N, Upadhyay J, Lampert C. 2025. Continual release moment estimation with differential privacy. Advances in Neural Information Processing Systems. NeurIPS: Neural Information Processing Systems, Advances in Neural Information Processing Systems, vol. 38, 64002–64047."},"article_processing_charge":"No","publication":"Advances in Neural Information Processing Systems","department":[{"_id":"GradSch"},{"_id":"ChLa"}],"year":"2025","type":"conference","doi":"10.52202/085713-1924","publisher":"Neural Information Processing Systems Foundation","language":[{"iso":"eng"}],"user_id":"2DF688A6-F248-11E8-B48F-1D18A9856A87","publication_status":"published","alternative_title":["Advances in Neural Information Processing Systems"],"intvolume":"        38","day":"02","oa_version":"Published Version","page":"64002-64047","oa":1,"title":"Continual release moment estimation with differential privacy","status":"public","volume":38,"month":"12","_id":"22828","date_published":"2025-12-02T00:00:00Z","scopus_import":"1","date_updated":"2026-09-10T11:17:06Z","corr_author":"1","quality_controlled":"1","OA_place":"publisher","date_created":"2026-09-06T22:02:00Z","das_tickbox":"0","acknowledgement":"We thank Monika Henzinger for her valuable feedback and insightful discussions on earlier versions\r\nof this draft. We are also grateful to Mher Safaryan for his contributions to discussions on DP-Adam.\r\nAdditionally, we thank Ryan McKenna for suggesting Joint Clipping as a baseline.\r\nJalaj Upadhyay’s research was funded by the Rutgers Decanal Grant no. 302918, NSF CNS 2433628,\r\nGoogle Research Scholar Award, and Google Seed Fund Grant. A part of this work was done while\r\nvisiting the Institute of Science and Technology Austria (ISTA).\r\nNikita Kalinin’s research was funded in part by the Austrian Science Fund (FWF) [10.55776/COE12]","fulldoi":"https://doi.org/10.52202/085713-1924","abstract":[{"text":"We propose Joint Moment Estimation (JME), a method for continually and privately estimating both the first and second moments of a data stream with reduced noise compared to naive approaches. JME supports the matrix mechanism and exploits a joint sensitivity analysis to identify a privacy regime in which the second-moment estimation incurs no additional privacy cost, thereby improving accuracy while maintaining privacy. We demonstrate JME’s effectiveness in two applications: estimating the running mean and covariance matrix for Gaussian density estimation and model training with DP-Adam.","lang":"eng"}],"publication_identifier":{"issn":["1049-5258"],"isbn":["9798331338275"]},"researchdata_availability":"no","main_file_link":[{"url":"https://doi.org/10.52202/085713-1924","open_access":"1"}],"OA_type":"free access","author":[{"id":"4b14526e-14d2-11ed-ba64-c14c9553d137","last_name":"Kalinin","full_name":"Kalinin, Nikita","first_name":"Nikita"},{"last_name":"Upadhyay","full_name":"Upadhyay, Jalaj","first_name":"Jalaj"},{"last_name":"Lampert","id":"40C20FD2-F248-11E8-B48F-1D18A9856A87","full_name":"Lampert, Christoph","first_name":"Christoph","orcid":"0000-0001-8622-7887"}],"supplementarymaterial":"no"}]