跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1865新增论文
+1473其中 AI 相关
+176已有论文修订
+95新增 AI 导读

arXiv 最近完成扫描:2026-11-17 04:58 · 全站最近入库:2026-11-17 17:12各信源进度 →

统计口径与数据新鲜度

统计 2026-11-16 17:28 — 2026-11-17 17:28(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-17 04:58最近发现新增版本:2026-11-17 17:12库内论文最新更新时间:2026-11-17 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

25,692 已收录论文21,414 AI 相关7,613 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 227 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-17 04:57(北京时间)
新论文 · v1arXiv:2611.21584

CertiSum: Bitwise-Identical MXFP4 Training Across Platforms and Parallelism Layouts

Steve Rhyner, Yubo Gao, Stefan de Lasa, Sankeerth Durvasula, Nandita Vijaykumar

原文摘要

Training with MXFP4, a 4-bit microscaling format, is non-deterministic: identical inputs can produce different results across parallelism layouts and GPU vendors. These differences complicate debugging and training audits and can cause training and inference engines to disagree. We present CertiSum to make MXFP4 training bitwise invariant across runs, parallelism layouts, and GPU vendors. Its key insight is that native FP32 accumulation of MXFP4 products is often already exact for every summation order. CertiSum collects statistics during quantization to certify native matrix multiplication results, recomputes uncertified outputs exactly, and uses exact kernels directly where certification rarely succeeds. CertiSum keeps quantization blocks unchanged across layouts and merges partial sums without rounding, including in attention and weight-gradient computation. In 200-step Qwen3-8B-Base experiments, CertiSum preserves training states bit for bit across AMD MI350X and NVIDIA GB200 and across changes in data, tensor, and context parallelism on MI350X. Across three 8B--14B models on FineWeb-Edu using MI350X, it adds 11.8--60.5% to training-step latency relative to its non-exact MXFP4 counterpart at the larger evaluated input shapes, with speedups of up to $1.24\times$ over BF16 and $1.70\times$ over deterministic BF16.

cs.LG · cs.DC

新论文 · v1arXiv:2611.21570

Trustworthy Numerical Software in the Age of AI: What Would Turing Do?

Kimon Fountoulakis, Maksim Melnichenko, Muhammad Fetrat Qharabagh, Georgios Alexandros Georgantzas, George Spyrou, Oleg Balabanov, Vasileios Georgiou, Riley Murray 等 9 位作者

原文摘要

AI coding agents can now produce numerical software faster than human experts can review it, raising serious concerns about usability and reliability. These and other AI tools are already transforming mathematics and software, and they will soon transform mathematical software. These concerns motivated us to assess AI-generated numerical linear algebra (NLA) kernels, which we did through an audit of a recent public GPU QR competition. All five leading kernels passed the competition's accuracy tests, and the winning kernel achieved a 107-fold speedup over the competition's baseline, which calls NVIDIA's QR routine. Yet each failed at least once on new inputs, even those from the competition's generator, under the same checker, while the baseline passed them all. In light of this, we want to develop reusable formalizations of numerical analysis methods, which remain sparsely represented in existing formal libraries. We describe VSCL, our Lean library of mathematical foundations for numerical computing. We also propose VERNA to connect these tasks by using formalized results to guide kernel assessment against explicit specifications. Drawing on Turing's work on rounding-error analysis and program checking, we discuss design principles for developing trustworthy numerical software when direct human checking is impractical.

math.NA · cs.MS

版本更新 · v2arXiv:2611.06207

Performance Portability of Out-of-Kernel Optimizations on GPUs: A Molecular Docking Case Study

Gianmarco Accordi, Leonardo Beltrame, Theresa Pollinger, Jens Domke, Davide Gadioli, Gianluca Palermo

原文摘要

Achieving performance portability is challenging for irregular scientific applications built as repeated multi-kernel GPU pipelines, because throughput depends on out-of-kernel decisions as much as on kernel efficiency. Using molecular docking as a representative case study, we propose an out-of-kernel tuning methodology that treats batching policy, worker concurrency, and per-worker memory allocation as explicit optimization parameters, replacing backend-dependent occupancy-driven choices with a more portable execution strategy based on explicit resource bounds. We implement this methodology in the muDock mini-app under a manager/worker execution model where multiple host workers share a GPU through independent streams. We then evaluate the methodology on NVIDIA, AMD, and Intel GPUs across CUDA, HIP, and SYCL implementations. Across all evaluated platform/backend pairs, the memory-bounded formulation exceeds the peak throughput of the occupancy-guided baseline. The fully tuned methodology improves the end-to-end throughput on all tested configurations, achieving an average speedup of 1.53x over the baseline.

cs.DC · cs.PF

新论文 · v1arXiv:2611.21274

SURMA: Efficient Fine-Grained Communication Offload for SuperPOD

Haoran Shang, Weifeng Zhang, Mengzhu Li, Huizheng Wang, Sihan Guan, Xinyang Huang, Bowen Liu, Ran Shu 等 17 位作者

原文摘要

GPU SuperPODs underpin large language model (LLM) serving, but growing bandwidth and round-trip times make inter-GPU communication consume increasing compute resources. SM-driven remote memory access occupies shared memory or registers for in-flight data and competes for SM-to-NoC bandwidth. Dedicated engines can free these resources, but must handle fine-grained transfers, down to a 576\,B KV-cache entry. NVIDIA's CopyEngine uses less than 1\% of NVLink bandwidth at 7\,KB granularity. We present SURMA, an on-die Scale-Up Remote Memory Access engine. Adapting RDMA requires overcoming slow doorbell submission, excessive per-application connection state, head-of-line blocking of completion notifications on shared connections, and costly ordering for multipath transfers. SURMA uses a credit-based shared send queue to reduce submission latency from 1.2\,$μ$s to 400\,ns; fused write-with-atomic semantics to share connections while independently notifying each transaction; and Network Per Path Ordering to ensure data arrives before the completion flag without Sender or Receiver Ordering. We implement SURMA in RTL on a two-FPGA prototype and evaluate it with two 8-GPU NVLink servers. At 512\,B granularity, SURMA saturates attainable bandwidth with a single SM, whereas NVLink needs over 14\% of the SMs. SURMA occupies under 0.5\% of GPU area and accelerates inference on four models by 1.05--1.26$\times$ over NVLink.

cs.AR

版本更新 · v2arXiv:2611.03330

The SIMPLE Law for Energy Cost of LLM Inference

Hiari Pizzini Cavagna, Andrea Proia, Daniele Cesarini, Zeynep Kiziltan, Andrea Bartolini

原文摘要

Generative AI is billed per token, yet tokens are not equally expensive. Each token enters the KV-cache and raises the cost of those that follow, so energy grows non-linearly along a sequence whose tokens look identical. No analytical model states what a token costs without per-LLM fitting, and no law names the factors that set it. From closed-form expressions of LLM inference's computational and memory-access complexity, we derive the SIMPLE Inference Model to Predict LLMs Energy (SIMPLE), a four-parameter law that separates prefill from decode, and decode into KV-cache and next-token computation cost. Over traces of 13 dense LLMs from several families on NVIDIA H100 GPUs, it attains R2 = 0.969 and 11.97% MAPE, 3.1-3.5x below the SOTA, without per-LLM calibration. SIMPLE explains as well as predicts: prefill energy grows quadratically with hidden dimension, so larger LLMs suffer most from long inputs; and higher GQA KV grouping reduces the per-token energy linearly relative to standard MHA. Increasing the reasoning token budget pushes output length beyond the analytically derived energy minimum, driving super-linear energy growth; accuracy rises but energy rises faster, so Intelligence per Joule peaks at an intermediate budget and falls by up to 3.14x across a 4.0x range. Moving to agentic serving raises energy per request by 13.0x, as both prefill and KV-cache terms grow with input length; prefix caching removes most of the prefill cost (-93%), but not KV-cache cost, so even a cached agentic request costs 4.8x more.

cs.AI · cs.LG · cs.PF

新论文 · v1arXiv:2611.20864

LiDSeg: Linear DehazeFormer Guidance for Efficient Semantic Segmentation in Adverse Weather Conditions

Yobsan Bayisa Leta, Christos Kyrkou, Maria K. Michael, Theocharis Theocharides

原文摘要

Robust semantic segmentation in adverse weather is essential for safety-critical autonomous systems. Existing restore-then-segment pipelines introduce latency and prioritize pixel-level restoration over semantic discriminability, while domain adaptation methods often face representation bottlenecks and computational complexity. We introduce LiDSeg, a framework that uses reconstruction-informed guidance to improve segmentation under adverse weather. During training, a lightweight guidance encoder is jointly optimized with the segmentation branch using segmentation and auxiliary reconstruction objectives. At inference, the reconstruction decoder is discarded, while the guidance encoder is retained to extract weather-robust features. A lightweight gated fusion module integrates these features into the segmentation branch, preserving clear-weather accuracy while improving adverse-weather robustness. On Foggy Cityscapes, LiDSeg limits the clear-to-dense-fog mIoU degradation to 2.5-2.8 percentage points, reducing the baseline performance drop by more than fivefold. On the official ACDC test server, LiDSeg-PIDNet achieves 75.07% pooled all-condition mIoU, comparable to supervised HRNet-W48 at 75.0%, while remaining substantially smaller. The gains generalize zero-shot to Foggy Driving and Foggy Zurich without training on real-fog images. LiDSeg adds only 1.9-3.1M parameters across several models, with total model sizes of 5.2-39.5M parameters and no inference-time reconstruction overhead. It delivers real-time inference on the NVIDIA Jetson AGX Orin. Code: https://github.com/yobsanb/LiDSeg

cs.CV

新论文 · v1arXiv:2611.20771

Nemotron 3.5 Bolt E4B 4GB: An Open and Efficient Small Language Model for Fast Conversational Agents

NVIDIA, :, Adi Margolin, Aditya Malte, Aishwarya Chhabra, Alex Hodson, Ameya Sunil Mahabaleshwarkar, Ambrish Dantrey 等 113 位作者

原文摘要

We introduce Nemotron 3.5 Bolt E4B 4GB, an open and efficient small language model further post-trained from Gemma-4-E4B-it and optimized for speed and intelligence per GB of memory used for fast conversational agents. The model carries forward Gemma's strong conversational capabilities and further improves multi-turn tool use, instruction following, and input grounding for fast, local voice/device assistants, interactive personas, and embedded/Edge interfaces. The model combines natural conversation with multi-turn agentic capabilities to power voice/device assistants, interactive characters and robotic/IoT interfaces running on GeForce RTX, Jetson, and RTX/DGX Spark. Experiments on 21 benchmarks show that Nemotron 3.5 Bolt E4B 4GB achieves the highest average accuracy while generating the fewest output tokens in both reasoning and non-reasoning modes, compared with Gemma-4-E4B-it and Qwen3.5-4B. The model is released with a DFlash drafter. With DFlash, Bolt E4B achieves $2.74\times$ the per-user decode speed of inference without speculative decoding at concurrency~1, and $1.21\times$ the generation throughput of Gemma-4-E4B-it with MTP at batch size 1.

cs.CL · cs.AI · cs.LG

新论文 · v1arXiv:2611.20501

A Testing Method for Detector Error Models in Quantum Error Correction Compilers

Junjie Luo, Yasunari Suzuki

原文摘要

A quantum error correction (QEC) compiler produces, from a code and a noise model, both the physical circuit to execute and a detector error model (DEM). The DEM lists the probability of each error mechanism and the measurements it flips. Decoders consume the DEM without verification, so a defect that describes the noise incorrectly calibrates the decoder against wrong rates and silently degrades the logical error rate. Existing techniques for testing quantum software judge correctness by crashes or by statistics of sampled outputs. DEM defects that only distort rates lie below what sampling can resolve at the error rates where QEC operates. We present DEMVal, a testing method that validates each model the compiler emits against a reconstructed reference, with rates obtained from outside the compiler under test. It compares the two models exactly, so its sensitivity is independent of the physical error rate. The reference can be built in three ways that jointly cover the pipeline from source program to emitted model. Applied to NVIDIA's CUDA-Q QEC, Riverlane's Deltakit, and the deq component of Microsoft's QDK-EC, DEMVal found six defects, all confirmed by the developers and fixed. Sampling could have exposed only the four defects that drop, merge, or misplace mechanisms, and exposing all four requires millions of samples per experiment. Even analytically optimal moment tests miss the two rate-level defects, while exact model comparison detects all six with no samples. To our knowledge, DEMVal is the first systematic testing method for the error-model semantics of QEC compilers.

cs.SE

版本更新 · v2arXiv:2609.06635

From JPEG to CUDA Batch: Measuring RGB Input Pipeline Throughput

Vladimir Iglovikov

原文摘要

We measure how many ready batches per second eight image input pipelines deliver to an NVIDIA L4. Each pipeline starts with RGB JPEG files and delivers synchronized CUDA float16 batches of 256 images. The benchmark contains 1314 measurements across 73 recipes. On the 34 recipes shared by all eight paths, DALI delivered a median of 19.7 batches/s and AlbumentationsX (AX) delivered 17.1. DALI was fastest on 34 of those recipes. Median peak process GPU memory was 1,872 MiB for DALI and 404 MiB for AX.

cs.PF · cs.CV

新论文 · v1arXiv:2611.20289

Benchmarking LLM Inference on Desktop-Scale NVIDIA Grace Blackwell Clusters

Kevin Bhimani, Christian Escobar, Daniel Mihovch, Rudolf Eigenmann, Sunita Chandrasekaran

原文摘要

Two NVIDIA DGX Spark GB10 systems connect directly over 200 GbE to serve larger language models and increase inference capacity, but the benefits depend on how they are deployed. We benchmark 17 open-weight checkpoints spanning dense and mixture-of-experts architectures, 8B to 235B parameters and four weight precisions on one and two RDMA-connected systems. Using vLLM, we compare tensor parallelism (TP=2), pipeline parallelism (PP=2) and a two-replica baseline at concurrency 1 to 64, with three independent launches per configuration. TP=2 delivers a median 1.59x peak-throughput gain over one Spark, versus 1.39x for PP=2 and has higher throughput in 93 of 96 matched comparisons with separated 95% Welch intervals. It reduces unloaded inter-token latency to 0.52x to 0.73x of one Spark, whereas PP=2 provides little reduction. TP=2 also achieves the highest goodput under a strict 100 ms inter-token latency objective wherever both strategies are supported and the objective is met. For models fitting on one Spark, TP=2 generally favors low-concurrency serving, while replication offers higher throughput at high concurrency. One Spark uses less GPU energy per output token at matched concurrency. Across three precision ladders, NVFP4 reduces unloaded latency more than adding a second Spark at BF16. NCCL microbenchmarks and comparisons with an earlier TCP campaign suggest that collective communication overhead contributes to the advantage of TP=2. These results provide deployment guidance for desktop-scale inference.

cs.DC · cs.AI

版本更新 · v2arXiv:2609.29067

TileBench: A Controlled Benchmark for Performance Evaluation and Bottleneck Diagnosis of Tile-Based Programming Models

Bowen Cui, Zhongchun Zhou, Hao Wu, Tejas Ramesh, Junyu Yin, Jialiang Gu, Keren Zhou

原文摘要

Tile-based programming models, such as Triton and cuTile, aim to simplify high-performance kernel development, but their practical performance, tuning behavior, and usability remain difficult to compare systematically. We present TileBench, a controlled benchmark for evaluating Triton and cuTile on NVIDIA B200 GPUs under matched operator semantics and comparable implementation structures. TileBench contains 45 operators covering diverse AI-kernel patterns and memory/computation behaviors. Each task provides a PyTorch reference, verified Triton and cuTile implementations, standardized data-types (dtype) and input-size sweeps, default and autotuned configurations, roofline-based metrics, and profiling-guided diagnosis. Our evaluation shows that performance gaps are workload-dependent: cuTile excels on a small cluster of Tensor-Core/TMA-friendly kernels, while Triton is stronger on many irregular, streaming, and bandwidth-bound operators. We further evaluate LLM-generated cuTile and Triton kernels and find that Triton is consistently more token-efficient than cuTile under the same iterative refinement protocol. TileBench is publicly available at https://github.com/Deep-Learning-Profiling-Tools/Tilebench.

cs.PF · cs.PL

新论文 · v1arXiv:2611.19922

Lao Siji: Adapting openpilot to Closed-Loop End-to-End Driving in Neural Reconstruction

Chengzhi Gao, Yongzhi Liu, Jinchang Xu, Weichao Zhuang, Guodong Yin

原文摘要

We present Lao Siji, an end-to-end driving policy that ranks third in the 2026 NVIDIA AlpaSim E2E Driving Challenge, ahead of most entries built on foundation models. Instead of training from scratch on a general-purpose vision backbone, we adapt openpilot, an open-source L2 production model. We freeze its vision encoder and fine-tune the policy layers with an ego-state adapter and a route memory. On closed-loop rollouts in neural-reconstructed scenes, we supervise the policy with a privileged map teacher for lateral self-correction and a body-extent loss for road and obstacle clearance, while distilling the base model's remaining outputs to retain car following. Ablations show that self-correction samples carry the gain, while the clearance loss substitutes for them where no teacher exists. Longitudinal safety instead rests on the base model's lead-vehicle estimate behind a rule-based speed layer. Longer training, imitating human logs, and human-log speed targets yield either no measurable gain or adverse effects. Without further training, Lao Siji also places among the leaders on HUGSIM and Odyssey. The final training takes only one GPU-hour and shifts the weights by 3.9% from the shipped model. Code will be released after review.

cs.RO · cs.CV

新论文 · v1arXiv:2611.19512

Sada-ASR: A Production Multidialectal Arabic Speech Recognition System

Mahmoud Khaled, Mohammed Abd-Elnaser, Mahmoud Mohsen, Hazem Ahmed, Abdelrahman Hafez

原文摘要

Callers to Arabic customer-service lines speak their own dialect, while speech recognizers are built and evaluated mainly on Modern Standard Arabic (MSA). A recognizer that mishears the dialect breaks the call before anything else can help. We present Sada-ASR, the speech recognizer of EchoNova's production Arabic voice agent. On a 120-hour test set of MSA and eight dialects, it has the lowest word and character error rates of twelve systems, against eleven from Microsoft, ElevenLabs, Google, OpenAI, Speechmatics, IFM, Cohere, NVIDIA and Audar: 11.69% word error rate against 13.51% for the strongest, Microsoft's MAI-Transcribe-2 (a lead of 0.74 points on clips unaffected by that system's fixed number and script conventions), and 15.52% against 16.96% averaged over varieties. Its largest margins are on Omani (4.30 points), Egyptian (3.78) and Saudi (2.19) speech; it is level on Syrian and loses on Lebanese (24.96% against 23.05%), mostly because it writes Lebanese words in MSA forms and is less verbatim. On six public test sets its average word error rate under the leaderboard's current scoring is 17.17%, below every published figure we found, among them Falcon-ASR (20.92%), Jais-ASR (22.85%) and Audar-ASR-V1-Turbo (23.17%), though none of these states that it used that scoring, so the comparison is indicative. We describe the data pipeline and methodology behind Sada-ASR and release a 20-hour subset of the test set with our scoring code.

cs.CL · cs.SD · eess.AS

新论文 · v1arXiv:2611.19475

StyleJudge: A Human-Validated, In-Context-Learning Vision-Language Evaluator for Artistic Style Transfer

Lokesh, Pramukto Mandal, Somanshu Juneja, Jashnoor Singh, Amit Kumar, Pranav Adlinge, Richa Thakur, Kunal Swami 等 9 位作者

原文摘要

Automatic evaluation of artistic style transfer remains an open problem: existing metrics score an image pair in isolation, without seeing an example of the intended style transformation, and correlate poorly with human judgment. We introduce StyleJudge, a compact vision-language evaluator conditioned on a small set of reference source-to-stylized pairs before scoring a query pair, mirroring how a human rater calibrates to a style before judging a candidate. On a human-annotated benchmark of 878 rated pairs across 12 artistic styles (728 from 10 training/validation styles, 150 from 2 unseen styles), our efficient 0.8B-parameter StyleJudge reaches a Pearson correlation of 0.291 (Spearman 0.325) with human ratings on seen styles and 0.381 on unseen ones. On seen styles, it achieves the highest positive correlation among tested metrics: Gram-matrix style loss and CLIP similarity correlate weakly ($r=0.041$, $r=0.168$), while LPIPS (scored higher when closer to the source) correlates negatively ($r=-0.404$). We achieve this by distilling evaluation knowledge from a 72B-parameter vision-language teacher used only during training, which reaches $r=0.327$, confirming that most of its human alignment survives compact distillation. For training, we build a dataset of source-to-stylized pairs across the 12 styles via LoRA-adapted diffusion, densified with graded intermediate examples from a high-resolution image-morphing pipeline so the evaluator recognizes partial and complete stylization. On two held-out validation styles, StyleJudge reaches 28.2% exact-match agreement with its teacher's structured scores versus 1.1% for the untuned base model, with 67.8% shorter reasoning traces, and scores a pair in 39 s (median) using 2.5 GB of GPU memory on one NVIDIA T4. We release the human-rating benchmark, annotation guideline, SDXL-adapted FreeMorph code, and model weights.

cs.CV

新论文 · v1arXiv:2611.19274

HospitalBench: Diagnosing Where and Why Embodied Agents Fail on Long Hospital Task Chains

Gaytri Jena, Kapil Wanaskar, Aman Chadha, Vasu Sharma, Vinija Jain, Amitava Das

原文摘要

Long-horizon embodied tasks fail in ways that end-task success cannot see. HospitalBench is a diagnostic benchmark of ten hospital chores in NVIDIA Isaac Sim, each two unscored warm-up steps followed by a causal chain of five scored sub-tasks in one continuous simulation: the state one step leaves (an object pose, a choice, a fact seen once) determines what the next step must do, unlike benchmarks that place sub-tasks side by side. Vision-language agents act through skill-level tools from camera images. A scripted robot that makes only the calls a model gets is the solvability check of a chore. From each trajectory the benchmark reads where a chain breaks, what the agent did after a failed call, and how far its own position estimate drifts from the true one. This version reports first results only: 15 of 60 planned runs (three models at high effort, one run each on chores T1 and T10 from both start states and T3 from one), with scripted-robot proofs on three of the ten chores. Of the 15 runs, 2 passed all five sub-tasks. Runs still going after sub-tasks 1 to 5 were 15, 8, 4, 2, 2. Of 189 failed calls, 168 were followed by a different call and 111 by a call that did not fail. These are first results and not a full benchmark. We will release the scenes, task chains, scripted robot, evaluation harness and trajectories.

cs.RO · cs.AI

新论文 · v1arXiv:2611.18145

Compiling Specialized GPU-Initiated Networking

Javid Baydamirli, Ismayil Ismayilov, Kaan Oktay, Didem Unat

原文摘要

Machine learning systems specialize computation to the workload and hardware, but communication remains in generic libraries whose device code and transport decisions are not visible to high-level compilers. We present Obaleth, a compiler that makes GPU-initiated communication part of a kernel's intermediate representation, co-optimizes it alongside computation, and lowers it to specialized transport code. Obaleth plans completions around buffer lifetimes and optimizes memory ordering, synchronization, and batching across NVIDIA and AMD GPUs. On H100 GPUs, Obaleth sends small messages up to 3.3$\times$ faster, and on 256 GPUs runs all-to-all 1.25$\times$ and 2.3$\times$ faster than NVSHMEM and NCCL GIN, with much smaller memory and register use. In vLLM, it serves Qwen3-235B on 16 B200 GPUs and answers requests up to 1.29$\times$ faster than with DeepEP.

cs.DC · cs.LG · cs.PL

版本更新 · v2arXiv:2611.10357

LEGIT: Live Eligibility Gating over Identifier Tries for Constrained Decoding of Heterogeneous Pages

Felix Fang, Zheng Liu, Robert Saliba

原文摘要

A generative recommender names each candidate with a short token sequence, a Semantic ID, and recommends by generating those tokens. Some generative recommenders, including the Uber Eats search engine we are building, write a whole results page, stores with their menu items, in one decoding pass. Every Semantic ID on such a page must name a real store or item that is eligible for the search, such as a store that delivers to the user's location, and that is not already on the page. Each item must also belong to the store above it. A trie over the catalog keeps identifiers real, but eligibility and repetition change with every request and token, and the grammar-constrained decoder we tried could not enforce these changing rules exactly and quickly. We call this constraint LEGIT, for live eligibility gating over identifier tries. Numbering trie leaves depth first turns each rule into a range lookup over one store trie and one item trie, and every generated page passes an independent catalog check. On one NVIDIA B200 GPU with one or two requests per decode step, a direct implementation that walks the trie on the GPU makes each token about 5 times as slow as unconstrained decoding. Across 36 implementations, hardware counters and ablations are consistent with kernel launches and CPU-GPU waits as the source of this slowdown. Our fastest implementation walks the trie on the CPU while the GPU runs the model, needs no custom kernel, and cuts time per output token by 65.9 percent against the direct GPU implementation, to about 1.7 times the unconstrained cost. This implementation is also the fastest on a page layout that puts each item before its store.

cs.IR · cs.LG · cs.PF

新论文 · v1arXiv:2611.17929

Evaluating Large Language Models for Scientific Paper Search: Retrieval Success and Inference Latency on Mac Studio and NVIDIA A100

Frederic Andriantseheno Rakoto, Yannis Bendi-Ouis, Xavier Hinaut

原文摘要

Scientific literature search can benefit from locally deployed LLM assistants, which reduce reliance on hosted model providers while giving research teams greater control over their data and infrastructure. This work investigates whether open-weight models can support web-assisted paper retrieval under practical computing constraints. We first characterize inference latency as the input context grows from 512 to 128k tokens on a Mac Studio and a server equipped with two NVIDIA A100 GPUs. We then evaluate eight models on the retrieval of 70 papers from seven scientific repositories, using queries that represent three levels of domain knowledge. The results show that retrieval success depends strongly on the selected model and the specificity of the query. And that A100 GPUs are generally faster than the Mac Studio, but its unified memory allows larger models that could not be loaded in the tested GPU configuration.

cs.AI

版本更新 · v4arXiv:2603.08036

Samyama: A Unified Graph-Vector Database with In-Database Optimization, Agentic Enrichment, and Hardware Acceleration

Madhulatha Mandarapu, Sandeep Kunkunuru

原文摘要

Modern data architectures fragment across graph databases, vector stores, analytics engines and optimization solvers, forcing ETL between them. We present Samyama, a graph-vector database in Rust that unifies these workloads in one engine: a RocksDB-backed store with MVCC, a vectorized executor, a cost-based planner, a CSR analytics engine, RDF and SPARQL, 22 metaheuristic solvers callable from the query language, HNSW vector indexing, and agentic enrichment that expands a graph via LLMs. It has been run to billion-edge scale: 74.3M nodes and 1.07B edges from four biomedical sources on one machine for $2.50 of spot compute. This version re-measures the system at release v1.8.0 on rented Linux hosts a reader can boot, replacing earlier figures taken on a Mac Mini that nobody could re-run. On a 16-vCPU cloud instance, ingestion reaches 123K-152K nodes/s and PageRank costs 6.7 ms per iteration at ten thousand nodes and 236 ms at a million. The previously reported 8.2x GPU speedup is withdrawn; in its place we measure the CUDA path, which no earlier version measured at all, on a rented NVIDIA A16: parity with the CPU at ten thousand nodes, 1.62x at a hundred thousand and 2.70x at a million, with 1.1-1.35 s of context initialisation on the first call. The earlier Cypher-throughput figure, its claim of near-constant index-driven scaling, and a vector-search figure are withdrawn as unreproducible. openCypher conformance, once estimated at ~90%, is now measured at 99.9% of evaluated TCK scenarios. Against Neo4j 5 and FalkorDB on one host over an identical SF10 extract, 14 of 14 SNB complex reads land within 5x of the best competitor and 7 of 14 are faster.

cs.DB

新论文 · v1arXiv:2611.17867

UltraBERT-JEV: Encoder models for System One Applications

Astarag Mohapatra

原文摘要

We study encoder models that pick one of several options in a single forward pass, without generating text. This technical report compares recipes for fine-tuning UltraBERT-380M with one mask per option. The four recipes differ in how an option is scored and how the scores become probabilities; one follows the expert gate of DeepSeek-V3. Trained on the same data as Qwen3-0.6B with the LLM-as-Jev recipe (67.6% test accuracy), the encoder reaches 69.7% to 72.1%. Scoring an option by "yes" minus "no" instead of "yes" alone gives the larger gain, from 69.7% to 71.3% on the test set and from 57.8% to 59.6% on external benchmarks (56.1% for the decoder). Sigmoid normalisation adds 0.8 test points; a square-root softplus adds none. One request at a time on an NVIDIA A100, the encoder takes 29 ms per decision against 68 ms, but uses more peak memory (2.8 against 2.2 GiB). An auxiliary head, trained on human-marked evidence and on labels selected with a reward adapted from SelfCite, predicts in the same pass the sentences annotated as support for the answer.

cs.CL · cs.AI

版本更新 · v2arXiv:2608.29583

Drive the Thoughts: Runtime Monitoring of VLA Reasoning-Trajectory Consistency

Tian Yu, Lu Feng, Sebastian Elbaum

原文摘要

While machine learning models help autonomous vehicles (AVs) navigate complex environments, their black-box nature complicates validation and verification (V&V). Vision-language-action (VLA) models offer a unique opportunity: alongside trajectories, they generate an explicit chain-of-thought (CoT) explaining their decisions. We posit that this CoT can serve as a candidate specification to cross-check trajectories and expose unsafe behavior. To assess this, we curate DriveAlignBench, a dataset of 150 CoT-trajectory pairs from NVIDIA's Alpamayo 1.5 VLA, manually annotated for CoT reliability, CoT-trajectory consistency, and trajectory safety. We find that 33.3% of CoTs are unreliable, and among reliable CoTs, 26% have inconsistent trajectories. To detect these cases, we propose integrating a CoT-trajectory consistency check into a runtime monitor. This check is nontrivial because CoTs express open-vocabulary, scene-relative maneuvers, while trajectories are low-level ego-motion sequences whose semantics depend on road geometry and motion context. To bridge this gap, we develop a family of automated consistency monitors. The strongest, the lane-aware monitor, adds lane-relative evidence to the input of the judge model scoring consistency. This raises F1 across all four evaluated judges from 0.56-0.62 with raw waypoints to 0.65-0.82, showing that evidence representation improves detection across judges. Moreover, for reliable CoTs, inconsistency is associated with unsafe behavior: 27% of human-labeled inconsistent trajectories are unsafe compared to 3% of consistent ones, and the lane-aware monitor flags most of these unsafe cases. We release DriveAlignBench, the monitor implementations, and per-run results at https://github.com/less-lab-uva/drive-the-thoughts.

cs.SE

版本更新 · v2arXiv:2601.22585

HetCCL: Efficient LLM Training on Heterogeneous Vendor GPUs

Heehoon Kim, Chris Jaehwan Lee, Taejeoung Kim, Jongwon Park, Jinpyo Kim, Pyongwon Suh, Ryan H. Choi, Sangwoo Lee 等 9 位作者

原文摘要

The rapid growth of large language models is driving organizations to expand their GPU clusters, often with GPUs from multiple vendors. However, current deep learning frameworks lack support for collective communication across heterogeneous GPUs, leading to inefficiency and higher costs. We present HetCCL, a collective communication library that unifies vendor-specific backends and enables RDMA-based communication across GPUs without requiring driver modifications. HetCCL introduces two novel mechanisms that enable cross-vendor communication while leveraging optimized vendor libraries, NVIDIA NCCL and AMD RCCL. Evaluations on a multi-vendor GPU cluster show that HetCCL matches NCCL and RCCL performance in homogeneous setups while uniquely scaling in heterogeneous environments, enabling practical, high-performance training with both NVIDIA and AMD GPUs without changes to existing deep learning applications.

cs.DC · cs.LG

新论文 · v1arXiv:2611.17565

Performance-Portable Exact Gaussian Process Regression in Python via SYCL and oneMath

Alexander Strack, Marcel Graf, Alexander Van Craen, Dirk Pflüger

原文摘要

Modern multi-vendor GPU environments force a choice between portable code and native performance. GPRat is an open-source Python library for exact Gaussian process regression built on the HPX asynchronous many-task runtime, tiled algorithms, and CPU/CUDA BLAS backends. We contribute a UXL Foundation oneMath backend with a hybrid HPX/SYCL design, extending GPRat to AMD and Intel GPUs. The new SYCL backend comes within 3% of native CUDA on an NVIDIA A30 at the largest problem size and outperforms the GPflow and GPyTorch libraries by up to 1.75x and 2.17x on the A30 and up to 1.79x and 4.62x on an AMD Instinct MI210. On an Intel Arc B580, a consumer-class card with FP64 throughput below our CPUs, the unmodified SYCL backend trails GPRat's CPU backend. Overall, oneMath's dispatch to vendor-native BLAS lets a single SYCL codebase match CUDA on NVIDIA and outperform established Python GP frameworks on AMD, without vendor-specific code.

cs.DC · cs.LG · cs.PF

版本更新 · v2arXiv:2610.02665

大规模语言连续扩散模型

Large Language Continuous Diffusion Models

Zhihan Yang, Wei Guo, Jean-Marie Lemercier, Simon Welker, Mohammad Mahdi Kamani, Yonggan Fu, Sajad Norouzi, Julius Berner 等 17 位作者

AI 速读 基于摘要

提出首个大规模(3B/8B)连续扩散语言模型 Nemotron-Labs-Sigma,通过可操控的低维 ODE/SDE 潜轨迹实现高效推理与加速。

cs.CL · cs.AI · cs.LG

查看原文摘要

Despite the success of discrete diffusion language models (dLMs) for fast parallel decoding, their non-smooth, high-dimensional space hinders trajectory steering for reasoning and inference acceleration. To overcome this, we present Nemotron-Labs-Sigma, the first family of large-scale (3B/8B) continuous dLMs built on steerable, low-dimensional ODE/SDE latent trajectories. Trained blockwise via likelihood optimization, Nemotron-Labs-Sigma jointly denoises Gaussian-corrupted token embeddings while learning an optimal embedding geometry, and leverages pretrained autoregressive (AR) weights for warm-starting. At inference, we identify classifier-free guidance and score temperature as essential for high-fidelity reasoning and coding. Against state-of-the-art discrete counterparts (masked dLMs and AR baselines), Nemotron-Labs-Sigma outperforms them on standard benchmarks after pretraining (e.g., GSM8K, Minerva, HumanEval, MBPP) and on challenging reasoning tasks after supervised fine-tuning (e.g., MATH-500, AIME), with the largest consistent gains on coding. Beyond accuracy, we exploit two structural properties unique to continuous dLMs. First, smooth trajectories degrade gracefully at low NFEs and enable efficient few-step distillation: distribution matching proves strongest among the methods we study, giving roughly 8x acceleration on math and 4x on coding (8 and 4 tokens per forward pass) with negligible accuracy loss and preserved diversity. In addition, the one-step speculative decoding on the (pre-SFT) distilled 8B model yields a further ~3.5x throughput gain over the AR counterpart (on GB200), outperforming masked dLM baselines. Second, embedding-space steering controls the quality--diversity trade-off, yielding strong pass@$k$ that keeps improving as $k$ grows. Together, these results establish continuous dLMs as a promising paradigm for efficient language generation.

新论文 · v1arXiv:2611.17212

稀疏卷积的理论与实证分析

Theoretical and Empirical Analysis of Sparse Convolution

Keshvi Tuteja, Yu-Hsiang Tsai, Gregor Olenik, Markus Götz, Achim Streit, Hartwig Anzt, Charlotte Debus

AI 速读 基于摘要

该工作通过roofline分析刻画稀疏权重标准卷积的性能特征,并提出高稀疏度下的高效稀疏卷积技术,初步在NVIDIA A100与Jetson Nano上相对稠密卷积取得加速。

cs.PF · cs.DC · cs.LG

查看原文摘要

Deep neural networks for image processing are increasingly deployed on resource-constrained edge devices such as drones, medical systems, and smartphones, where limited computational resources and memory bandwidth pose significant challenges for efficient inference. Pruning during training often produces highly sparse convolutional kernels, offering the potential to significantly reduce computation and memory requirements. However, exploiting this sparsity efficiently at runtime remains challenging due to irregular memory access patterns and workload imbalance across processing elements. In this work, we investigate the standard convolution with sparse weights and analyze its performance characteristics through roofline analysis, showing that a 3 x 3 sparse convolution achieves an arithmetic intensity of approximately 10 FLOP/byte compared to roughly 96 FLOP/byte for its dense counterpart. We propose techniques for an efficient sparse convolution at high sparsities. Preliminary results demonstrate that our approach achieves upto 1.65x speedup as compared to dense convolution on NVIDIA A100 and upto 1.52x on NVIDIA Jetson Nano (Tegra X1) , motivating further exploration of direct convolution as a practical path to efficient sparse inference on modern accelerators/edge devices.

新论文 · v1arXiv:2611.17152

AeroZoo:面向空气动力学基础模型的良定流动问题基准

AeroZoo: A Benchmark of Well-Posed Flow Problems for Foundation Models of Aerodynamics

Peter Sharpe, Carmelo Gonzales, Corey Adams, Mehdi Ataei, Sanjay Choudhry

AI 速读 基于摘要

作者提出 AeroZoo 基准与训练语料,用 3,815 个 RANS 时均解覆盖多几何族、多外罩配置与宽雷诺数范围,并设计按几何族、外罩、边界条件布局和雷诺数范围留出的冻结划分,以检验空气动力学基础模型的跨分布泛化能力。

cs.LG · cs.CE · physics.flu-dyn

查看原文摘要

Neural surrogates for computational fluid dynamics (CFD) are trained and ranked on datasets that each vary one design family in one flow domain. Such datasets cannot test the least that an aerodynamic foundation model must do: predict CFD-quality incompressible flow in seconds for new shapes in new enclosures. On DrivAerML, larger models and longer training barely lower Transolver's error; doubling the training cases cuts it by about 12% up to 400 cases, exhausting the dataset. We introduce AeroZoo, a benchmark and training corpus of 3,815 time-mean solutions to the unsteady Reynolds-averaged Navier--Stokes (RANS) equations with a transition model, across seven geometry families, eight enclosure configurations and Reynolds numbers from $10^3$ to $10^8$. Each sample poses a well-posed PDE problem in the form a solver receives it: the boundary meshes grouped by condition type, the prescribed values and the fluid properties. The 2,816 samples with a stable time mean form the canonical benchmark: two tasks, engineering metrics and 22 frozen splits, disjoint in geometry, that hold out geometry families, enclosures, boundary-condition layouts and Reynolds ranges. On the random split, boundary-condition tokens cut GeoTransolver's median velocity error from 76.9% to 8.4% of the inlet speed in single-seed runs (uniform inlet flow: 11.0%). ISLA, a transformer for boundary-driven PDEs, reaches 5.3% over four seeds, about 4 times the change from a finer mesh. On the same test samples, a held-out geometry family barely changes ISLA's error relative to uniform flow, and a held-out enclosure or boundary-condition layout raises it 2.4 to 9.2 times. The data and the scoring code will be released.

新论文 · v1arXiv:2611.16856

可持续的LoRA联邦微调:环境影响与性能足迹

Sustainable Federated Fine-Tuning with LoRA: Environmental Impact and Performance Footprint

Pedro Augusto Londe Ribeiro, Rodrigo Moreira, Larissa Ferreira Rodrigues Moreira, Joberto S. B. Martins, Tereza C. Carvalho, Flávio de Oliveira Silva

AI 速读 基于摘要

作者报告在三种GPU上评估联邦LoRA微调的能耗与碳排放,发现本地步数主导能耗变化,较少步数可在保持精度的同时大幅降低能耗。

cs.LG · cs.NI

查看原文摘要

Federated Learning (FL) with Parameter-Efficient Fine-Tuning (PEFT) enables efficient adaptation of Large Language Models (LLMs), yet its energy and carbon costs across hardware remain poorly understood. We evaluated federated Low-Rank Adaptation (LoRA) fine-tuning on NVIDIA A40, RTX 6000 Ada, and A100 GPUs using a $2^2$ full factorial design with 135M/360M models and 10/100 local steps. CodeCarbon tracked energy and estimated CO$_2$-equivalent emissions. Descriptive allocation of variation assigned 84.8-91.2% of energy and emissions variation to local steps within these ranges. On A40 and RTX 6000 Ada, configurations with fewer steps achieved comparable accuracy with roughly 6-8-fold lower energy consumption and higher accuracy per kWh. These findings support smaller local budgets for energy-efficient federated fine-tuning within the evaluated setting.

版本更新 · v2arXiv:2611.09195

InstantHMR : Real-Time Human Mesh Recovery on Your Phone

Mohamed Adjel, Kahina Chalabi, Vincent Bonnet

原文摘要

Real-time 3D human mesh recovery from a single image on mobile devices is a challenging problem with many applications, including robotics, biomechanics, AR/VR, and exergames. Current state-of-the-art models are either too slow to run efficiently on mobile devices or sacrifice accuracy for speed. We present InstantHMR, a family of open-source single-image regressors suited for mobile inference, based on the Momentum Human Rig (MHR) parametric model and trained exclusively on publicly available annotations from the SAM 3D Body dataset. Our approach achieves remarkable performance on a Samsung Galaxy S23 NPU, recovering a 6,890-vertex body mesh in 5.5 ms per person, and 0.98 ms on an NVIDIA RTX 4070 GPU. This speed surpasses all existing methods we evaluated. On mobile, it is 12x faster than the single mobile-compatible method that slightly outperforms ours in accuracy, while achieving competitive accuracy with 40.0 mm PA-MPJPE on 3DPW, 48.8 mm on EMDB and 41.7 mm on RICH. To ensure fair comparisons, we re-evaluate 11 state-of-the-art methods under a unified protocol aligned with the closest MHR-based competitors, including Meta's official MHR-to-SMPL conversion, and report their reproduction results. Additionally, we measure latency across eight different targets, ranging from desktop TensorRT to two phone NPUs. Code and models are available at https://github.com/Naturalpad-studio/InstantHMR.

cs.CV

版本更新 · v2arXiv:2611.12304

Evolving Infra: Agent-Adaptive LLM Serving at Scale

Hao Kang, Weili Xu, Ziyang Li, Zheng Du, Hanchen Li, Tian Xia, Song Bian, Tushar Krishna 等 11 位作者

原文摘要

Agentic workloads are a major source of LLM inference demand. Existing prefill--decode (PD) disaggregated serving systems offer limited flexibility for runtime adaptation of compute allocation and KV cache management, and rely on coarse signals that only partly capture the diverse and dynamic agent workload. These limitations lead to load imbalance between prefill and decode nodes and eviction of KV caches that agents will reuse. We present Evolving Infra, an agent-adaptive serving framework that adapts compute allocation to agent workloads and KV cache management to agent context management at runtime. We decouple compute allocation from key--value (KV) cache management, allowing nodes to switch efficiently between prefill and decode. Building on this design, an online planner matches the PD split to workload composition, while an agent-aware distributed KV manager guides cache eviction and transfer with signals from agent context management. Across coding, math, and computer-use tasks with GLM-5.2 and Kimi K2.6 on up to 256 GB300 GPUs, Evolving Infra reduces mean end-to-end request latency and P95 TTFT by up to 4.1x and 14x over Dynamo, NVIDIA's production framework. We open-source the system implementation of Evolving Infra at https://github.com/mit-han-lab/EvolvingInfra.

cs.DC

新论文 · v1arXiv:2611.16363

Vector-Space Projection and Unified Complex-Valued Acceleration for Scaling Matrix-Based Brain-Inspired Representations

William Youngwoo Chung, Hyunwoo Oh, Calvin Yeung, Hansen Jin Lillemark, Hamza Errahmouni Barkam, Mohsen Imani

原文摘要

Symbolic reasoning at scale requires representations that are both expressive and hardware-efficient. Hyperdimensional Computing (HDC) provides a promising brain-inspired framework, encoding concepts as high-dimensional vectors composed through efficient algebraic operations. However, the widely used Fourier Holographic Reduced Representation (FHRR) relies on commutative binding, which cannot encode order or directionality. Prior FPGA/ASIC accelerators recover this through costly permutation logic, incurring irregular memory access and high control overhead. Generalized Holographic Reduced Representations (GHRR) instead redefine binding as non-commutative matrix multiplication, naturally encoding sequences, graphs, and hierarchies without permutations-but at the cost of turning every element into a matrix, introducing compute and storage overhead that limits scalability. In this work, we propose a flattening methodology that preserves GHRR's matrix-based encoding while executing training and inference directly in vector space, functionally equivalent to FHRR inference yet free of permutation logic. We further design a custom 28 nm ASIC that fuses binding and similarity into a unified complex-valued datapath, with dual-DMA streaming and runtime normalization for accurate inference. Against a PyTorch baseline on an NVIDIA RTX 4090 GPU, the prototype delivers 1.36x-1.56x higher throughput and 16.2x-18.6x better energy efficiency.

cs.AR