SumCheck is a central prover primitive in zero-knowledge proofs, exposing table-parallel work that maps naturally to GPUs but becomes increasingly arithmetic-intensive as finite fields widen. We study how this shift affects the cost of programming SumCheck in Triton rather than CUDA. Our benchmark, zkDuel, provides matched implementations of 25 zkPHIRE-derived constraints alongside diagnostic polynomials, covering 32-, 64-, 128-, and 256-bit prime fields, with independent CPU-reference validation and cross-backend checksum gating. Of 800 configurations attempted on an NVIDIA H100, 758 complete with matching checksums. CUDA's median advantage over Triton falls from 3.5x at 32 bits to 1.4x at 256 bits, and Triton is faster on 19 large-table configurations; the same convergence appears on an A100, where the gap falls from 4.4x to 1.6x. Profiling shows that the small-field gap largely reflects host launch overhead and redundant modular additions in Triton's original block reduction, whose relative costs diminish as multi-limb Montgomery work grows. Replacing that reduction with tl.reduce brings Triton's median kernel time within 8% of CUDA on both GPUs, while specialized point interpolation improves 256-bit evaluation kernels by 1.55-1.75x. At cryptographic widths, these results show that a Python-embedded DSL can approach CUDA performance, although the largest fully unrolled constraints still expose compiler-scalability limits.
Direct-policy World Action Models (WAM) have demonstrated the benefits of decoupling action generation from explicit future visual prediction for efficient inference in autonomous driving and embodied control. However, they still require tens of denoising steps to achieve competitive action quality, incurring substantial latency and hindering real-time deployment. Simply reducing the number of denoising steps will collapse the nonlinear multi-step denoising trajectory into coarse updates, substantially degrading action quality. To reduce the denoising steps while enabling real-time action generation, we present an effective world action model with one-step action generation, namely OSWAM, for autonomous driving and embodied control. We argue that it is the absence of structured action priors in random noise initialization that forces direct-policy WAMs to recover coherent action sequences through iterative denoising. Hence, we construct imagination tokens to initialize the diffusion process, providing physical priors and action-conditioned representations. We further introduce privileged information distillation with future conditioning to enhance future-aware action generation. As a result, OSWAM can efficiently and effectively generate high quality actions in just one diffusion step. Our experiments demonstrate that OSWAM achieves comparable or even better results in both autonomous driving and embodied control than direct-policy WAM and Joint-WAM that require tens of steps. It obtains approximately $3\times$ speedup over direct-policy WAM baselines and supports real-time compiled inference with latency as low as 40ms on an NVIDIA RTX 5090 GPU. The project page is available at https://gwenzhang.github.io/OSWAM/
High-resolution video matting requires preserving fine details, such as hair strands and soft transparency, while maintaining consistency across frames. Memory-based methods achieve strong matting quality by matching current-frame features with information stored from earlier frames, but this process becomes increasingly costly as resolution grows. We present MatAnyone-E, an efficient extension of MatAnyone 2 that focuses computation where it is most needed. Our approach first performs low-resolution propagation to identify uncertain regions that require closer inspection and the memory candidates most relevant to them. It then performs high-resolution matching only for these regions using the selected memory, while reusing low-resolution matching results elsewhere. This strategy reduces both the number of locations requiring detailed matching and the amount of memory searched, while preserving the information needed to produce a complete matte. We further reduce the cost of generating high-resolution mattes with a lightweight decoder designed around measured runtime bottlenecks and trained to reproduce the original decoder's representations. Experiments show that MatAnyone-E maintains matting quality comparable to MatAnyone 2 while achieving an estimated 7.4x inference speedup and reducing peak GPU memory usage to 7.6% of the usage of MatAnyone 2 at 4K. With compilation-enabled inference on a single NVIDIA H200 GPU, it reaches 73.07 FPS at 1080p and 28.62 FPS at 4K.
RISC-V is emerging as a candidate architecture for high-performance computing (HPC), but its competitiveness against established x86 and Arm processors remains difficult to assess because existing studies use different platforms, workloads, and methodologies, and rarely consider energy efficiency. Our work provides an evaluation of contemporary HPC RISC-V, Arm, and x86 processors using a common software stack and methodology. We characterize the Sophgo SG2044 and compare it with the NVIDIA Grace CPU Superchip and Intel Xeon Platinum 8480+ across four HPC benchmarks (STREAM, HPL, HPL-MxP, and HPCG) and four representative applications (iPIC3D, PLUTO, OpenFOAM, and llama$.$cpp). The version of HPL-MxP linked to OpenBLAS built and used for this work is made available as open-source software. The SG2044 reaches 8.02 GFLOP/s of FP64 performance and 7.8 GB/s of memory bandwidth per core, increasing to 294 GFLOP/s and 107 GB/s at 64 cores. L2-cache-aware process placement enables near-linear scaling up to 16 cores, beyond which memory-bandwidth saturation limits further gains. Across generations of RISC-V hardware, in the context of the Monte Cimone (MC) project, the SG2044-based MCv3 doubles both single-core performance and peak memory bandwidth relative to MCv2, while improving memory bandwidth, node-level FP64 performance, and energy efficiency by approximately 100x, 158x, and 10x, over MCv1. Against the Arm and x86 systems, the SG2044 shows 8-15x less absolute performance, reflecting its narrower vector resources and less advanced technology node, but delivers competitive scaling and energy efficiency on several workloads. Its most notable result is LLM decode, where 64-core execution provides the highest relative scaling and energy efficiency among the evaluated platforms while exceeding Intel in throughput. Our results show that RISC-V processors are closing the gap with the competition.
WebGPU's security-focused design imposes per-operation validation that compounds across the many small dispatches in neural network inference, yet the true cost of this overhead is poorly characterized. We present a systematic characterization of WebGPU dispatch overhead for LLM inference at batch size 1, spanning four GPU vendors (NVIDIA, AMD, Apple, Intel), two native implementations (Dawn, wgpu-native) and three browsers (Chrome, Safari, Firefox), and two model sizes (Qwen2.5-0.5B and 1.5B). Our primary contribution is a sequential-dispatch methodology that reveals naive single-operation benchmarks overestimate dispatch cost by ${\sim}20\times$. The true per-dispatch cost of WebGPU API overhead alone is 24-36 $μ$s on Vulkan and 32-71 $μ$s on Metal, while the total per-operation overhead including Python cost is ${\sim}95$~$μ$s, which turns out to be a distinction critical for optimization. On Vulkan, kernel fusion improves throughput by 53%, while CUDA fusion provides no benefit, confirming that per-operation overhead is a primary differentiator. LLM inference was tested across three major operating systems (Linux, Windows, macOS). We built $\texttt{torch-webgpu}$, a PrivateUse1-based out-of-tree PyTorch backend and an FX-to-WebGPU compiler, which on our reference platform achieves 11--12% of CUDA performance. At dtype-matched float32, RTX PRO 2000 achieves 1.4$\times$ WebGPU's throughput despite ${\sim}6\times$ less compute than RTX 5090. For dispatch overhead, backend choice is the dominant factor, although implementation choice also matters substantially within a backend (2.2$\times$ for Metal). In terms of dispatch vs kernel compute efficiency, we conclude that at batch=1 with the current dispatch-heavy pipeline, per-operation overhead dominates regardless of kernel quality. All code, benchmarks, and raw data are open source.
Rotationally equivariant architectures improve the physical fidelity of machine-learned (ML) universal materials models based on graph neural networks (GNNs). However, their scalability is limited by the use of non-trivial, high-precision tensor operations during both training and inference. Adopting the reduced-precision formats of matrix multiplication units (MMUs), e.g., FP8 instead of FP32, provides massive increases in throughput, up to 30x (60x) on NVIDIA Hopper (Blackwell) GPUs, but may lead to drastic loss of accuracy. Here, we introduce tailored 'emulated' low-precision operations under the integer Ozaki scheme to accelerate the 'SO(2)-equivariant' processing layers of GNNs, which dominate the runtime of (1) many state-of-the-art ML interatomic potentials (MLIPs) and (2) most universal ML Hamiltonian (MLH) models. Our framework, FlashSO(2), outperforms TF32, the most accurate native low-precision MMU format, in throughput, accuracy, and memory footprint at once. At the SO(2) layer level, it reaches 40x (7x) the throughput of MALOQ's (UMA's) reference implementation, an MLH (MLIP) model, sustaining up to 236 (232) effective Tflop/s on an NVIDIA GH200, 3.5x the FP32 peak. At the model level, FlashSO(2) helps train MALOQ 3.9x faster than its pure-FP32 version while reducing peak memory by 58% at no loss in accuracy. In molecular dynamics simulations with UMA, the speedup factor is 3.3x, and the peak memory is reduced by 51%. FlashSO(2) thus pushes the speed-accuracy Pareto frontier of equivariant materials models beyond the native low-precision formats and substantially lowers the resources required to exploit their scaling laws.
Sizhe Chen, Sanjay Kariyappa, Liwei Jiang, Chong Xiang, Jihye Choi, Chaowei Xiao, G. Edward Suh
The recent incident in which OpenAI agents autonomously compromised Hugging Face's infrastructure illustrates the importance of instruction hierarchy (IH): an LLM should follow lower-priority instructions, such as user requests, only when they do not conflict with higher-priority instructions, such as system policies. IH can be effectively improved by fine-tuning LLMs on simulated examples of instruction conflicts, whose diversity is key to teaching LLMs to properly handle conflicts at test time. However, our decomposition of a task into goals and constraints reveals a conflict type -- goal-goal conflicts -- that is absent from existing IH training sets from OpenAI and NVIDIA. We show the importance of this conflict type: fine-tuning Qwen3.5-9B and Gemma4-12B-it on examples of this conflict type significantly improves IH on a wide range of real-world tasks unseen in training; adding these examples to existing IH training sets for fine-tuning also gives better IH adherence. The IH gains come with trivial mean utility changes on no-conflict tasks across eight benchmarks. Our work advocates including this new conflict type in IH training, even if these examples do not reflect typical LLM use.
Christin David Bose, Cesar Avalos, Junrui Pan, Timothy Rogers
The Tensor Memory Accelerator (TMA), introduced on NVIDIA Hopper, is used pervasively in production GPU kernels yet its usage is conspicuously selective: expert kernel developers in CUTLASS, FlashAttention-3, and FlashInfer apply TMA to some tensor operands within a kernel but fall back to cp.async, direct global loads, or distributed shared memory (DSM) for others. Vendor documentation describes the API, and recent microbenchmarking work characterizes TMA throughput in isolation, but neither answers the question practitioners actually face: given a memory operand in a kernel, should it use TMA, and if so, with or without cluster multicast? We address this gap with a workload-grounded characterization study on H100. We catalog the TMA, multicast, DSM, and cp.async choices made inside FlashAttention-3 forward, isolate the discriminating features through targeted microbenchmarks (transfer-size crossover, multicast scaling), and produce rules of thumb for the load-mechanism choice on Hopper. We further compare TMA multicast against DSM as competing cluster-broadcast primitives, and isolate the FlashAttention-3 paged-key-value (KV) TMA-vs-cp.async dispatch on H100 across page size, datatype, model shape, and execution phase. Our results yield concrete, predictive guidance on when TMA pays off and when multicast helps or hurts on Hopper, and we extend the analysis to server Blackwell (SM100) via a CUTLASS GEMM study on B200 that shows the qualitative dispatch rules carry across generations while absolute crossovers shift with tensor-core and bandwidth scaling.
Large language model agents that use tools could take on part of the electronic health record and desk work that occupies physicians' time, but in many hospitals the medical information systems run inside closed networks, so the model and the agent harness must also be inside the network. We designed, implemented, and deployed LLENS, a clinical support AI agent platform that runs inside the closed network of Hokkaido University Hospital on a single host with 8 NVIDIA H200 GPUs, the open-weight model Kimi K2.6 (~1 trillion parameters), and a web application rebuilt from Open WebUI. The harness has 10 components, including page image reading, full-text attachment input, a path-addressed document tree without a vector index, a network-isolated sandbox, artifact return, self-checking of generated output, subagents, and saved procedures. Kimi K2.6 was chosen from 3 open-weight candidates and scored 480/500 on the 2026 Japanese Medical Licensing Examination; it generated 119 tokens/s for a single user and served 16 concurrent requests at 908 tokens/s in total. On 180 questions over hospital documents (3 seeds each), mean accuracy was 10.7% with no documents, 81.3% with fixed retrieval (top 10 chunks of one lexical search), and 97.2% with interactive access, in which the model reads with document tools and the sandbox (15.9 percentage points above fixed retrieval, 95% CI 10.9-21.3); the difference lay in filtering/counting questions (16.7% vs 87.8%). In an automated test of 13 work scenarios (4 runs each), the expected components were triggered in all runs for 12, and 5 scenarios were run with synthetic data on the deployment host. The platform has run on the hospital network since September 14, 2026.
LLMs on edge devices are constrained by limited memory and bandwidth. Although 1.58-bit ternary weights greatly reduce model storage, long-context inference is increasingly bottlenecked by KV-cache traffic and attention. Existing accelerators also use separate datapaths for ternary linear layers and attention, incurring hardware and execution overhead. We present SPARTA, a sparse ternary LLM accelerator for efficient long-context edge inference. SPARTA integrates: (1) a phase-aware sparse dataflow exploiting heterogeneous sparsity across compute-bound prefill and memory-bound decode; (2) a unified compute fabric mapping ternary SpMM/SpMV and bit-serial attention onto shared processing elements to exploit zero-weight and zero-bit sparsity; and (3) an adaptive two-stage attention engine combining page-level Top-k selection to reduce K-cache accesses with token-level Top-p pruning to reduce V-cache traffic. SPARTA reduces KV traffic by 4.1-4.8x over full attention and 1.3-1.5x over a state-of-the-art sparse-attention scheme, with task-dependent quality loss. Compared with NVIDIA Jetson AGX Orin, SPARTA achieves 3.9x higher generation throughput and 8.3x higher end-to-end energy efficiency on average; versus AMD Ryzen 7 9700X, the gains are 5.0x and 81.7x. It also delivers 1.6-12.0x higher throughput, 2.0-4.2x higher energy efficiency, and 1.3-51.4x higher area efficiency than FIGLUT, TeLLMe, and MCBP.
Stephen Abkin, Aldric Yang, Alex Flesher, Nitheesh Kodarapu, Deepak Panigrahy, Aakash Tyagi, David Kebo Houngninou
Quantization reduces the memory and energy needed to serve large language models (LLMs) and AI agents for coding and data analysis can now run on locally hosted LLMs. In practice, this benefit is commonly benchmarked per output token. Yet an agent calls its model repeatedly, runs tools, and revises its work, and quantization can change both how much the agent generates and whether it succeeds. Existing energy studies either compare quantized models on single requests or measure agents at one fixed precision, overlooking how a change in precision plays out across a complete agent task. To address this gap, this paper presents the first characterization of the task-level energy trade-offs of model quantization in locally hosted agents. We develop 29 agent tasks with automatic pass/fail checks in six families, ranging from single-answer SQL analytics to multi-step repository bug fixing, and measure CPU and GPU energy per output token and per completed task for 4-bit and 8-bit versions of two open-weight models on bare-metal NVIDIA A100 nodes with llama$.$cpp. Our findings reveal that 4-bit precision generally saves energy per token, though whether it saves energy per completed task is task dependent and changes with the model and the execution budget (i.e., limits on turns, tokens, and time). We trace this discrepancy to the extra tokens needed to complete a task in full, which translates into using up to 3.5 times as much energy per completed task. These findings offer actionable guidance for developers and operators choosing model precision for agents, showing how tokens per completed task, set against the per-token saving, reveal whether lower precision completes the work with less energy.
Can AI agents autonomously train useful language models and improve them using feedback from users? We introduce RSI Arena, a two-stage study of autonomous post-training as a step toward recursive self-improvement. Ten frontier models act as training agents, each starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base. During a six-day first stage, each agent has 1,000 GPU-hours and $300 in API credit on a shared 64-GPU cluster, independently selecting data, implementing training procedures, and choosing checkpoints. We evaluate their models on 20 benchmarks and through blind human evaluations. All ten agents produce working assistants, with the strongest model, trained by GPT-6 Astra, closing a median of 76% of the base-to-Instruct gap across 12 capability benchmarks. Its human-preference rating is also close to NVIDIA Instruct after statistical adjustment for answer length and formatting. In Stage 2, each agent receives the comparison records and votes involving its Stage 1 model and is given three additional days for further training. Using NVIDIA Instruct as a common reference across the two evaluation rounds, we observe an improvement from Stage 1 to Stage 2 in the style-adjusted relative rating of eight of the ten models. We show that benchmark gains were modest, with no clear link to preference gains, while some agents encouraged longer responses and one model showed increased harmful compliance on StrongREJECT. RSI Arena provides an auditable setting for studying how human feedback can guide autonomous model development. It takes a step toward recursive self-improvement with human feedback (RSIHF), in which people help define what improvement means for the AI systems they use.
Guangwei Zhang, Jianing Zhu University of Texas at Austin, Cheng Qian University of Illinois at Urbana-Champaign, Neil Zhenqiang Gong Duke University, Rada Mihalcea University of Michigan, Zhaozhuo Xu Workato and Stevens Institute of Technology, Jingrui He University of Illinois at Urbana-Champaign, Jiaqi W. Ma University of Illinois Urbana-Champaign 等 14 位作者
We present Copyright Detective, the first interactive forensic system for detecting, analyzing, and visualizing potential copyright risks in LLM outputs. The system treats copyright infringement versus compliance as an evidence discovery process rather than a static classification task due to the complex nature of copyright law. It integrates multiple detection paradigms, including content recall testing, paraphrase-level similarity analysis, persuasive jailbreak probing, and unlearning verification, within a unified and extensible framework. Through interactive prompting, response collection, and iterative workflows, our system enables systematic auditing of verbatim memorization and paraphrase-level leakage, supporting responsible deployment and transparent evaluation of LLM copyright risks even with black-box access. In our experiments with GPT-4o-mini, we demonstrate that the specific persuasive strategy “Pathos” shifts the leakage distribution from about 0.1 (ROUGE-L) to 0.7. Our live system is hosted on Streamlit server, with a demonstration video included as supplementary material.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表
Generative AI—powered by Large Language Models (LLMs)—is increasingly deployed in industry across healthcare decision support, financial analytics, enterprise retrieval, and conversational automation, where reliability, efficiency, and cost control are critical. In such settings, models must satisfy strict constraints on energy, latency, and hardware utilization—not accuracy alone. Yet prevailing evaluation pipelines remain accuracy-centric, creating a Deployment–Evaluation Gap—the absence of operational and economic criteria in model assessment. To address this gap, we present EDGE-EVAL—a industry-oriented benchmarking framework that evaluates LLMs across their full lifecycle on legacy NVIDIA Tesla T4 GPUs. Benchmarking LLaMA and Qwen variants across three industrial tasks, we introduce five deployment metrics—Economic Break-Even (Nbreak), Intelligence-Per-Watt (IP W ), System Density (ρsys), Cold-Start Tax (Ctax), and Quantization Fidelity (Qret)—capturing profitability, energy efficiency, hardware scaling, serverless feasibility, and compression safety. Our results reveal a clear efficiency frontier—models in the < 2B parameter class dominate larger baselines across economic and ecological dimensions. LLaMA-3.2-1B (INT4) achieves ROI break-even in 14 requests (median), delivers 3× higher energy-normalized intelligence than 7B models, and exceeds 6,900 tokens/s/GB under 4-bit quantization. We further uncover an efficiency anomaly—while QLoRA reduces memory footprint, it increases adaptation energy by up to 7× for small models—challenging prevailing assumptions about quantization-aware training in edge deployment.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表
Layer-aligned distillation and convergence-based early exit represent two predominant computational efficiency paradigms for transformer inference; yet we establish that they exhibit fundamental incompatibility under standard deployment conditions for convergence-based early exit. Distillation objectives that align intermediate student layers to teacher representations suppress the representational convergence that early-exit mechanisms exploit, rendering such mechanisms ineffective on distilled models.We introduce LEAP (Layer-wise Exit-Aware Pretraining), an auxiliary training objective that reconciles this incompatibility. LEAP requires no architectural modifications; it augments standard distillation with a single constraint ensuring intermediate layers approximate final-layer representations. LEAP-MiniLM achieves 1.61\times measured wall-clock speedup (batch = 1, NVIDIA L4) at \theta = 0.95, with 91.9% of samples exiting by layer 7 and 1.80\times theoretical layer reduction, where standard distilled models achieve zero effective speedup. We validate across sentence similarity (STS-B: 0.760 \pm 0.006) and retrieval benchmarks (BEIR), providing operational guidance including latency measurements, decision thresholds, and deployment criteria.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表
Navya Gupta Singapore Institute of Technology and Indraprastha Institute of Information Technology, Delhi, Rishitej Reddy Vyalla, Avinash Anand, Chhavi Kirtani, Erik Cambria Nanyang Technological University, Zhengchen Zhang, Zhengkui Wang, Timothy Liu NVIDIA 等 11 位作者
Curriculum learning helps language models tackle complex reasoning by gradually increasing task difficulty. However, it often fails to generate consistent step-by-step reasoning, especially in multilingual and low-resource settings where cross-lingual transfer from English to Indian languages remains limited. We propose IRIS: Interleaved Reinforcement with Incremental Staged Curriculum, a two-axis framework that combines Supervised Fine-Tuning on progressively harder problems (vertical axis) with Reverse Curriculum Reinforcement Learning to reduce reliance on step-by-step guidance (horizontal axis). We design a composite reward combining correctness, step-wise alignment, continuity, and numeric incentives, optimized via Group Relative Policy Optimization (GRPO). We release CL-Math, a dataset of 29k problems with step-level annotations in English, Hindi, and Marathi. Across standard benchmarks and curated multilingual test sets, IRIS consistently improves performance, with strong results on math reasoning tasks and substantial gains in low-resource and bilingual settings, alongside modest improvements in high-resource languages. Our code and dataset will be publicly available at https://github.com/avinanand/IRIS-Interleaved-Reinforcement-
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Binxing Xu, Hao Gu The Hong Kong University of Science and Technology, Lujun Li, Hao Wang, Bei Liu The Hong Kong University of Science and Technology, Jiacheng Liu, Qiyuan Zhu, Xintong Yang 等 11 位作者
Training LLMs at ultra-low precision remains a formidable challenge. Direct low-bit QAT often suffers from convergence instability and substantial training costs, exacerbated by quantization noise from heavy-tailed outlier channels and error accumulation across layers. To address these issues, we present \text{Bit-by-Bit}, a progressive QAT framework with outlier channel splitting. Our approach integrates three key components: (1) block-wise progressive training that reduces precision stage by stage, ensuring stable initialization for low-bit optimization; (2) nested structure of integer quantization grids to enable a “train once, deploy any precision” paradigm, allowing a single model to support multiple bit-widths without retraining; (3) rounding-aware outlier channel splitting, which mitigates quantization error while acting as an identity transform that preserves the quantized outputs. Furthermore, we follow microscaling groups with E4M3 scales, capturing dynamic activation ranges in alignment with OCP/NVIDIA standards. To address the lack of efficient 2-bit kernels, we developed custom operators for both W2A2 and W2A16 configurations, achieving up to 11\times speedup over BF16. Under W2A2 settings, \text{Bit-by-Bit} significantly outperforms baselines like BitDistiller and EfficientQAT on both Llama2/3, achieving a loss of only 2.25 WikiText2 PPL compared to full-precision models.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yixin Xiang Nanjing University of Science and Technology, Yunshan Ma Singapore Management University, Xiaoyu Du Nanjing University of Science and Technology, Yibing Chen Nanjing Pami Intelligent Technology Co., Ltd., Yanxin Zhang NVIDIA, Jinhui Tang Nanjing Forestry University
Document Question Answering (DQA) involves generating answers from a document based on a user’s query, representing a key task in document understanding. This task requires interpreting visual layouts, which has prompted recent studies to adopt multimodal Retrieval-Augmented Generation (RAG) that processes page images for answer generation. However, in multimodal RAG, visual DQA struggles to utilize a large number of images effectively, as the retrieval stage often retains only a few candidate pages (e.g., Top-4), causing informative but less visually salient content to be overlooked in favor of common yet low-information pages. To address this issue, we propose a Multi-Armed Bandit–based DQA framework (MAB-DQA) to explicitly model the varying importance of multiple implicit aspects in a query. Specifically, MAB-DQA decomposes a query into aspect-aware subqueries and retrieves an aspect-specific candidate set for each. It treats each subquery as an arm and uses preliminary reasoning results from a small number of representative pages as reward signals to estimate aspect utility. Guided by an exploration–exploitation policy, MAB-DQA dynamically reallocates retrieval budgets toward high-value aspects. With the most informative pages and their correlations, MAB-DQA generates the expected results. On four benchmarks, MAB-DQA shows an average improvement of 5%-18% over the state-of-the-art method, consistently enhancing document understanding. Codes are available at https://github.com/ElephantOH/MAB-DQA.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Mehrzad Samadi NVIDIA, Aleksander Ficek NVIDIA, Sean Narenthiran NVIDIA, Siddhartha Jain NVIDIA, Wasi Uddin Ahmad NVIDIA, Somshubra Majumdar NVIDIA, Vahid Noroozi NVIDIA, Boris Ginsburg NVIDIA
Competitive programming has become a rigorous benchmark for evaluating the reasoning and problem-solving capabilities of large language models (LLMs). The International Olympiad in Informatics (IOI) stands out as one of the most prestigious annual competitions in competitive programming and has become a key benchmark for comparing human and AI-level programming ability. While several proprietary models have been claimed to achieve gold medal-level performance at the IOI, often with undisclosed methods, achieving comparable results with open-weight models remains a significant challenge. In this paper, we present GenCluster, a scalable and reproducible test-time compute framework that attains IOI gold-level performance using open-weight models. It combines large-scale generation, behavioral clustering, ranking, and a round-robin submission strategy to efficiently explore diverse solution spaces under limited validation budgets. Our experiments show that the performance of our proposed approach scales consistently with available compute, narrowing the gap between open and closed systems. Notably, we will show that GenCluster can achieve a gold medal at IOI 2025 for the first time with an open-weight model gpt-oss-120b, setting a new benchmark for transparent and reproducible evaluation of reasoning in LLMs.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Ahmed Heakl LG Corporation, Gustavo Bertolo Stahl, Sarim Hashmi Mohamed bin Zayed University of Artificial Intelligence, Seung Hun Eddie Han Inception AI and Mohamed bin Zayed University of Artificial Intelligence, Mukul Ranjan MBZUAI, Arina Kharlamova, Salman Khan Mohamed bin Zayed University of Artificial Intelligence and Australian National University, Abdulrahman Mahmoud Mohamed bin Zayed University of Artificial Intelligence
Cross-architecture GPU code transpilation is essential for unlocking low-level hardware portability, yet no scalable solution exists. We introduce CASS, the first dataset and model suite for source- and assembly-level GPU translation (CUDA ↔ HIP, SASS ↔ RDNA3). CASS contains 60k verified host-device code pairs, enabling learning-based translation across both ISA and runtime boundaries. We generate each sample using our automated pipeline that scrapes, translates, compiles, and aligns GPU programs across vendor stacks. Leveraging CASS, we train a suite of domain-specific translation models that achieve 88.2% accuracy on CUDA → HIP and 69.1% on SASS → RDNA3, outperforming commercial baselines including GPT-5.1, Claude-4.5, and Hipify by wide margins. Generated code matches native performance in 85% of cases, preserving both runtime and memory behavior. To support rigorous evaluation, we introduce CASS-Bench, a curated benchmark spanning 18 GPU domains with ground-truth execution. All data, models, and evaluation tools will be released as open source to support progress in GPU compiler tooling, binary compatibility, and LLM-guided code translation.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jiandong Shao, Raphael Tang Microsoft and University College London, University of London, Crystina Zhang, Karin Sevegnani NVIDIA, Pontus Stenetorp University College London, Jianfei Yang, Yao Lu University College London
Multilingual large language models achieve impressive cross-lingual performance despite largely monolingual pretraining. While bilingual data in pretraining corpora is widely believed to enable these abilities, details of its contributions remain unclear. We investigate this question by pretraining models from scratch under controlled conditions, comparing the standard web corpus with a monolingual-only version that removes all multilingual documents. Despite constituting only 2% of the corpus, removing bilingual data causes translation performance to drop 56% in BLEU, while behaviour on cross-lingual QA and general reasoning tasks remains stable, with training curves largely overlapping the baseline. To understand this asymmetry, we categorize bilingual data into parallel (14%), code-switching (72%), and miscellaneous documents (14%) based on the semantic relevance of content in different languages. We then conduct granular ablations by reintroducing parallel or code-switching data into the monolingual-only corpus. Our experiments reveal that parallel data almost fully restores translation performance (91% of the unfiltered baseline), whereas code-switching contributes minimally. Other cross-lingual tasks remain largely unaffected by either type. These findings reveal that translation critically depends on systematic token-level alignments from parallel data, whereas cross-lingual understanding and reasoning appear to be achievable even without bilingual data.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yifeng Ding Meta FAIR and Department of Computer Science, University of Illinois at Urbana-Champaign, Hung Le, Songyang Han NVIDIA, Kangrui Ruan AWS AI Lab, Zhenghui Jin NVIDIA, Varun Kumar Amazon, Zijian Wang Meta Superintelligence Labs, Anoop Deoras
Training Large Language Models (LLMs) for multi-turn Tool-Integrated Reasoning (TIR) – where models iteratively reason, generate code, and verify through execution – remains challenging for existing reinforcement learning (RL) approaches. Current RL methods, exemplified by Group Relative Policy Optimization (GRPO), suffer from coarse-grained, trajectory-level rewards that provide insufficient learning signals for complex multi-turn interactions, leading to training stagnation. To address this issue, we propose Group Turn Policy Optimization (GTPO), a novel RL algorithm specifically designed for training LLMs on multi-turn TIR tasks. GTPO introduces three key innovations: (1) turn-level reward assignment that provides fine-grained feedback for individual turns, (2) return-based advantage estimation where normalized discounted returns are calculated as advantages, and (3) self-supervised reward shaping that exploits self-supervision signals from generated code to densify sparse binary outcome-based rewards. Our comprehensive evaluation demonstrates that GTPO outperforms GRPO by 3.0% across diverse math reasoning benchmarks, establishing its effectiveness. GTPO also improves GRPO by 3.9% on commonsense reasoning and program synthesis tasks, demonstrating its generalizability to non-math domains. Importantly, GTPO incurs negligible overhead, ensuring its practicality for real-world scenarios.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Radin Shayanfar, Chu Fei Luo, Rohan V Bhambhoria Evolver Transformation, Samuel Dahan Queen's University, Xiaodan Zhu Queen's University
Building Task-Oriented Dialogue (TOD) systems that generalize across different tasks remains a challenging problem. Data-driven approaches often struggle to transfer effectively to unseen tasks. While recent schema-based TOD frameworks improve generalization by decoupling task logic from language understanding, their reliance on neural or generative models often obscures how task schemas influence behaviour and hence impair interpretability. In this work, we introduce a novel framework, CoDial (Code for Dialogue), at the core of which is converting a predefined task schema to a structured heterogeneous graph and then to popular programmatic LLM guardrailing code, such as NVIDIA’s Colang. The pipeline enables efficient and interpretable alignment of dialogue policies during inference. We introduce two paradigms for LLM guardrailing code generation, CoDial-free and CoDial-structured, and propose a mechanism that integrates human feedback to iteratively improve the generated code. Empirically, CoDial achieves state-of-the-art (SOTA) performance on the widely used benchmark datasets, while providing inherent interpretability in the design. We additionally demonstrate CoDial’s iterative improvement via manual and LLM-aided feedback, making it a practical tool for human-guided alignment of LLMs in unseen domains.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhen Wan, Chao-Han Huck Yang NVIDIA Research, Jinchuan Tian, Hanrong Ye NVIDIA, Ankita Pasad NVIDIA, Szu-Wei Fu NVIDIA and Microsoft, Arushi Goel NVIDIA Research, Ryo Hachiuma NVIDIA 等 18 位作者
We introduce a voice-agentic framework that learns one critical omni-understanding skill: knowing when to trust itself versus when to consult external audio perception. Our work is motivated by a crucial yet counterintuitive finding: naively fine-tuning an omni-model on both speech recognition and external sound understanding tasks often degrades performance, as the model can be easily misled by noisy hypotheses. To address this, our framework, Speech-Hands, recasts the problem as an explicit self-reflection decision. This learnable reflection primitive proves effective in preventing the model from being derailed by flawed external candidates. We show that this agentic action mechanism generalizes naturally from speech recognition to complex, multiple-choice audio reasoning. Across the OpenASR leaderboard, which includes seven domain-diverse speech datasets, Speech-Hands consistently outperforms strong baselines by 12.1% WER on the OpenASR benchmark. The model also achieves 77.37% accuracy and high F1 on audio QA decisions, showing robust generalization and reliability across diverse audio question answering datasets. By unifying perception and decision-making, our work offers a practical path toward more reliable and resilient audio intelligence.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Alejandro Hernández-Cano EPFL - EPF Lausanne, Alexander Hägele EPFL - EPF Lausanne, Allen Hao Huang EPFL - EPF Lausanne, Angelika Romanou, Antoni-Joan Solergibert, Barna Pásztor ETHZ - ETH Zurich, Bettina Messmer EPFL - EPF Lausanne, Dhia Garbaya 等 102 位作者
Open LLMs enable AI practitioners to control development costs by building on an existing foundation for downstream applications. While offering substantial promise, current models often fail to meet the needs of users needing open solutions aligned with responsible AI principles, including data compliance, transparency, and inclusivity. In this work, we present Apertus, a fully open suite of large language models (LLMs) designed to address responsibility shortcomings in today’s open model ecosystem, namely data responsibility and global representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of data memorization, we also adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. Apertus also drastically expands multilingual coverage, training on 15T tokens from over approximately 1800 languages, with about 40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivaling or surpassing open-weight counterparts.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Can Jin Rutgers University, Rui Wu, Tong Che NVIDIA, Qixin Zhang Nanyang Technological University, Hongwu Peng Adobe Research, Jiahui Zhao University of Connecticut, Zhenting Wang Mohamed bin Zayed University of Artificial Intelligence, Wenqi Wei Fordham University 等 13 位作者
Ensuring that Large Language Models (LLMs) adhere to safety principles without refusing benign requests remains a significant challenge. While OpenAI introduces deliberative alignment (DA) to enhance the safety of its o-series models through reasoning over detailed “code-like” safety rules, the effectiveness of this approach in open-source LLMs, which typically lack advanced reasoning capabilities, is understudied. In this work, we systematically evaluate the impact of explicitly specifying extensive safety codes versus demonstrating them through illustrative cases. We find that referencing explicit codes inconsistently improves harmlessness and systematically degrades helpfulness, whereas training on case-augmented simple codes yields more robust and generalized safety behaviors. By guiding LLMs with case-augmented reasoning instead of extensive code-like safety rules, we avoid rigid adherence to narrowly enumerated rules and enable broader adaptability. Building on these insights, we propose CADA, a case-augmented deliberative alignment method for LLMs utilizing reinforcement learning on self-generated safety reasoning chains. CADA effectively enhances harmlessness, improves robustness against attacks, and reduces over-refusal while preserving utility across diverse benchmarks, offering a practical alternative to rule-only DA for improving safety while maintaining helpfulness.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Ziyang Wang, Jiangfeng Xiao Shenzhen University, Chuan Xiao Osaka University, Ruoxiang LI Shenzhen University, Rui Mao Shenzhen Institute of Computing Sciences and Shenzhen University, Jianbin Qin Shenzhen University
Large language models (LLMs) are expensive to serve because dense FFN blocks, multi-head attention, and KV caches dominate memory, making structured pruning a natural way to reduce serving costs under tight parameter and memory budgets. We present GRASPrune, a global budgeted structured pruning framework applied post-hoc to a pretrained model that jointly prunes FFN channels and attention KV head groups under a single global parameter budget. GRASPrune attaches lightweight learnable gates to prunable units and optimizes only these gates on a small unlabeled language-modeling calibration set, keeping all backbone weights frozen while enforcing the target sparsity at every step. A final budget-preserving scaling calibration reweights the surviving channels and heads to correct scale shifts introduced by pruning. On LLaMA-2-7B, GRASPrune removes 50% of parameters and achieves 12.18 perplexity on WikiText-2 while maintaining competitive average zero-shot accuracy on five downstream benchmarks, using a short calibration run of four epochs on 512 unlabeled sequences on a single NVIDIA A100 80GB GPU, all without any full-model fine-tuning.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Recent years have witnessed remarkable progress in Text-to-Audio Generation (TTA), providing sound creators with powerful tools to transform inspirations into vivid audio. Yet despite these advances, current TTA systems often suffer from slow inference speed, which greatly hinders the efficiency and smoothness of audio creation. In this paper, we present MeanAudio, a fast and faithful text-to-audio generator capable of rendering realistic sound with only one function evaluation (1-NFE). MeanAudio leverages: (i) the MeanFlow objective with guided velocity target that significantly accelerates inference speed, (ii) an enhanced Flux-style transformer with dual text encoders for better semantic alignment and synthesis quality, and (iii) an efficient instantaneous-to-mean curriculum that speeds up convergence and enables training on consumer-grade GPUs. Through a comprehensive evaluation study, we demonstrate that MeanAudio achieves state-of-the-art performance in single-step audio generation. Specifically, it achieves a real-time factor (RTF) of 0.013 on a single NVIDIA RTX 3090, yielding a 100x speedup over SOTA diffusion-based TTA systems. Moreover, MeanAudio also shows strong performance in multi-step generation, enabling smooth transitions across successive synthesis steps.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Sen Fang, Yalin Feng, Hongbin Zhong, Yanxin Zhang NVIDIA, Dimitris N. Metaxas
Sign Language Production (SLP) is the process of converting the complex input text into a real video. Most previous works focused on the Text2Gloss, Gloss2Pose, Pose2Vid stages, and some concentrated on Prompt2Gloss and Text2Avatar stages. However, this field has made slow progress due to the inaccuracy of text conversion, pose generation, and the rendering of poses into real human videos in these stages, resulting in gradually accumulating errors. Therefore, in this paper, we streamline the traditional redundant structure, simplify and optimize the task objective, and design a new sign language generative model called Stable Signer. It redefines the SLP task as a hierarchical generation end-to-end task that only includes text understanding (Prompt2Gloss, Text2Gloss) and Pose2Vid, and executes text understanding through our proposed new Sign Language Understanding Linker called SLUL, and generates hand gestures through the named SLP-MoE hand gesture rendering expert block to end-to-end generate high-quality and multi-style sign language videos. SLUL is trained using the newly developed Semantic-Aware Gloss Masking Loss (SAGM Loss). Its performance has improved by 48.6% compared to the current SOTA generation methods, which is a significant increase in the SLP field.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
António Loison H Company, Quentin Macé Illuin Technology, Antoine Edy, Victor Xing Illuin Technology, Tom Balough NVIDIA, Gabriel de Souza P. Moreira NVIDIA, Bo Liu NVIDIA, Manuel Faysse Facebook and CentraleSupelec 等 10 位作者
Retrieval-Augmented Generation (RAG) pipelines must address challenges beyond simple single-document retrieval, such as interpreting visual elements (tables, charts, images), synthesizing information across documents, and providing accurate source grounding. Existing benchmarks fail to capture this complexity, often focusing on textual data, single-document comprehension, or evaluating retrieval and generation in isolation. We introduce ViDoRe V3, a comprehensive multimodal RAG benchmark featuring multi-type queries over visually rich document corpora. It covers 10 datasets across diverse professional domains, comprising ~26,000 document pages paired with 3,099 human-verified queries, each available in 6 languages. Through 12,000 hours of human annotation effort, we provide high-quality annotations for retrieval relevance, bounding box localization, and verified reference answers. Our evaluation of state-of-the-art RAG pipelines reveals that visual retrievers outperform textual ones, late-interaction models and textual reranking substantially improve performance, and hybrid or purely visual contexts enhance answer generation quality. However, current models still struggle with non-textual elements, open-ended queries, and fine-grained visual grounding. To encourage progress in addressing these challenges, the benchmark is released under a commercially permissive license.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表