The paper introduces IDeaL, a data‑free multi‑teacher distillation technique that generates teacher‑specific, improved samples using decorrelation losses at patch and image levels. By tailoring noise to each teacher, IDeaL produces strong student models that capture complementary teacher information and achieve results close to those distilled from real images. Experiments demonstrate that with only 1,000 images, students trained on IDeaL samples match or exceed the performance of students distilled from a 1,000‑image subset of ImageNet.
By Feyza Yavuz, Mert B\"ulent Sar{\i}y{\i}ld{\i}z, Diane Larlus
Minima-KV introduces a retention‑preserving hierarchy for mixed‑format paged attention that keeps recent and protected anchor pages in FP8 while older pages are compressed into packed TQ3, allowing every live‑request page to remain addressable. The approach uses format‑specific kernels and a globally normalized online‑softmax merge to compute partial attention states, enabling direct heterogeneous decoding without a dense shadow cache. Experiments on Qwen3.6‑27B on a 96‑GB NVIDIA RTX PRO 6000 Blackwell GPU show 3.50× compression over BF16 and 1.75× over FP8, with minimal impact on performance across long‑context benchmarks.
By Sergii Kozyrev (Minima AI, Inc), Davyd Maiboroda (Minima AI, Inc)
PROOF-Gen is a method that improves distillation of tool‑calling models by recovering successful trajectories from teacher failures. It uses per‑scenario prompt optimization to generate corrective guidance that steers the teacher to a passing trajectory, then removes this guidance before training so the student learns from clean demonstrations. On τ2‑bench, PROOF-Gen recovers 93% of failed scenarios, boosting Qwen3‑4B‑Instruct‑2507’s Pass^1 from 0.132 to 0.529 and improving Gemma 4 E4B‑it by 7.2pp on BFCL v4 multi‑turn, while also raising deployed on‑device model performance by up to 5.0pp across response‑quality metrics.
By Anh Ta, Junjie Zhu, Shahin Shayandeh
arXiv:2608.24207v1 Announce Type: new
Abstract: Semantic identifiers (SIDs) represent entities as hierarchical token sequences for generative retrieval and recommendation. Residual-quantization token...
By Yunxiao Luo, Siyuan Wang, Ben Chen, Chenyi Lei
arXiv:2608.24696v1 Announce Type: cross
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) and on-policy distillation (OPD) have become two widely adopted paradigms for post-training lar...
By Wenze Lin, Jiale Zhao, Xitai Jiang, Songde Rao, Yining Li, Shenzhi Wang, Bingxiang He, Gao Huang
QABBA is a quantized variant of the Adaptive Brownian Bridge-based Aggregation (ABBA) method that compresses long numerical time‑series into short symbolic sequences using integer‑quantized symbolic centers. It reduces parameter storage and enables integer arithmetic while providing rigorous error bounds—including a dimension‑free segment error, a time‑domain reconstruction error, a stability condition for symbolic assignment, and a bit‑allocation rule. The resulting symbolic strings can be fed directly into pretrained large language models, and experiments on several benchmark archives show a practical trade‑off among storage, reconstruction accuracy, and downstream predictive performance.
By Erin Carson, Xinye Chen, Fei He, Cheng Kang
The paper introduces an elastic key‑value (KV) cache for large language model (LLM) serving that dynamically reclaims a pre‑allocated reserve during decode‑heavy phases and restores it before prefill, using a userspace CUDA virtual‑memory trick that requires no driver changes. The authors implement this mechanism, test it under realistic workloads, and find that it offers only marginal benefits—about a 1 % difference in time‑to‑first‑token for large prefill chunks—and that simpler strategies such as lowering the maximum batch size can achieve similar results. The study also notes that the reserve’s impact diminishes with higher tensor‑parallelism levels.
whyItMatters":"The work demonstrates that a dynamic KV cache reclamation strategy can be implemented without driver patches and that its practical benefits are limited, guiding future LLM serving optimizations toward simpler approaches."
By Sathishkumar Sivashanmugam
The Transformer Accelerator (TFA) is a synthesizable, parameterizable INT8 memory‑to‑memory engine designed for transformer inference and machine translation. It features a one‑time‑multiplexed datapath that handles prompt processing and autoregressive generation, and implements key operations such as matrix multiplication, softmax, RMSNorm, and elementwise functions through eight 512‑bit macro‑op descriptors. In extensive verification, TFA achieved zero mismatches across 25 tests and 34 constrained‑random runs, matched floating‑point references on multiple translation tasks, and delivered a 20× speedup over a 22‑thread CPU while projecting significant energy reductions in larger designs.
By Shashank
arXiv:2606.22975v2 Announce Type: replace
Abstract: Text-attributed graphs (TAGs) are widely used in many real-world domains, and learning on TAGs requires jointly modeling text semantics and graph s...
By Yeongho Kim, Yeonje Choi, Kijung Shin
RT-NeuS is a neuro‑symbolic framework for long‑form video question answering that retains the accuracy and formal guarantees of temporal‑logic‑guided methods while dramatically reducing inference latency. It achieves this by using coarse‑to‑fine adaptive sampling to focus on query‑relevant frames and batched proposition detection with KV‑cache reuse, enabling all propositions to be evaluated in a single forward pass. Experiments on LongVideoBench, Video‑MME, and MLVU show up to a 13× speed‑up on an NVIDIA H200 GPU while matching or surpassing prior neuro‑symbolic accuracy.
By Shawn Liang, Sahil Shah, Chengwei Zhou, S P Sharan, Harsh Goel, Arnab Sanyal, Sandeep Chinchali, Gourav Datta
arXiv:2608.23879v1 Announce Type: cross
Abstract: Cardiac cine-MRI serves as a direct visual indicator of cardiovascular hemodynamics by capturing the continuous wall motion of the aorta. Quantifying...
By Dexter Wen Jie Teo, Nairouz Shehata, Herve Lombaert
arXiv:2605.30972v2 Announce Type: replace
Abstract: Accurate 3D medical image segmentation requires both fine spatial detail and long-range volumetric context. Although Mamba provides efficient long-...
By Bakht Zada, Chao Tong, Qile Su, Shuai Zhang
arXiv:2608.24549v1 Announce Type: new
Abstract: Persistent entropy is the Shannon entropy of a persistence-based probability measure defined on a persistence diagram. However, its cross-entropy versi...
By Sijin Yeom, Jae-Hun Jung
arXiv:2605.28791v2 Announce Type: replace-cross
Abstract: On-policy self-distillation (SD) improves LLM reasoning by using teacher-side privileged information (PI) to turn sparse verifier outcomes in...
By Jiazhen Huang, Xiao Chen, Xiao Luo, Yong Dai, Senkang Hu, Yuzhi Zhao
The paper compares two strategies for handling memory limits in large language model (LLM) serving: tensor parallelism, which distributes weights and KV cache across multiple GPUs, and KV compression, which reduces cache size via quantisation and eviction on a single GPU. Using a cost‑normalised simulator calibrated on A100, A40, and H100 hardware, the authors find that across two models (Llama‑2 7B and 70B) and various GPU configurations, compression consistently outperforms tensor parallelism in cost per million tokens, offering 1.20× to 2.00× savings. The study identifies a model‑size threshold (~36B parameters on an 80 GB card) where compression dominates, while tensor parallelism becomes necessary only for larger models where weights alone exceed a single GPU’s capacity.
By Srikanta Datta Tumkur, Mehar Simhadri, Anshu Bansal, Jay Iyer, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly, Raj Dandekar
VisCache introduces a two-stage, plug‑and‑play framework for pruning visual key‑value caches in Vision Large Language Models without retraining. The first stage filters out temporally redundant keyframes, while the second stage, PruneKV, applies a parabolic layer‑wise budget and asymmetric update to selectively prune keys and fuse values, preserving essential context. Experiments show up to 2.35× speedup and significant memory savings with only 19–28% of the original cache retained, outperforming existing baselines.
By Lyuke Wang, Zhuo Li, Guangxu Zhu
arXiv:2304.10041v2 Announce Type: replace
Abstract: This work investigates formal policy synthesis for continuous-state stochastic dynamic systems subject to high-level specifications expressed in li...
By Lening Li, Zhentian Qian, Jianan Xia, Qiren Geng, Huasheng Zhang, Liang Hu, Qishuang Li, Junqiang Lou
SenseShift is an encoder-based framework that enables fine‑grained, sentence‑level sentiment control in text generation. It uses bidirectional attention, quantized sentiment signals, and iterative mask infilling to generate local sentences conditioned on target sentiment intensity. Experiments on story and review generation show that SenseShift delivers stronger sentiment controllability while preserving text quality and robustness to out‑of‑domain inputs compared to larger decoder‑based baselines.
By Shahed Masoudian, Markus Frohmann, Emmanouil Karystinaios, Navid Rekabsaz, Markus Schedl
arXiv:2608.24674v1 Announce Type: new
Abstract: Joint text-to-video-audio generation produces synchronized visual and acoustic content, but the long sampling trajectories and heterogeneous multimodal...
By Xiaoda Yang, Yuxiang Liu, Kaiwen Zheng, Yuan Liu, Yibo Lai, Shengpeng Ji, Kai Jiang, Jianfei Chen, Xiaobin Hu, Shuicheng Yan, Jintao Zhang, Jun Zhu, Zhou Zhao
Masked diffusion language models (dLLMs) promise faster text generation by denoising multiple tokens simultaneously, yet their real‑world serving behavior has been largely unexamined. Using LLaDA‑8B‑Instruct on a single NVIDIA H200 GPU, the study finds that request difficulty is discretized into 11 step‑count levels, short‑budget benchmarks underestimate serving variance, and only 24% of single‑request time is GPU computation, with batching mainly reducing CPU dispatch overhead. The authors also demonstrate that output quality remains stable across batch sizes and propose a batch‑timeout rule for synchronized batching under Poisson arrivals.
By Farhana Amin, Sabiha Afroz, Mona Moghadampanah, Dimitrios S. Nikolopoulos