The paper introduces Reward‑Tilted On‑Policy Distillation (RT‑OPD), a method that enhances acoustic grounding in audio‑language models by using a frozen teacher to generate a reward based on the contrast between token predictions with and without audio. This reward reshapes the teacher distribution for reverse‑KL distillation, encouraging students to rely more on acoustic evidence. Experiments on two compact students across three benchmarks show that RT‑OPD consistently outperforms vanilla OPD, and a 3B model trained with RT‑OPD achieves 72.72% accuracy on the MMAU benchmark, surpassing other 3B models and rivaling larger 7B and 8B models.
By Kaiyang Li, Shaobo Han, Yue Tian, Shihao Ji
The paper introduces Pre-hoc Sparsity (PrHS), a method that selects key-value (KV) cache entries before attention scoring to avoid posterior bias in large language model inference. By bounding mutual‑information loss through the dropped attention mass, PrHS offers explicit accuracy control and implements three orthogonal selectors across time, depth, and layer. Experiments on LLaMA and Mistral models show that PrHS cuts retrieval overhead by over 90%, achieves higher sparsity than HShare, and delivers significant speedups and reduced FLOPs on NVIDIA A100 GPUs while maintaining near‑dense accuracy.
By Yifei Gao, Lei Wang, Rong-Cheng Tu, Qixin Zhang, Jun Cheng, Dacheng Tao
The paper introduces Smol‑VL‑BLV, a compact vision‑language model designed for blind and low‑vision users. It employs a 500M decoder transformer with teacher‑student distillation and Group Relative Policy Optimization to add spatial detail, directional cues, and hazard detection to post‑training. After a lightweight finetuning step, the model achieves significant gains on spatial, social, OCR, and VQA benchmarks while remaining under 450 MB and running entirely offline on a mid‑range Android phone.
By Rishabh Choudhary, Shreyansh Raj, Umesh Goyal, Shubh Kashyap, Shrestha Kumar, Sushovan Jena, Komal Kumar, Hisham Cholakkal, Aditya Nigam
This study evaluates the instance‑segmentation performance of YOLOv11 and YOLOv8 on immature green apples in orchard settings. YOLO11n‑seg achieved the highest mask precision (0.831), while YOLO11m‑seg and YOLO11l‑seg excelled in non‑occluded and occluded fruitlet segmentation. YOLOv8n, however, outperformed the YOLO11 series in inference speed, reaching 3.3 ms compared to 4.8 ms for the fastest YOLO11 model.
By Ranjan Sapkota, Manoj Karkee
RotVLA introduces a Vision‑Language‑Action framework that replaces discrete latent action encoding with a continuous rotational latent action representation on the group SO(n). This design provides continuity, compositionality, and structured geometry that better capture real‑world action dynamics, and a triplet frame learning scheme enforces meaningful temporal dynamics while preventing degeneration. Trained with 1.7 B parameters on large cross‑embodiment datasets, RotVLA achieves state‑of‑the‑art performance on LIBERO and RoboTwin2.0 benchmarks and shows strong real‑world manipulation results.
By Qiwei Li, Xicheng Gong, Xinghang Li, Peiyan Li, Quanyun Zhou, Hangjun Ye, Jiahuan Zhou, Yadong Mu
Task-Aware Spectral Pruning (TASP) is a post‑training framework that tailors sparse masks to specific tasks by calibrating module‑level spectral descriptors against task‑specific ablation effects. It constructs masks that close grouped‑query‑attention and SwiGLU dependencies, routing each user turn to a single compiled mask that remains fixed during prefill and decoding. In experiments, TASP achieves a 43% active‑FLOP reduction while preserving 97.7% of the dense BF16 performance on Llama‑3‑70B, and delivers a 1.44× speedup on an A100 80GB with INT8‑weight/BF16‑compute, reducing decode latency from 45.2 to 31.3 ms/token.
By Ibne Farabi Shihab, Fariya Afrin, Sanjeda Akter, Anuj Sharma
The paper introduces COSA-GS, a new compression method for 3D Gaussian Splatting that avoids spatial aggregation by using anchor-wise causal factorization. It builds a compact learnable anchor latent from geometry context and fuses it with the geometry context to create an anchor context for attribute coding, employing only linear transformations and activations. The method is trained with rate–distortion optimization, adaptive Gaussian pruning, and quantization-aware training to ensure bit‑exact entropy decoding across platforms, achieving state‑of‑the‑art compression performance with fast, consistent cross‑platform decoding.
By Pengpeng Yu, Yueru Chen, Fei Song, Tai Qin, Qi Zhang, Jing Wang, Yulan Guo
The paper introduces a framework for intrinsic‑extrinsic coupling in learning dynamics, where a learner’s current observations do not solely dictate its future training responses. It formalizes this coupling through a continuation‑conditioned value of a constrained learning‑state intervention and employs an executable finite‑frame classifier‑head to protect current logits while adjusting historical margins. Experiments across CLINC‑derived class‑incremental settings, output distillation with RoBERTa, and SGDW dynamics reveal that coupling can produce both positive and negative interactions, and that coordinated interventions can match or exceed development gains while reducing cross‑entropy loss compared to standard replay.
Accelerating Video Diffusion via Training-Free Trajectory Routing (TRACK) introduces a heterogeneous denoising strategy that switches between large and small diffusion models at selected steps, determined by a calibration process that measures disagreement between model predictions. By routing quality-sensitive steps to the large model and low-disagreement steps to the small model, TRACK achieves significant speedups—up to 2.73×—across several video diffusion benchmarks while maintaining comparable quality and diversity. The method requires no retraining, architectural changes, or online dual-model evaluation, making it a practical acceleration paradigm for video diffusion.
The paper examines on‑policy self‑distillation (OPSD) for multi‑turn agents, showing that using privileged information (PI) in the loss can make agents appear confident yet underperform plain RL, sometimes worse than the untrained base model. To address this, the authors propose Privileged Self‑Practice (PSP), which keeps PI in the prompt and uses it only during sampling, not in the loss. PSP consistently outperforms plain GRPO across AppWorld and SWE‑bench Verified, improving task‑goal completion by up to 65% and resolved rate by up to 61%.
The paper introduces αp-LoRA, a rank-allocation strategy for low-rank adaptation (LoRA) in large language models that uses π-regularization (0 < p < 1) to induce sparsity in rank-one components. By regularizing the energy of each component, redundant parts are encouraged to vanish while important ones are retained, and the authors derive a proximal subproblem that reduces the matrix optimization to a two‑dimensional thresholding criterion. Experiments on natural language understanding and question‑answering tasks show that αp-LoRA achieves performance competitive with existing LoRA baselines.
DeltaS is a query‑agnostic, training‑free method for evicting key‑value cache entries in hybrid video‑language models that combine linear and full attention. It uses the change in the recurrent state of gated‑delta linear attention—called state drift—to decide which video chunks to keep, selecting those that induce larger normalized state changes. In experiments with a fixed memory budget, DeltaS outperforms position‑, attention‑, and key‑value‑based eviction signals, improving performance by 2.1 points on average across six long‑video benchmarks and 5.6 points on the longest benchmark, while adding only 1.9% of the forward‑pass cost.
By Taeyoun Kwon, Seungjin Kim, Hyeonyu Kim, Moon Hwan Kim
The paper introduces a task‑aware quadratic unconstrained binary optimization (QUBO) surrogate for mixed‑precision quantization, separating weight and activation profiles and incorporating a bit‑operation (BOP) cost and structural priors. Using this surrogate, a network‑wide allocation is refined via a direct validation‑based PROTES search, achieving a 37.192 dB PSNR on a compact NAFBlock denoiser with 4.035% routed‑layer BOPs, slightly better than a HAWQ‑style baseline. The study also shows that LSQ+ refinement narrows the quality gap and that the benefit of refinement varies with architecture and recovery strategy.
By Osama Orabi, Artur Zagitov, Hadi Salloum, Viktor A. Lobachev, Yaroslav Kholodov
The paper presents a new evaluation protocol for post‑training quantization of speech language models that separates lexical output, transcript‑insufficient endpoints, and packed implementations. In a Qwen2‑Audio case study, a 6‑bit allocation selected for translation improves chrF scores but degrades emotion recognition, while uniform and front‑layer controls perform better on emotion tasks. Similar patterns hold at 7 bits, and a 4‑bit study shows consistent emotion deficits across all low‑bit allocations, with no advantage for the selected scheme. The study highlights a precision‑dependent mismatch between lexical output, waveform‑dependent behavior, and nominal precision, without claiming a general failure of low‑bit models or a deployment benefit for the selected allocation.
By Mengzhe Geng, Jinxi Jin, Junhao Xu
UVU is a vision-language unified autoregressive framework that integrates visual supervision directly into the pre-training stage of multimodal large language models. By using continuous visual encoding and a large-scale iterative hierarchical clustering algorithm to build a pixel-level visual codebook, UVU enables lossless representation of visual inputs and autoregressive generation of pixel-level image tokens alongside textual tokens. This approach synergizes pixel-level visual perception with semantic-level visual understanding, allowing models to internalize visual reconstruction capabilities and improve multimodal understanding performance.
By Zhehan Kan, Xinghua Jiang, Yubo Zhu, Yanlin Liu, Xiaochen Yang, Zhixiang Wei, Shifeng Liu, Qingmin Liao, Wenming Yang, Xin Li, Yinsong Liu, Deqiang Jiang, Xing Sun
The paper introduces an automated high‑throughput microscopy system for melissopalynology that combines H∞ robust mechanical control with deep learning pipelines. It uses U^2‑Net for salient object detection and a DINOv2 Vision Transformer trained via deep metric learning for pollen grain classification, augmented with Gradient‑Weighted Attention for interpretable texture annotations. The system reports a 95.8% classification recall and at least a six‑fold speedup over manual expert analysis.
By J. Staforelli-Vivanco, R. Jofr\'e, P. Coelho, I. Sanhueza, L. Viafora, C. Toro, J. Troncoso, M. Rondanelli-Reyes, I. Lamas, Andy Banegas-Medina, Isis-Yelena Montes, B. Mu\~noz-Cepeda, V. Salamanca-Levi, M. Gonz\'alez-Ortiz, E. Vera
The paper proves that training a binary quantized neural network (2-QNNT) is W[1]-hard when parameterized solely by the sum of input and output dimensions, α+ω. This hardness result holds even for zero training error on a specially constructed dataset where each input equals its target and the examples form a coordinate‑wise prefix chain. The proof reduces from DAG edge‑disjoint paths, employing a one‑flip routing equivalence that links activation transitions to vertex‑disjoint paths in the network.
By Tao Jiang, Minbo Gao, Shaowei Cai
PipeLive introduces a method for live, in‑place reconfiguration of pipeline parallelism in large language model serving. By redesigning the KV cache layout and extending PageAttention, it enables dynamic resizing of the cache without interrupting inference. The system also uses an incremental KV patching mechanism to keep KV states consistent during reconfiguration, achieving significant reductions in reconfiguration time and improvements in latency metrics.
By Xu Bai, Muhammed Tawfiqul Islam, Chen Wang, Adel N. Toosi
The paper "Reachable Global Optimization in AI Systems: How Global Is Global?" argues that claims of AI systems optimizing various components (prompts, policies, architectures, etc.) are underspecified unless they define the region actually reachable by the system. It introduces Reachability-Induced Optimization (RIO), a framework where a generator, verifier, controller, memory, tools, and budget determine a reachable candidate region, and proves several theoretical results about reachable-optimality and related concepts. Extensive benchmarks (66,150 trials across 270 landscapes) demonstrate that control can alter reachability, and that optimization quality, reachability quality, and control reliability must be reported separately.
By Wesley Shu
SlackDrive is a pre‑inference compute allocator that dynamically selects the compute budget for each driving control step by reusing the realized latency from previous inferences. By profiling a small set of discrete budgets once, it estimates the current compute state online and chooses the highest‑utility budget that stays within the admissible latency envelope. On the NAVSIM v2 benchmark with DriveDreamer‑Policy, SlackDrive boosts latency‑constrained EPDMS performance by 21.7% compared to the best baseline, while full‑budget and token‑pruning approaches exceed the latency limits under runtime contention.
By Xiaohuan Pei, Hengguang Zhou, Yuanhao Ban, Justin Cui, Jiaqi Feng, Haoyu Xie, Tao Huang, Pichao Wang, Yanchao Yang, Cho-Jui Hsieh