Quantization, distillation, pruning and serving work aimed at the same accuracy for less memory, latency and money.
arXiv:2609.23606v1 Announce Type: new Abstract: Mesh texture compression typically relies on 2D UV atlases, whose chart discontinuities and mapping overhead can limit coding efficiency. To tackle thi...
arXiv:2609.23715v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) incur substantial computational overhead due to the reliance on hundreds of visual tokens to represent images....
arXiv:2609.23733v1 Announce Type: new Abstract: Feed-forward visual geometry models such as the Visual Geometry Grounded Transformer (VGGT) have recently enabled direct 3D reconstruction from multi-v...
arXiv:2609.23815v1 Announce Type: new Abstract: Medical image segmentation models typically rely on large amounts of densely annotated volumetric data, limiting their scalability across tasks and ima...
arXiv:2609.24872v1 Announce Type: new Abstract: Two-stage oriented detectors achieve high localization accuracy in synthetic aperture radar (SAR) ship detection, but their large backbones, feature py...
arXiv:2506.11784v2 Announce Type: replace Abstract: Vision Transformers (ViTs) are essential in computer vision but are computationally intensive, too. Model quantization, particularly to low bit-wid...
The paper presents a discrete generative model for neuronal spiking activity recorded on microelectrode arrays. It uses a shared vocabulary of spatiotemporal motifs learned by a residual vector‑quantized autoencoder and predicts motif occurrence with a factorized masked transformer. Evaluated on 31 assays from human brain organoids and ex vivo hippocampal tissue, the model achieves superior reconstruction and generation performance compared to baselines and shows that motifs are largely reused across assays.
NAVIR is an end‑to‑end audio‑visual speech recognition system designed for the BrainChip Akida neuromorphic processor, which only supports sequential 2‑D convolutions. The architecture separates spatial and temporal encoding into three AkidaNet modules—per‑frame visual, temporal video, and spectrogram audio encoders—fused by a lightweight predictor and decoded with constrained beam search. Trained with CTC on noise‑augmented audio and fine‑tuned via quantization‑aware training, the quantized model achieves 14.0% WER on GRID’s unseen‑speaker split and 3.3% on overlapped‑speaker split, outperforming audio‑only baselines, and delivers 98.6% command accuracy at 1.5% WER on an industrial‑command corpus, while offering a 13‑fold energy advantage over conventional ANNs and roughly 5‑fold lower energy per inference than a Raspberry Pi CPU.
The paper presents a causal analysis of a compressed VLA policy that performs well in offline tests but fails in closed‑loop execution on a simulated pick‑and‑place task. An 8‑layer distillation of Octo‑Base retains most parameters and passes all offline metrics, yet collapses during deployment, with early stages degrading gradually and final transport failing entirely. The failure is traced to a negative, late‑heavy residual in the action trace, and standard remedies (continued training, offline data, command‑level compensation, clamping) do not restore performance; only a minimal‑pair intervention that mixes deployment‑distribution rollouts with teacher data restores parity with the teacher. whyItMatters":"The study demonstrates that offline validation metrics alone are insufficient to guarantee closed‑loop success for compressed policies, highlighting the need for targeted deployment‑time testing and interventions."
The paper presents an end‑to‑end workflow for deploying a recurrent probabilistic neural network (RPNN) on FPGA hardware to enable real‑time plasma state estimation for Tokamak control. By reducing architecture size and applying quantization‑aware training with QKeras, the model is synthesized with hls4ml for a Xilinx Alveo U50 device, achieving deterministic sub‑10 µs single‑timestep latency while staying within all four resource budgets (DSP, LUT, FF, BRAM).
GraphSVR is a q‑space‑aware graph‑based framework for 4D slice‑to‑volume registration in diffusion MRI. It models slice groups as nodes in a graph whose edges encode temporal, spatial, and diffusion‑encoding relationships, and uses a graph neural network to predict globally consistent stack‑wise rigid motion in a self‑supervised, zero‑shot manner. In synthetic and realistic simulations, GraphSVR reduces grid and rotation errors by up to 73% compared to the standard FSL eddy method, especially under severe motion and sparse‑direction regimes.
Teacher Should Think Ahead: Adaptive Continuations for Reliable On-Policy Distillation explores how on‑policy distillation (OPD) can be improved by addressing teacher uncertainty that contracts as the teacher continues from a student‑generated prefix. The authors identify Teacher Uncertainty Contraction (TUC) and theoretically analyze its variance‑bias trade‑off, leading to the proposal of Adaptive‑Continuations On‑Policy Distillation (AC‑OPD). Experiments on mathematical reasoning and code generation show that AC‑OPD consistently outperforms standard OPD, with controlled‑continuation and matched‑budget analyses supporting the adaptive‑continuation design.
arXiv:2609.22295v1 Announce Type: new Abstract: Event pipelines often discretize asynchronous timestamps before learning. This step looks harmless, but its stability depends directly on temporal reso...
arXiv:2609.22566v1 Announce Type: cross Abstract: Knowledge distillation (KD) aims to compress high-performance teacher LLMs into lightweight students. However, distilled students often exhibit subst...
arXiv:2609.22913v1 Announce Type: cross Abstract: Talking-head and dyadic models now achieve real-time inference, yet fast motion generation alone does not produce an interactive conversation. A live...
ValueDiff introduces a value‑geometric KV cache eviction strategy for large language models that suppress attention sinks. It ranks tokens by the L2 deviation of their value vectors from the cache mean, a score that aligns with minimal‑disturbance eviction under a max‑entropy assumption. Across several benchmarks—RULER, LongBench, and MATH‑500—ValueDiff consistently retains a higher proportion of useful tokens than prior methods, especially under tight cache budgets.
RPMem introduces a two‑stage architecture that compiles each session into a model‑independent latent memory and then consolidates it with retained memory via a task‑trained recurrent gate. The consolidated memory is mapped to backbone‑specific low‑rank adaptation (LoRA) parameters, enabling the memory to transfer when the backbone is replaced. Across three long‑term memory benchmarks and five diverse backbones, RPMem achieves broad generalization with near‑constant update cost and memory footprint, outperforming existing parametric and text‑based baselines on the PERMA benchmark.
arXiv:2609.23587v1 Announce Type: new Abstract: Large reasoning models (LRMs) incur high inference costs, often mitigated by efficiency techniques like quantization and pruning. However, the impact o...
arXiv:2609.23590v1 Announce Type: new Abstract: Battery energy storage system (BESS) dispatch must preserve operational feasibility while declining price spreads reduce the margin available to pay fo...
arXiv:2609.23892v1 Announce Type: new Abstract: Mechanistic interpretability defines features as the fundamental units of a neural network and circuits as the weighted subgraphs that carry out its co...