Tag
论文提出 FOCUS——一个基于 Mamba-2 的超轻量 EEG 认知负荷分类模型，采用固定 FFT 频带分解与权值共享编码，在仅 0.04M 参数下于 11 人试点实验中达到 74.75% 准确率与 0.08ms 前向延迟。
Vik Paruchuri 介绍 Indra，一个仅 1M 参数的 OCR 模型，在 CPU 上实现 49 页/秒，速度比 PaddleOCR 快 11 倍、比 Tesseract 快 17 倍，并在 OmniParseBench 文本识别上取得更高精度。
ej 是一个仅 11.4MB 的本地模型，专为运行时定义选项的类型化决策设计，可在 CPU 上单次前向推理输出概率分布，涵盖选择、是非与有序评分任务，并附带训练、评估与基准测试工具链。
作者用一个仅126万参数的轴向Transformer模型，将htop、vim、emacs等TUI界面的单元格解析为角色标签，再确定性地转换为A2UI声明式UI组件，客户端完全不再需要终端模拟器。模型在真实屏幕上的mIoU为0.51，约40%的屏幕可直接命中模板而无需运行模型，代码与标签已开源。
Sopro V2 Turbo is a 120M-parameter open-source TTS model with zero-shot voice cloning that streams with ~300ms time-to-first-audio on laptop CPU and runs in-browser via ONNX, reaching SOTA-level intelligibility despite its small size.
SpaFactor is a lightweight framework that predicts spatial transcriptomics from histology images by integrating spatial context and gene programs, demonstrating improved accuracy and biological fidelity across multiple cohorts.
LightMIS presents a scalable family of ultra-lightweight convolutional networks for 2D binary medical image segmentation, achieving high accuracy with reduced parameters and GFLOPs, suitable for on-device execution.
This paper analyzes the privacy-utility trade-off in LLM interactions, uncovering underlying mechanisms and introducing an intent-driven local protection framework with a lightweight model to enhance privacy while maintaining response utility.
RelateAnything is a lightweight, real-time open-vocabulary relation prediction model that accepts arbitrary predicate vocabularies and region sources, trained on a large geometrically verified dataset and evaluated on new cross-dataset benchmarks, showing significant performance gains over comparable methods.
A lightweight TTS implementation, replicating Audio8's training with 2000 hours of data, achieving a SIM metric of 0.72 in under 10 hours of training on H200.
This paper proposes LA-ReduNet, a lightweight adaptive architecture that uses hyperspherical manifold learning and adaptive step sizes to significantly reduce the number of layers needed for the MCR2 objective in neural networks, achieving comparable classification accuracy with far fewer parameters.
LiLiCorr is a lightweight likelihood-based model that improves speculative decoding by correlating per-position marginal distributions to enhance token coherence, increasing acceptance length and throughput in language model inference.
Built and released BetterGPT-150M, a compact 150M parameter causal language model that outperforms GPT-2 Small with low resource footprint. Includes live Hugging Face Space demo for text completion.
TurboVLA introduces a new Vision-Language-Action paradigm that directly maps vision and language to action, achieving 97.7% success on LIBERO with only 0.2B parameters and real-time inference at 32 Hz on consumer GPUs, significantly reducing computational cost.
The paper proposes Light-MER, a lightweight multimodal emotion recognition framework that uses knowledge distillation from an 8B teacher model to a sub-1B student, achieving state-of-the-art performance with significantly higher inference efficiency, challenging the necessity of models larger than 1B parameters.
This paper proposes a tokenizer transplantation pipeline for lightweight ASR models like Moonshine to address autoregressive collapse in Bengali. By replacing the English-centric tokenizer with a BanglaBERT WordPiece vocabulary, token fertility drops from 9.16 to 1.30 and sequence length by 85.8%, achieving 21.54% WER on the Lipi-Ghor dataset.
This paper introduces SPAM (S3M-based Phonological Activation Mapping), a method that leverages self-supervised speech models to perform both phone segmentation and recognition simultaneously using lightweight, gradient-descent-free prediction heads requiring minimal phonetic transcriptions.
MIThinker proposes a lightweight reasoning model for motivational interviewing counseling agents, trained via supervised fine-tuning and reinforcement learning to generate therapeutic thoughts, achieving MI competency comparable to state-of-the-art systems with lower computation.
This paper systematically evaluates data referencing errors (DREs) in LLMs processing tables and proposes critic-based filtering and rejection sampling to improve answer accuracy. A lightweight 4B-parameter critic model achieves high detection accuracy for DREs.
This paper proposes a lightweight multi-agent framework using AutoGen for automated concrete barrier design, achieving over 98% accuracy and showing that smaller models can outperform massive ones in this domain.