Les modèles de langage simplifiés désignent les systèmes de traitement du langage naturel basés sur des transformateurs fonctionnant avec moins de 7 milliards de paramètres environ — un seuil défini moins par une limite formelle que par la contrainte pratique de déploiement sur du matériel grand public, des appareils mobiles et des systèmes embarqués sans infrastructure d'inférence dans le cloud.
Ce domaine est apparu en réponse directe aux coûts de calcul et économiques des modèles à grande échelle : bien que les architectures à plus d'un milliard de paramètres démontrent une large capacité générale, leur empreinte mémoire, leur latence d'inférence et leur consommation d'énergie les rendent structurellement incompatibles avec le déploiement sur appareil, les applications sensibles à la confidentialité et les contextes opérationnels à faible bande passante ou hors ligne.
Le programme de recherche central vise à combler l'écart de capacité entre les modèles compacts et les modèles de pointe grâce à une combinaison de distillation des connaissances — l'entraînement d'un modèle étudiant plus petit par rapport aux distributions de sortie d'un modèle enseignant plus grand —, d'élagage structuré et non structuré, de quantification agressive des poids jusqu'aux représentations INT4 et INT8, et de méthodes de réglage fin efficaces en termes de paramètres telles que LoRA et QLoRA qui adaptent un modèle de base compressé à des tâches spécifiques au domaine à un coût de calcul supplémentaire minimal.
Les publications et brevets indexés ci-dessous traitent des techniques de compression de modèles, des algorithmes de quantification, des protocoles de distillation, des architectures de transformateurs efficaces, de l'optimisation de l'inférence sur l'appareil et des pipelines de réglage fin spécifiques au domaine :
Voici notre dernière sélection de publications et de brevets mondiaux en anglais sur les petits modèles de langage (SLM), parmi de nombreuses revues scientifiques en ligne, classées et axées sur petit modèle de langage, SLM, modèle de langage sur appareil, modèle de langage à la périphérie, transformateur compact, modèle à paramètres sub-7B, compression de modèle de langage, distillation de connaissances NLP, modèle de langage à élagage structuré, modèle de langage à quantification de poids, quantification NLP INT4, quantification NLP INT8, ajustement fin efficace des paramètres, ajustement fin LoRA, QLoRA, modèle linguistique d'élagage non structuré, modèle linguistique de quantification des poids, quantification INT4 NLP, quantification INT8 NLP, réglage fin efficace des paramètres, réglage fin LoRA, réglage fin QLoRA, modèle linguistique d'adaptation, inférence sur appareil, inférence NLP de bord, décodage spéculatif, transformateur de distillation de modèle, format de quantification GGUF et modèle compact de mélange d'experts.
Deformable high-strength aluminum alloy compositions and methods of making the same
Patent published on the 2026-06-04 in US under Ref US20260152827 by PURDUE RES FOUNDATION [US] (Zhang Xinghang [us], Wang Haiyan [us], Stegman Benjamin Thomas [us], Shang Anyu [us])
Abstract: [0000] An alloy comprising 92 at % aluminum, 2 at % titanium, 2 at % iron, 2 at % cobalt, and 2 at % nickel. A method of making an alloy is disclosed. The method contains the steps of providing particles of desired composition, utilizing a selective leaser melting (SLM) apparatus producing a first layer of the particles on a substrate and melting and solidifying a first group selected areas of the layer of particles, wherein the melting and the solidification results in an alloy of desired compo[...]
Our summary: The content describes a high-strength aluminum alloy with specific composition percentages. It outlines a method for creating the alloy using selective laser melting to achieve desired thickness and shape. The process involves layering particles, melting, and solidifying selected areas to form intermetallic structures.
aluminum alloy, selective laser melting, intermetallic lamellae, high-strength
Patent
Quantization-aware lora fine-tuning for llm
Patent published on the 2026-06-04 in US under Ref US20260154540 by MEDIATEK SINGAPORE PTE LTD [SG] (Lim Jia Yao Christopher [sg], Huang Ya-lin [tw], Li Huai-ting [tw], Wong Wai Mun [sg], Liang Jen-wei [tw], Lee Timothy Jun Jie [sg])
Abstract: [0000] In an aspect of the disclosure, a method of using a LoRA for inference with a FC layer of a LLM is provided. The method includes: dequantizing an INT input to an FP output; processing the FP output from the DQ and a first FP input from first weights of a down projection module of the LoRA, to output a first FP output; processing the first FP output from the first BMM and a second FP input from second weights of an up projection module of the LoRA, to output a second FP output; quantizing [...]
Our summary: The method describes using LoRA for inference in a fully connected layer of a large language model. It involves dequantizing inputs, processing them through down and up projection modules, and quantizing outputs. The final output is an INT inference result derived from the LoRA adjustments.
Quantization, LoRA, fine-tuning, LLM
Patent
Systems and methods for assisting operation and maintenance of marine machine equipment
Patent published on the 2026-06-03 in EP under Ref EP4752805 by ALFA LAVAL CORP AB [SE] (Karlsson Jimmie [se], Boman Jesper [se])
Abstract: [0001] The present invention relates to a method of operating and maintaining a piece of marine machine equipment. The piece of marine machine equipment is connected to a local processor. The method comprising the steps of obtaining a set of training data specific to the piece of marine machine equipment and training a Small Language Model (SLM) with the set of training data specific to the piece of marine machine equipment. The method further comprising the step of executing the trained SLM on [...]
Our summary: The invention describes a method for operating and maintaining marine machine equipment using a local processor. It involves training a Small Language Model (SLM) with specific training data for the equipment. The trained SLM provides offline operational advice utilizing real-time data from the equipment.
marine machine equipment, operational advice, Small Language Model, real-time data
Patent
Parameter-free method for efficient and accurate llm inference acceleration via speculative decoding
Patent published on the 2026-05-07 in WO under Ref WO2026092843 by MARZOLLO MICHELE [DE] (Marzollo Michele [de], Mueller Lorenz [de], Zhuang Jiawei [de], Roemer Niklas [de], Cavigelli Lukas [de])
Abstract: In some examples, apparatus and methods are provided for selecting a draft token sequence for verification by using a large language model, LLM. Different sources of statistics on text data (prompt, generated output, large dataset of text data) can be utilized in order to choose candidates to use for speculative decoding via look-ups.[...]
Our summary: This method accelerates LLM inference without parameters by using speculative decoding. It selects draft token sequences for verification through statistical analysis of text data. The approach utilizes various sources of statistics to optimize candidate selection for decoding.
speculative decoding, LLM inference, token sequence selection, text data statistics
Patent
Automated synthesis of planar linkage mechanisms with diverse joint types via spring-connected link models and contrastive graph learning
Published on 2026-03-28 by @OXFORD
Abstract: AbstractThe automated synthesis of planar linkage mechanisms has long been a challenge in mechanism design, requiring both geometric feasibility and motion accuracy. Recent advances in data-driven and neural network–based methods have shown promise in automating linkage synthesis, improving efficiency and scalability compared to traditional analytical or optimization-based techniques. Nevertheless, existing data-driven approaches remain limited in handling diverse joint configurations and ofte[...]
Our summary: This study presents a framework for automating the synthesis of planar linkage mechanisms using deep learning and physics-based modeling. It employs a spring-connected link model for diverse joint configurations and utilizes contrastive graph learning for efficient linkage retrieval. The method demonstrates improved accuracy and optimization stability compared to traditional approaches.
mechanism synthesis, deep learning, contrastive graph learning, optimization stability
Publication
Enhancing Whisper Fine-Tuning with Discrete Wavelet Transform-Based LoRA Initialization
Published on 2026-01-29 by Liang Lan, Molin Fang, Yuxuan Chen, Daliang Wang, Wenyong Wang @MDPI
Abstract: In low-resource automatic speech recognition (ASR) scenarios, parameter-efficient fine-tuning (PEFT) has become a crucial approach for adapting large pre-trained speech models. Although low-rank adaptation (LoRA) offers clear advantages in efficiency, stability, and deployment friendliness, its performance remains constrained because random initialization fails to capture the time–frequency structural characteristics of speech signals. To address this limitation, this work proposes[...]
Our summary: This work introduces a structured initialization mechanism combining LoRA with discrete wavelet transform for fine-tuning in low-resource ASR. The proposed DWTLoRA method enhances convergence speed, stability, and accuracy by aligning with speech signal characteristics. Experimental results show DWTLoRA outperforms standard LoRA and other PEFT methods in character error rate and training efficiency.
Fine-Tuning, Discrete Wavelet Transform, Low-Rank Adaptation, Automatic Speech Recognition
Publication
Influence and Optimization of Process Parameters on Surface Roughness of Selective Laser Melting of 316L Stainless Steel
Published on 2026-01-20 by Pin Dong, Kamonpong Jamkamon, Suppawat Chuvaree @MDPI
Abstract: To achieve better surface quality in selective laser melting (SLM), this study used 316L stainless steel powder and conducted a systematic design experiment to investigate the influence mechanism of process parameters on the surface roughness of the top and vertical surfaces. Response surface methodology (RSM) was then used for parameter optimization. The results showed that scanning speed has the greatest impact on surface roughness, followed by laser power, while scanning spacing has the least[...]
Our summary: This study investigates the impact of process parameters on the surface roughness of 316L stainless steel in selective laser melting. Scanning speed significantly affects surface quality, with optimal conditions identified for minimal roughness. The findings validate the effectiveness of the response surface methodology used for parameter optimization.
Selective Laser Melting, Surface Roughness, Process Parameters, Response Surface Methodology
Publication
A Lightweight LLM-Based Semantic–Spatial Inference Framework for Fine-Grained Urban POI Analysis
Published on 2026-01-16 by Zhuo Huang, Yixing Guo, Shuo Huang, Miaoxi Zhao @MDPI
Abstract: Unstructured POI name texts are widely used in fine-grained urban analysis, yet missing labels and semantic ambiguity often limit their value for spatial inference. This study proposes a large language model-based semantic–spatial inference framework (LLM-SSIF), a lightweight semantic–spatial pipeline that translates POI texts into interpretable, fine-grained spatial evidence through an end-to-end workflow that couples scalable label expansion with scale-controlled sp[...]
Our summary: This study introduces LLM-SSIF, a lightweight framework for translating unstructured POI texts into spatial evidence. It employs LoRA-based fine-tuning for efficient adaptation and enhances label coverage. The model demonstrates strong performance in urban analysis, revealing cultural differences between cities.
LLM, semantic inference, spatial analysis, fine-grained POI
Publication
