Gemma 3
PulseAugur coverage of Gemma 3 — every cluster mentioning Gemma 3 across labs, papers, and developer communities, ranked by signal.
6 天有情绪数据
-
LLM 混合方法提高 PDF 数据提取准确性
研究人员评估了从表格 PDF 文档中提取信息的三个方法，以学术课程注册表为例。策略包括仅使用大型语言模型 (LLM)、结合确定性方法和 LLM 的混合方法，以及使用 Camelot 并带有 LLM 回退的管道。实验表明，混合方法提高了元数据提取的效率，而带有 LLM 回退的 Camelot 管道实现了最高的准确性和计算效率，每份文档的提取时间不到一秒。
-
Gemma 3 微调用于卡纳达语语言模型
一位个人微调了 Google 的 Gemma 3 模型，创建了一个拥有 40 亿参数、专门针对卡纳达语的语言模型。此举旨在弥合大型语言模型在印度语言方面的能力差距。该过程涉及调整现有 Gemma 3 模型，使其能更好地理解和生成卡纳达语文本。
-
LLMs struggle with Bangla medical visual questions, new dataset shows
Researchers have developed BanglaMedVQA, a new dataset designed to evaluate Large Language Models (LLMs) and Large Vision Language Models (LVLMs) on medical visual question answering in the Bangla language. Their benchm…
-
LLM judge circuits revealed in Gemma, Qwen, Llama models
Researchers have identified a generalized 'Latent Evaluator' sub-graph within large language models like Gemma-3, Qwen2.5, and Llama-3 that is responsible for making judgments. This sub-graph is located in the mid-to-la…
-
New dataset evaluates Chinese ambiguity understanding in LLMs
Researchers have developed CHA-Gen, a new dataset designed to evaluate how well large language models understand linguistic ambiguity in Chinese. This dataset, grounded in Potential Ambiguity Theory, includes over 5,700…
-
SLMs emerge as enterprise alternative to LLMs for specific tasks
In 2026, Small Language Models (SLMs) are emerging as a viable alternative to Large Language Models (LLMs) for enterprise workloads. SLMs are suitable for narrow, well-defined tasks, data privacy concerns, edge device d…
-
AI agents' tool failures predicted; Spec Kit + Claude Code claims 90% code acceptance
A new paper introduces a method using Scale-Activation Effects (SAEs) to predict when AI agents might fail when using tools, offering internal observability. Separately, a tool called Spec Kit, combined with Anthropic's…
-
团队利用 LLM 和集成方法进行 SemEval-2026 多语言在线极化检测
研究人员为 SemEval-2026 Task 9 开发了系统，这是一项涵盖 22 种语言的多语言极化检测挑战。一种方法使用低秩适配 (LoRA) 微调 Gemma 3 模型，并使用了 GPT-4o-mini 生成的增强数据，取得了 0.811 的平均宏 F1 分数，位列第二。另一种方法侧重于使用 QLoRA 和数据增强技术（如匿名化和同形异义词替换）来微调中型 LLM，以提高鲁棒性。
-
HeadQ: 模型可见失真与分数空间校正用于KV缓存量化
研究人员正在开发几种新颖的方法来优化大型语言模型中的键值（KV）缓存，这是长上下文处理的主要瓶颈。这些方法包括训练模型内在生成可压缩表示（KV-CAT）、操纵潜在注意力空间以实现高效引导（Memory Inception）以及采用先进的量化技术，如int4和谱去噪（eOptShrinkQ、HeadQ）。此外，用于多模态模型的WindowQuant和用于分布式KV缓存管理的tierKV等新策略旨在减少延迟和内存使用，其中tierKV甚至…
-
Google releases open-weight Gemma 4 multimodal models with long context
Google DeepMind has released Gemma 4, a new family of open-weight models licensed under Apache 2.0, marking a significant advancement in their open-source AI offerings. The models are designed for reasoning and agentic …
-
Google DeepMind expands Singapore presence to advance AI in Asia-Pacific
Google DeepMind is expanding its presence in Singapore by opening a new research lab to advance AI in the Asia-Pacific region. This expansion aims to foster linguistic and cultural inclusivity for AI models, enhance Gem…
-
Meta releases Llama 3.1, Google launches Gemma 3
Meta has released Llama 3.1, an updated open-source large language model available in 405B, 70B, and 8B parameter sizes. Google has also launched Gemma 3, a new multimodal and multilingual model with a long context wind…