Gemma-3-4B-IT
PulseAugur coverage of Gemma-3-4B-IT — every cluster mentioning Gemma-3-4B-IT across labs, papers, and developer communities, ranked by signal.
-
New diagnostic suite reveals record-use errors in multimodal AI models
Researchers have developed a new diagnostic suite called RecordAuth-Diag to evaluate how personalized multimodal models handle record use. The suite identifies instances of visual memory misbinding, where a record is ap…
-
New audit method reveals complex weight changes in specialized medical AI models
Researchers have developed a new method to audit the weight changes in AI models when they are specialized for medical tasks. This weight-delta audit approach was applied to two pairs of models: Gemma-3-4B-IT to MedGemm…
-
New LoRA and Activation Steering Methods Improve Gender-Inclusive Language Generation
Researchers have developed a new method using parameter-efficient Low-Rank Adaptation (LoRA) for gender-inclusive language generation, achieving a score of 80.00% on the LT-EDI 2026 Shared Task. Additionally, they intro…
-
IndicGuard: New safety model and dataset for Indic languages launched
Researchers have developed IndicGuard, a new multilingual safety model and dataset designed to address the limitations of English-centric safety mechanisms for Large Language Models (LLMs) in the Indic region. The model…
-
New IndicGuard model enhances LLM safety for Indic languages
Researchers have developed IndicGuard, a new multilingual safety guard model and dataset specifically designed for Indic languages. This model addresses the limitations of existing English-centric safety mechanisms by c…
-
New Residual Paving method enhances LLM editing and control
Researchers have developed a new method called Residual Paving to improve the control and editing of large language models. This technique separates the decision of whether to intervene in a model's output from the actu…
-
New AI research tackles multimodal finetuning, image editing, and verification
Researchers have developed TRACER, a novel method for robust multimodal finetuning that addresses catastrophic forgetting by using a Weighted Moving Average (WMA) teacher. This approach improves out-of-distribution accu…