Entity · model
Gemma
modelactive
gemma-f6a2f419·20 events·first seen Aliases: Gemma
Co-occurring entities
Google DeepMindGoogleHugging FaceQwenLlamaGemma 3nInstruction-Tuned Models Locally Reuse Human Syntax More Than Humans DoLoRAPortLLMMistralInternVLMuSciClaimsSciVerGRPO (Group Relative Policy Optimization)ToolSciVerMESAMESA: Prioritizing Vulnerable Communication Channels for Securing Multi-Agent SystemsLangGraphSafeVecRAS: Measuring LLM Safety Through Refusal Alignment
More like this (12)
Recent events (20)
Instruction-tuned LLMs show stronger syntactic convergence toward human interlocutors than humans do
A new arXiv preprint measures syntactic convergence in 16 open-weight Llama and Gemma models (1B–70B, pretrained and instruction-tuned) by substituting model outputs into pre-existing human dialogues and measuring context-free grammar rule reuse. Every model showed greater CFG-rule overlap with the preceding human turn than a random baseline, and instruction-tuned models exceeded the syntactic convergence of the human responses they replaced. Instruction tuning increased local syntactic and semantic alignment with interlocutors, though it also raised baseline overlap with unrelated primes, suggesting a more globally accommodating output style rather than purely targeted priming.
Theoretical and empirical analysis of long-term temporal portability of LoRA patches across continual pretraining updates
This arXiv paper investigates PortLLM, a training-free and data-free scheme for adapting LLMs after continual pretraining, extending prior short-term results to 10 continual pretraining steps across Mistral, Gemma, and Qwen base models. The authors find that LoRA patches remain portable across longer update horizons, suggesting repeated fine-tuning is unnecessary when base models are periodically updated. Two theoretical analyses are offered, identifying near-orthogonality of high-dimensional vectors as the geometric mechanism underlying temporal portability. The work has practical implications for reducing fine-tuning overhead in production deployments with frequently updated base models.
ToolSciVer: Tool-augmented reinforcement learning for multimodal scientific claim verification
Researchers introduce ToolSciVer, a framework that equips vision-language models with three type-aware visual tools (table focus, chart-to-structure parsing, high-resolution zoom) to verify scientific claims grounded in figures, tables, and charts from papers. The policy is trained using Group Relative Policy Optimization (GRPO) with a composite reward covering correctness, format, tool-use efficiency, and validity. Experiments across five VLMs from three model families (Qwen, InternVL, Gemma) on SciVer and MuSciClaims benchmarks show improvements over prompting-based and RL-based baselines. The work is notable as the first tool-augmented framework specifically targeting multimodal scientific claim verification.
MESA framework proactively ranks vulnerable communication channels in multi-agent systems
Researchers introduce MESA, a label-free framework for prioritizing security-critical communication edges in multi-agent systems (MAS) before attacks are observed. The framework combines six graph-theoretic metrics with two dynamic probes (ablation and masking) to rank edges by compromise risk, without requiring attack traces. Evaluated across three MAS scenarios, eight network topologies, and five open-source LLMs, MESA achieves mean Spearman ρ=+0.60 correlation with empirical per-edge attack success, and monitoring the top 10% of ranked edges intercepts roughly 3x more successful attacks than random allocation. The work highlights that attack impact in MAS is highly concentrated — a single compromised edge can account for up to 75% of total attack success.
SafeVec and RAS: White-box LLM safety evaluation via internal refusal representations
Researchers introduce SafeVec, a white-box safety evaluation procedure that measures LLM safety from internal hidden-state representations rather than generated outputs. The method extracts layer-wise refusal directions from a safety-aligned reference model, identifies stable layers where safe and unsafe behaviors are separable, and scores target models via a calibrated 0-100 Refusal Alignment Score (RAS). Evaluated across Llama, Gemma, and Qwen model families, RAS distinguishes aligned from uncensored/abliterated variants and correlates with output-level attack success rates while being substantially faster than judge-based evaluation. The approach addresses key limitations of output-level safety evals: cost, judge sensitivity, and dependence on fixed question banks.
Contrastive-Difference CKA reveals concept-specific structural alignment across LLM architectures
Researchers introduce CKA_Delta (contrastive-difference CKA), a training-free diagnostic that isolates concept-specific representational convergence from generic similarity across LLM architectures. The method reveals a geometric-functional universality dissociation: moderate geometric alignment coexists with near-perfect functional transfer across six concept domains and multiple architectural families. CKA_Delta also functions as an architectural outlier detector, flagging Gemma as a notable outlier (d=1.08, AUC=0.79). The work provides a practical tool for cross-architecture concept monitoring without requiring model training.
DeepMind announces DiffusionGemma with 4x faster text generation
DeepMind published a blog post introducing DiffusionGemma, a diffusion-based variant of the Gemma model family claiming 4x faster text generation. The announcement suggests a departure from standard autoregressive decoding in favor of diffusion-based generation. If the claims hold, this could represent a meaningful inference efficiency advance for the Gemma line.
Welcome Gemma - Google's new open LLM
Google released Gemma, a family of open-weight large language models, announced via the Hugging Face blog. The models are positioned as Google's entry into the open-weights LLM space, following the success of models like Llama 2. This release marks a significant strategic move by Google to compete in the open-source AI ecosystem.
Fine-Tuning Gemma Models in Hugging Face
Hugging Face published a guide on fine-tuning Google's Gemma models using parameter-efficient fine-tuning (PEFT) techniques. The post covers practical workflows for adapting Gemma to downstream tasks within the Hugging Face ecosystem. This represents part of the broader tooling support rollout accompanying Gemma's release in February 2024.
CodeGemma - Google's Official Code-Focused LLM Release
Google has released CodeGemma, a family of code-specialized large language models, announced via the Hugging Face blog. CodeGemma builds on the Gemma model family and is targeted at code generation and understanding tasks. The release represents Google's continued push into open-weights code LLMs to compete with models like Code Llama and DeepSeek Coder.
PaliGemma – Google's Cutting-Edge Open Vision Language Model
Google released PaliGemma, an open-weights vision-language model built on the PaLI architecture combined with Gemma language components. The model is hosted and documented on Hugging Face, making it accessible for research and fine-tuning. PaliGemma targets multimodal tasks including image captioning, visual question answering, and object detection.
Welcome EmbeddingGemma, Google's new efficient embedding model
Google has released EmbeddingGemma, a new embedding model announced via the Hugging Face blog. The model appears to be positioned as an efficient option for generating text embeddings, likely derived from or related to the Gemma model family. Details on architecture, benchmarks, and use cases are expected in the full post.
Introducing Gemma 3
Google DeepMind has released Gemma 3, described as the most capable model runnable on a single GPU or TPU. The announcement comes from DeepMind's official blog, indicating a new generation of the open-weights Gemma model family. Specific capability details, parameter counts, and benchmark results are not included in the provided body text.
DolphinGemma: Google DeepMind LLM for Decoding Dolphin Communication
Google DeepMind has developed DolphinGemma, a large language model designed to help scientists analyze and decode dolphin communication patterns. The model is being applied to the scientific challenge of understanding cetacean vocalizations. This represents a novel application of LLM-based sequence modeling to non-human animal communication research.
Announcing Gemma 3n Preview: Powerful, Efficient, Mobile-First AI
Google DeepMind has released a preview of Gemma 3n, an open-weights model optimized for on-device multimodal inference. The model features a 2-in-1 architecture for flexible deployment and adds audio understanding to its multimodal capabilities. It is designed for mobile and edge environments, targeting developers building real-time interactive applications.
DeepMind Launches 27B Parameter Gemma-Based Foundation Model for Single-Cell Analysis
DeepMind has released a new 27 billion parameter foundation model built on the Gemma open-model family, specifically designed for single-cell biological analysis. The model contributed to the discovery of a new potential cancer therapy pathway. This represents a significant application of large language model architecture to computational biology and genomics research.
VaultGemma: The world's most capable differentially private LLM
DeepMind introduces VaultGemma, a large language model trained from scratch using differential privacy (DP), claiming it as the most capable DP-trained model to date. The announcement positions VaultGemma as a significant advance in privacy-preserving AI, combining strong utility with formal privacy guarantees. The blog post is brief and likely precedes a more detailed technical disclosure.
Introducing Gemma 3n: The Developer Guide
Google DeepMind has published a developer-focused guide introducing Gemma 3n, a new model in the Gemma open-weights family. The announcement is directed at the developer community and appears to describe architecture, usage, and integration details for the new release. As a Tier 1 source announcement, this represents a notable addition to Google's open-weights model lineup.
MedGemma: DeepMind releases most capable open models for health AI development
Google DeepMind has announced new multimodal models in the MedGemma collection, described as their most capable open models for health AI development. The release expands the MedGemma family with enhanced multimodal capabilities targeting medical and clinical AI applications. As open models, they are intended to support developers building health AI systems.
T5Gemma: A new collection of encoder-decoder Gemma models
DeepMind has announced T5Gemma, a new collection of encoder-decoder large language models under the Gemma family. The release extends the Gemma model line beyond its existing decoder-only architecture to include encoder-decoder variants, following the T5 paradigm. Further technical details are sparse in the announcement but the models represent a notable architectural expansion of the open Gemma ecosystem.