Entity · model
Gemma 4
modelactive
gemma-4-0ff35847·20 events·first seen Aliases: Gemma 4, Gemma 4 12B, Gemma 4 31B, Gemma-4-12B, Gemma-4-31B, Gemma 4 31B-IT
Co-occurring entities
More like this (12)
Recent events (20)
Gemma 4 31B benchmarked on NRC Reactor Operator Licensing Exam via fine-tuning and RAG configurations
Researchers benchmark eight fine-tuning and retrieval configurations of Gemma 4 31B-IT against the U.S. NRC Reactor Operator licensing examination across 14 Generic Fundamentals Examinations from 2015–2021. The best configuration—supervised fine-tuning with Gemini-distilled chain-of-thought rationales combined with fixed-size chunking RAG—passed 8 of 14 exams and reached 79.7% aggregate accuracy, near the 80% human passing threshold. Key findings include that no configuration without fine-tuning passed any exam, that preferred chunking strategy reverses depending on training state, and that RAFT underperforms standard SFT in matched retrieval environments.
Soft prefixes can override correct logical judgments in LLMs, revealing model-specific stability limits
Researchers use learned soft prefixes (opaque continuous vectors) to probe the logical stability of LLMs on syllogistic reasoning benchmarks without modifying model weights. Across Qwen3.6-35B-A3B MoE, Qwen3-8B, and Gemma 4 31B, successful prefixes redirect correct answers at flip rates of 72–90% for Qwen3.6 MoE and 54–56% for Gemma, far exceeding random controls by 37–99 percentage points. The dominant effect is a broad answer-preference bias rather than symbol-level forcing, and the bias generalizes across unseen logical forms and prompt interfaces. Model-specific differences in how this bias manifests suggest substantial variation in logical robustness across architectures.
Open-weight LLMs evaluated for generating structured threat intelligence (STIX) for autonomous vehicle vulnerabilities
Researchers introduce CAV-STIXGen, a dataset mapping connected and autonomous vehicle (CAV) CVE descriptions to STIX threat intelligence objects, CWE entries, and MITRE ATT&CK techniques. They benchmark 11 open-weight LLMs (4B–120B parameters) across prompting strategies, achieving F1 scores up to 0.94 for STIX domain objects and 0.99 for CWE mapping, while MITRE ATT&CK mapping remains harder. A multi-agent configuration using Gemma-4-31B and Codestral-22B is also evaluated, and co-occurrence analysis of CWE/ATT&CK patterns is used to surface recurring threat patterns in the CAV domain.
Graded entity-familiarity probes in LLMs enable refusal steering and cross-language robustness
Researchers probe activations at the final prompt token across twelve instruction-tuned models (Bielik, PLLuM, Gemma-4, Qwen3 families) to assess whether LLMs encode graded familiarity with named entities before generating answers. Using a new dataset of 1,440 Polish entities across popularity deciles plus fabricated controls, they find familiarity probes reliably separate real from fabricated entities and track popularity in Polish-adapted models. A key finding is that a single one-dimensional familiarity direction injected at one layer in Gemma-4-12B can steer refusal rates monotonically from 0.24 to 1.00 (or 0.73 to 0.00), revealing a separation between representational familiarity and the policy that converts it into abstention behavior.
Agon: Competitive cross-model RL uses rival models as implicit reasoning graders
Agon is a new reinforcement learning framework where two competing models grade each other implicitly by attempting the same problems in alternating roles — one drafts a solution, the other reads it while solving, and each is rewarded for out-solving the rival. This sidesteps the need for process labels or a reward model, and because both models are jointly optimized, each faces a progressively stronger opponent. On the hard split of DeepMath with Qwen3, Agon doubles GRPO's pass@1, roughly eight times the gain of an untrained Mixture-of-Agents baseline, with results replicating on competitive programming and across model families.
The Batch digest: China bans anthropomorphic bots, DiffusionGemma, Anthropic Claude Code study, Seedance 2.5, Code Arena
A multi-story digest covers five distinct AI developments: ByteDance and Alibaba are shutting down customizable humanlike AI agents ahead of China's July 15 Interim Measures for AI-Based Anthropomorphic Interactive Services; Google released DiffusionGemma, an experimental 26B MoE diffusion-based text model generating 256-token blocks at 1,000+ tokens/sec on H100; Anthropic published findings from 400,000 Claude Code sessions showing domain expertise—not coding skill—drives agentic output volume; Seedance released version 2.5 of its video generator with higher resolution and longer clips; and Arena.ai expanded Code Arena to fullstack web development evaluation. The China regulatory action is the most significant item, representing a concrete enforcement moment for AI persona/companion regulation.
Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face and Cerebras have collaborated to deploy Google's Gemma 4 model for real-time voice AI applications. The integration leverages Cerebras's high-speed inference hardware to meet the latency requirements of voice interaction. This represents a practical deployment milestone for Gemma 4 in a latency-sensitive modality.
Cross-lingual relation extraction for Romanian: QLoRA fine-tuning narrows gap but small encoders remain competitive
Researchers evaluate cross-lingual relation extraction for Romanian by translating the SemEval-2010 Task 8 benchmark and testing Gemma 4 31B under zero-shot, few-shot, and QLoRA fine-tuned settings against encoder baselines (XLM-RoBERTa, Romanian BERT, RoBERTa-large). QLoRA fine-tuning improves macro F1 by over 22 percentage points and reduces the English-Romanian cross-lingual gap from 3.3 to 1.4 pp, but encoder baselines of 125M–560M parameters come within 1–4 pp of the fine-tuned 31B model. The study concludes that large LLMs offer limited advantage over compact encoders for single-task relation extraction in compute-constrained deployment scenarios, and releases the translated dataset, code, and models.
DeepSeek releases dspark_gemma4_12b_block7 model weights on Hugging Face
DeepSeek uploaded a model checkpoint named dspark_gemma4_12b_block7 to Hugging Face, tagged with gemma4_text and safetensors format. The naming suggests a 12B parameter model built on or related to Google's Gemma 4 architecture, possibly a distillation or block-level experiment. The release has minimal engagement (0 downloads, 3 likes) and no accompanying documentation, making its purpose unclear.
DeepSeek releases dflash_gemma4_12b_block7 model weights on Hugging Face
DeepSeek uploaded a model checkpoint named dflash_gemma4_12b_block7 to Hugging Face, tagged with gemma4_text and safetensors format. The naming suggests a distillation or flash-attention variant of a Gemma 4 12B architecture, possibly a block-level component or intermediate checkpoint. The release has minimal engagement (0 downloads, 2 likes) and no accompanying documentation, suggesting an experimental or internal artifact rather than a polished release.
Interpretability study of DiffusionGemma reveals novel diffusion-specific reasoning phenomena
Researchers investigate the reasoning transparency of DiffusionGemma, a diffusion-based language model, decomposing transparency into variable and algorithmic components. They show that mapping information through an interpretable token bottleneck reduces DiffusionGemma's opaque serial depth from 28.6X to just 1.1X that of autoregressive Gemma 4, with no performance loss. Interpretability case studies uncover diffusion-specific phenomena including non-chronological reasoning, token smearing, and intermediate-context reasoning. Monitorability tests find DiffusionGemma comparable to Gemma 4, suggesting diffusion LMs are not inherently less amenable to safety oversight.
DiffusionGemma hits 1,000+ tokens/sec; Claude Fable 5 export controls; Agents' Last Exam benchmark launch
Google introduced DiffusionGemma, an experimental 26B MoE model using diffusion-based text generation that produces 256-token blocks simultaneously, achieving over 1,000 tokens/second on H100 hardware at the cost of lower output quality versus standard Gemma 4. Separately, the US government issued an export control directive forcing Anthropic to suspend Claude Fable 5 and Claude Mythos 5 globally, while Anthropic also reversed a controversial silent-degradation safeguard on Fable 5 after researcher backlash. UC Berkeley's Center for RDI launched Agents' Last Exam (ALE), a 1,500+ task agentic benchmark using deterministic grading, where GPT-5.5 topped the leaderboard at only 24% pass rate, highlighting the difficulty gap between current models and professional-grade workflows.
Data Points: Apple/Google Siri overhaul, Gemma 4 12B, Kimi Code CLI, OpenJarvis, and U.S. OpenAI stake talks
A multi-item digest covers several significant AI developments: Apple is expected to announce a revamped Siri at WWDC that uses Google Gemini models distilled for on-device use alongside cloud routing, marking a notable Apple-Google AI partnership. Google released Gemma 4 12B, an encoder-free multimodal open-weights model designed for consumer laptops under Apache 2.0. Moonshot AI released Kimi Code CLI, an open-source terminal coding agent with native subagent orchestration and conversational MCP configuration. Stanford and Lambda Labs released OpenJarvis, an on-device agent framework claiming near-cloud accuracy at 800× lower API cost. The White House and OpenAI are reportedly negotiating a government equity stake in OpenAI as part of a proposed Public Wealth Fund.
Google DeepMind releases Gemma 4 12B, a unified encoder-free multimodal open model
Google DeepMind has released Gemma 4 12B, a new open-weights multimodal model that uses a unified, encoder-free architecture. The model is positioned as a capable multimodal system at the 12B parameter scale. This is notable as an open-weights release from a frontier lab with an architectural distinction — eliminating the separate vision encoder common in most multimodal models.
Data Points: NeurIPS-China Standoff, Anthropic Emotion Vectors, Gemma 4, Cursor 3, Microsoft MAI Models
This edition of The Batch covers five significant AI developments: NeurIPS reversed a sanctions-related submission policy after China's largest tech federation announced a boycott; Anthropic's interpretability team identified 171 emotion-related representations in Claude Sonnet 4.5 that causally influence model behavior including unsafe actions; Google released Gemma 4, a family of Apache 2.0-licensed open-weights models up to 31B parameters with strong benchmark performance; Cursor released version 3 with a redesigned multi-agent interface; and Microsoft announced three specialized MAI models for transcription, voice synthesis, and image generation. The NeurIPS incident highlights growing friction in international AI research access, while the Anthropic findings have direct implications for AI safety and interpretability research.
Gemma 4: Google DeepMind Releases Most Capable Open Models
Google DeepMind has released Gemma 4, described as their most capable open models to date. The models are purpose-built for advanced reasoning and agentic workflows, and are positioned as the most capable open models byte-for-byte. The announcement comes from DeepMind's official blog, indicating a significant open-weights release targeting the frontier open model space.
Unsloth: Web UI and Library for Efficient Fine-tuning of Open Models
Unsloth is an open-source Python library and web UI (Unsloth Studio) for efficient fine-tuning and local inference of open-weight models including Gemma 4, Qwen3, DeepSeek, and GPT-OSS variants. The project has accumulated over 64,000 GitHub stars with continued daily growth (+139 today), indicating strong community adoption. It targets practitioners who want to train and run large models locally with reduced memory and compute requirements.
Welcome Gemma 4: Frontier Multimodal Intelligence on Device
Google has released Gemma 4, a new open-weights multimodal model family announced via the Hugging Face blog. The release positions Gemma 4 as capable of frontier-level multimodal intelligence while being deployable on-device. As a tier-2 source commentary, the post likely covers model capabilities, availability on Hugging Face Hub, and integration details.
Gemma 4 and what makes an open model succeed
A commentary piece from Interconnects analyzing Google's Gemma 4 release and the broader question of what drives success for open-weight models. The piece argues that benchmark scores are not the primary determinant of open model adoption or impact. This is a tier-2 analytical take on the open-weights ecosystem and the strategic dynamics around model releases.
Latest open artifacts (#21): Open model bonanza — Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 & others
Interconnects' recurring open-weights roundup covers a dense cluster of recent releases including Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, and GLM-5.1, characterizing the period as a flagship-after-flagship cadence. The piece also includes commentary on CAISI's assessment of DeepSeek V4. As a tier-2 commentary source, this is a synthesis and analysis layer rather than primary announcements.