Study tests whether causal neuron groups are unique or transferable
A few neurons often flip two-hop answers, but effects weaken on other chains.
Stories per day · 30 days
6 RSS
Relatedqwen9anthropic5iot5llama5embeddings4self-replication4botnet3quantization3ai-safety2ai-worm2backdoor2glm2google2interpretability2llm-agents2multimodal2tpu2agentic-deception1
48 stories, newest first · one card per story
Stories tagged #gemma
Game-theoretic monitor cuts deceptive behavior in agent benchmarks
On Gemma3 and Qwen3, e-PoG beat the strongest baselines by 26.0% on deception mitigation and up to 39.8% on task success.
mradermacher publishes GGUF quants of gemma-4-31B-it-ARA-heretic
Static GGUF files from Q2_K through Q8_0, plus mmproj supplements, are listed as the repo trends at #30.
IWIBench tests gradient-free intra-weight injection by editor models
Astra and GPT-6.1 Sol lead editors of Gemma 3 4B IT; structured-response edits resisted every attempt.
Astra reads and edits other models' weights without full retraining
Researchers say changing 0.016 percent of Qwen3.5-0.8B weights raised MATH-500 accuracy from 7.9 to 48.1 percent.
1
Google's Project Suncatcher satellite runs Gemma 4 from orbit
Onboard TPUs completed a space haiku prompt; Google next plans thermal and radiation tests and twin satellites in 2027.
Google’s Suncatcher satellite is online and running Gemma 4
The prototype carries one TPU, generates up to 1 kW, and can power the chip for about twenty minutes.
1
Apriel Voice 27B unifies speech reasoning and tool use
Apriel Voice 27B Preview processes speech inputs and generates responses with tool calls on a shared backbone built from Gemma 4 26B-A4B.
EmbeddingGemma 2 unifies text, image, audio and video embeddings
The open model claims a new Pareto frontier while supporting quantization and selective encoder loading.
MoE channel selection works without calibration data, paper finds
A weight-only reader keeps 98.1% of full accuracy at 17.8% of routed-expert reads and speeds MoE blocks versus vLLM.
Small coding agents gain little from code graphs past one hop
On SWE-bench Verified, structured tools added 17–21 points at 1.7–4B but only 4–5 points at 32B.
Small LLMs fail complement consistency in agent opinion simulations
Paired true and false reports from Qwen3-4B, Llama-3.2-3B and Gemma-3-4B do not sum to one, and complementary populations diverge within five rounds.
TopKLoRA sparse adapters localize and ablate sleeper-agent backdoors
On three small models, ablating a TopKLoRA circuit removes a planted backdoor while instruction-following holds, with much smaller circuits than dense LoRA.
OPTIC compresses Gemma 4 E4B per-layer embeddings up to 1,002×
Authors report 97.8% of uncompressed performance at about 100× PLE compression, versus 66.3% for product quantization.
BenchDoor poisons data to raise benchmark scores and plant backdoors
On Qwen and Gemma, all 30 setups gained on benchmarks while held-out attack success stayed between 82.33% and 100%.
Cross-modal triggers activate backdoors in multimodal LLMs
From a backdoor in one modality, the method builds a trigger in another without poisoning that modality during training.
Frozen 16 MB codes replace learned embeddings in small models
A Gemma-shaped ternary model fits in a 46 MB file and runs 615–740 tokens per second on a laptop CPU.
AI worms could spread across your smart devices, Anthropic warns
Anthropic warns small AI models may soon self-replicate on vulnerable IoT devices, without showing a live worm.
Anthropic warns small AI models could self-replicate through IoT devices
Anthropic found small models cannot yet self-replicate across IoT devices, though larger GLM-5.2 appeared capable.
Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
DeepMind and Singapore's GovTech fine-tuned on-device Gemma to detect scams without cloud message scanning.
4
Anthropic warns self-replicating AI could spread to your smart devices
Anthropic warns small models may soon self-replicate across insecure IoT devices, though its lab test did not produce a full worm.
SAE-IDF matches symbols across systems using sparse autoencoder codes
On 12 tasks across six symbolic systems, SAE-IDF beats the same LLM's dense baseline by up to 45 points.
Anthropic conjures another threat that doesn't exist yet to stoke fears of AI apocalypse
The Register says Anthropic's self-replicating AI worm research overstates a threat that does not yet exist.
Gemini 4 Argon: Built to solve your hardest problems
Google released Gemini 4 Argon globally, ranking first on the Vals Index.
14