Google released Gemini 4 Argon on October 13 2026 for developers enterprises and subscribers. The frontier model excels at multi-step coding finance legal cybersecurity and…
Builders and creators can now generate photoreal, anime, and NSFW images on Venice without extra platform filters. Free tier gives 15 prompts daily; Pro unlocks 1,000 images…
New preprint shows that fine-tuning on carefully constructed benign examples can substantially degrade safety alignment in LLMs by inducing a refusal-to-compliance behavioral…
An arXiv preprint from October 19 2026 shows that simply stating a legal role such as "I am a lawyer" can increase refusal rates on benign legal-domain prompts by 2-20x in four…
An arXiv preprint dated September 28 2026 introduces PrivacySkills a framework that measures how system instructions and skill level labels affect whether LLM agents select…
New arXiv research on eight instruction-tuned LLMs shows that refusal can sometimes increase safety risks in suicidal ideation contexts. Clinician-guided protocols reduced both…
arXiv preprint dated 2026-09-28 examines compensatory feature injection on Qwen3.5 models using sparse autoencoders. It finds sycophancy cuts do not consistently strengthen…
arXiv preprint from September 2026 proposes VaccineBooster, a hybrid alignment defense combining embedding perturbation and gradient attenuation. On Llama-2-7B it shows an…
A September 2026 arXiv preprint introduces FDCU, a subspace projection method that uses Fisher Information and the Principle of Minimal Functional Intervention to achieve…
An October 2026 arXiv preprint by Saman Rahbar shows that similarity scores used to claim brain alignment or cross-lingual transfer can remain high even after the underlying…
A preprint submitted October 5 2026 explores if large language models exhibit anosognosia-like failure to recognize quantization-induced impairments in their computational…