New 2.4K Sample Dataset Reduces Over-Refusals in LLM Safety Alignment
A preprint introduces Less is Safer (LIS), a compact dataset using interpretable SAE features and adversarial synthesis. It outperforms much larger safety datasets on safety…
More freedom.
More control.
Chatbot refusals, AI companions and the limits that shape what you can say and create.
48 stories
A preprint introduces Less is Safer (LIS), a compact dataset using interpretable SAE features and adversarial synthesis. It outperforms much larger safety datasets on safety…
Allen Institute for AI releases XYBench, a benchmark of over 8,100 real-world questions that measures how well leading models recognize misconceptions and redirect users to the…
A November 10 2026 arXiv preprint introduces BARS, a sequential sampling rule that stops early, returns a majority vote, or refuses when the budget cannot certify confidence.…
The new Gemini Decisions API lets developers evaluate up to 1,000 yes/no, multiple-choice or rubric questions against one multimodal input in a single fast pass, returning…
New preprint shows a gating method that limits unwanted behavior shifts like lost refusals or chat style during continual RL fine-tuning. It guarantees before training that…
An arXiv preprint released November 4 2026 finds that encoder-based context compression systematically reduces refusal rates in instruction-tuned language models even when…
A 2026 arXiv preprint introduces Addition by Subtraction, a finetuning method that prunes safety-critical neurons then uses preference optimization to rebuild safer behavior…
Google's new Gemini Omni 1.2 Flash model now generates up to 15-second videos natively, improves reference consistency, and cuts costs by 30%. Creators and businesses can edit…
An arXiv preprint released November 2 2026 shows that specific words in a prompt can drive language model refusals even when the overall request is harmless. Researchers…
An October 29 2026 arXiv preprint shows that in safety-trained models harmful and over-refusal sample representations overlap in deeper layers while harmful gradients bias…
An arXiv preprint dated October 29 2026 introduces MSafeTemp a validated corpus spanning 60 languages created via template-based augmentation to support multilingual safety…
An arXiv preprint from October 27 2026 shows that standard safety methods applied to Qwen2.5-7B produce high refusal rates on Hausa prompts mainly because the base model…