Boundary-aware self-distillation tunes LLM safety refusals to deployment-specific subsets
A new arXiv paper introduces narrow-boundary safety, training models to refuse only harmful subsets of a topic rather than whole topics. Using Qwen3-8B, escalating retries cut untraceable prompts from 19.88% to 0.20%, raised target-domain refusal from 9.47% to 84.75%, and reduced unsafe responses from 26.26% to 0.14%, but increased over-refusal on XSTest from 2.00% to 74.00%.
Why this appeared
- Multiple sources
These labels describe how the story was ranked for the feed — not editorial endorsements.
Sources and evidence
Summary last validated Sep 8, 2026
Reader actions
Report an issue
Use this for an incorrect summary, wrong source, duplicate story, or wrong category. Submissions are private and do not change the story automatically.
Related coverage
- research
NIST Study Maps How Toxic Adulterants in Fentanyl Vary Across US
NIST researchers analyzed fentanyl samples nationwide and found that the toxic substances mixed into the drug vary by region and shift over time. The study aims to help first responders and law enforcement identify dangerous adulterants in local supplies, which the agency says can help save lives.
- research
NVIDIA teaches robots to assemble GB300 tester trays
NVIDIA describes how it taught robots to assemble tester trays for its Grace Blackwell GB300 superchip, which powers AI training and inference. The work required skilled physical labor in factories worldwide. NVIDIA says the project offered lessons in robot learning, mechanical intelligence, and traditional engineering.
- models
Liquid AI releases d1-3B and d1-omni-600M open decision models for edge
Liquid AI released two open decision models: d1-3B and experimental d1-omni-600M. d1-3B scores 48.57 on Decision Index 0.2.1, topping sub-10B models and Decider 35B-A3B, and answers in 16 ms on NVIDIA Jetson AGX Thor. d1-omni-600M handles text with image or audio. Both are built on Liquid Foundation Models.
- models
TII Releases Falcon ASR Speech Recognition Model
TII, the institute behind the Falcon language models, introduced Falcon ASR on Hugging Face. The blog post announces the automatic speech recognition model, but the provided source contains no further details on architecture, training data, languages, or benchmark results.
- research
Ai2's Bolmo byte-level language model technique published in Nature
Ai2 announced that the technique behind Bolmo, its fully open byte-level language models, has been published in Nature. New checkpoints show the approach generalizes beyond the Olmo model family to other model families, according to the Allen Institute for AI.