NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 28·4 min read
Where Does Toxicity Live? Mechanistic Localization and Targeted Suppression in Language Models
TL;DR
Researchers identified a gap in LLM evaluation benchmarks. They built a synthetic dataset with 10k adversarial prompts targeting reasoning failures.
✦ Why It Matters
Use this benchmark to audit LLM robustness before deploying in production reasoning pipelines.
Key Takeaways
- Run new adversarial benchmark before production LLM deployments.
- Expect 15-25% accuracy drop versus MMLU on reasoning tasks.
- Open-source models degrade faster — prefer fine-tuned variants.
- Dataset available under CC-BY for custom evaluation pipelines.
Full Summary
Prior LLM benchmarks relied on static multiple-choice datasets that saturate quickly. The team constructed adversarial prompts by combining chain-of-thought traps with factual distractors.
GPT-4 scored 71% on the new benchmark versus 89% on MMLU. Open-source models showed steeper drops, suggesting benchmark overfitting.
The dataset is released under CC-BY for reproducibility.