New paper: adversarial examples aren't bugs, they're superposition, with four lines of evidence
livgorton · x · 2026-10-22
Liv Gorton and Owen Lewis's paper "Adversarial Examples Are Not Bugs, They Are Superposition" (arXiv:2508.17456) argues that superposition from mechanistic interpretability is the primary cause of adversarial examples, supported by four lines of evidence: theory, toy-model interventions on superposition controlling robustness, the reverse direction via adversarial training, and replication in ResNet18. The authors contend superposition has remarkable predictive power over adversarial phenomena, offering a unified interpretability-based framework for robustness research.
Related event: Anthropic researcher rebuts "Neural Networks Are Already Clean" claim(11 posts)→
More from Research
- 130K-sample RAG claim verification dataset lands on Hugging Face — MaziyarPanahi · 2026-10-22
- Tencent Hunyuan's WorkForge builds 16.7K verifiable real-world environments for work agents — TencentHunyuan · 2026-10-22
- Gauss vs the machine: what the 1-to-100 sum trick says about ML and computation — tobias_rees · 2026-10-22
- Radiolab's 'Math Vs Machine': OpenAI claims a Millennium Prize Problem solve — stevenstrogatz · 2026-10-22
- UK/Nigeria driver-coaching NLG study wins Best Evaluation award at INLG 2026 — EhudReiter · 2026-10-22
- 31,000-question study finds LLMs across 24 languages are secretly obsessed with Japan — dioscuri · 2026-10-22