Understanding Emergent Misalignment via Feature Superposition Geometry
Samenvatting
Understanding Emergent Misalignment via Feature Superposition Geometry. arXiv:2605.00842v1 Announce Type: new Abstract: Emergent misalignment, where fine-tuning on narrow, non-harmful tasks induces harmful behaviors, poses a key challenge for AI safety in LLMs. Despite growing empirical evidence, its underlying mechanism remains unclear. To uncover the reason behind this phenomenon, we propose a geometric account based on the geometry of feature superposition. Because features are encoded in overlapping representations, fine-tuning that amplifies a target feature also unintentionally strengthens nearby harmful features in accordance with their similarity. We give a simple gradient-level derivation of this effect and empirically test it in multiple LLMs (Gemma-2 2B/9B/27B, LLaMA-3.1 8B, GPT-OSS 20B). Using sparse autoencoders (SAEs), we identify features tied to misalignment-inducing data and to harmful behaviors, and show that they are geometrically closer to each other than features derived from non-inducing data. This trend generalizes across domains (e.g., health, career, legal advice). Finally, we show that a geometry-aware approach, filtering training samples closest to toxic features, reduces misalignment by 34.5%, substantially outperforming random removal and achieving comparable or slightly lower misalignment than LLM-as-a-judge-based filtering. Our study links emergent misalignment to feature superposition, providing a basis for understanding and mitigating this phenomenon.
Waarom dit ertoe doet
Deze technologische ontwikkeling kan de manier waarop AI in de zorg wordt ingezet fundamenteel veranderen.
Context (AI-duiding)
Klik op “Toon context” om AI-duiding op te halen.
Nieuwsbrief
Wekelijks dit soort signalen in je inbox
De nieuwsbrief bundelt nieuwe signalen, relevante verschuivingen en korte duiding zodat je minder afhankelijk bent van incidentele sitebezoeken.
Scores
De mate waarin dit signaal de Nederlandse gezondheidszorg kan beïnvloeden (1 = minimaal, 5 = transformatief).
Hoe snel actie of aandacht nodig is (1 = kan wachten, 5 = onmiddellijke aandacht vereist).
De mate van onzekerheid over de uitkomst of timing (1 = zeer voorspelbaar, 5 = zeer onzeker).
Tags
Bronnen
- Understanding Emergent Misalignment via Feature Superposition GeometryarXiv - Artificial Intelligence (cs.AI) —
Pipeline versie: 0.2.0 | Gegenereerd door: pipeline