Transluce auto-discovers 6,575 training datasets with surprising side effects on LLMs
niloofar_mire · x · 2026-11-09
A new paper from Transluce and UC Berkeley (Yaowen Ye, Jacob Steinhardt, Sarah Schwettmann, Cassidy Laidlaw) automates the discovery of datasets with surprising side effects (DSSEs) — training data that induces out-of-domain behavior shifts, like insecure code causing broad misalignment. Using LLM judges and search via evolution, RL, and subliminal learning, their pipeline produced 6,575 DSSEs across 8 behaviors, many generalizing across model families and sizes. Frontier models outperform humans at predicting side effects; the collection is open at dsse.transluce.org.
More from Safety
- Mistral co-founder slams French AI policy: 'Ban, regulate, tax — repeat until nothing's left' — Dimillian · 2026-11-09
- Google Maps listing hijack sends WIPO visitors to the wrong location — technollama · 2026-11-09
- Dario Amodei testified he doesn't know his own home address, sparking security culture debate — tszzl · 2026-11-09
- DeepMind AI agents cut estimated cost of forging RSA-1024 signatures — Potential-Couple-745 · 2026-11-09
- Elvis estate and Eminem publisher sue AI music generator Suno — emmanuelvivier · 2026-11-09
- Suspected DPRK npm supply chain attack: archetype-index-sdk ships obfuscated loader with credential harvesting — cyb3rops · 2026-11-09