Bergson: open-source library unifies data attribution methods to study LLM generalization and misalignment
zetalyrae · x · 2026-10-05
A new paper from luciaquirke's team tackles a key question: how do LLMs generalize, and why do they sometimes become misaligned? Their angle: the training data.
The team released Bergson, an open-source library that brings together data attribution methods across a wide range of compute budgets, letting researchers trace model behavior back to training samples and study generalization and alignment issues with whatever resources they have.
More from Research
- ICML 2026 proceedings go live on PMLR as Volume 306 from Seoul conference — lawrennd · 2026-10-05
- Neuroscientist Anil Seth's TED talk: current AI won't become conscious — we're seeing faces in clouds — anilkseth · 2026-10-05
- CMU ships cua-speedrun: standardized benchmark finally measures computer-use agent speed and cost — rsalakhu · 2026-10-05
- TernaryQuench: open-source ternary quantization trainer for Qwen3 with MLX export — casper_hansen_ · 2026-10-05
- Xaira unveils AI drug discovery platform: 10x medicines goal, early results on hard GPCR target — BoWang87 · 2026-10-05
- The Agent Simulates Predictor problem: should one-boxing survive a weaker Omega? — jessi_cata · 2026-10-05