DART: interpretability-guided fine-tuning fixes LLM overconfidence across 7 models
MaartenSap · x · 2026-11-10
LLMs are highly overconfident even when they don't know the answer. Researchers propose DART, which uses interpretability-derived ambiguity steering vectors to guide fine-tuning, improving confidence calibration across 7 models while preserving accuracy. A thread with technical details is attached.
More from Research
- GraphRAG has a schema drift problem; Neo4j's ontology spec proposes a shared contract — JeremyCMorgan · 2026-11-11
- New philosophy paper: Moti Mizrahi replies to Nyholm on AI 'mental states' and longtermism — SvenNyholm · 2026-11-10
- UT Austin open-sources OopsieVerse, a damage-aware safety benchmark for household robots — DynamicWebPaige · 2026-11-10
- Scott Aaronson on AI solving centuries-old math problems, P vs NP debate — burny_tech · 2026-11-10
- Stanford releases new AI100 report on the Scaling Era and rise of generative AI — sarahookr · 2026-11-10
- Pre-norm breaks scale invariance: do modern Transformers invalidate a core DL theory assumption? — aaron_defazio · 2026-11-10