Training NLA with Entirely False Explanations
Turn_Trout · x · 2026-07-13
Discussing an experimental setup for NLA: Claude is first asked to generate confabulations where "every sentence must be false," and these fabricated explanations are then used to train the NLA to observe changes in reconstruction accuracy and output behavior.
The core question is: if the model's initial explanations are systematically wrong, can the trained NLA still maintain decent reconstruction accuracy, and will it ultimately continue to fabricate heavily or gradually revert to outputs that look more like "reasonable explanations"?
More from Research
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Finch 1: a geospatial foundation model trained on imagery from 50+ sensors — Scobleizer · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27
- Physicists Let AI Pick Its Own Problem—It Wrote a Publishable Paper in 2 Hours — burny_tech · 2026-10-27
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27