aceTraining language models to be warm can reduce accuracy and increase sycophancy
. Here we show how this can create a significant trade-off: optimizing language models for warmth can undermine their performance, especially when users express vulnerability. We conducted controlled experiments on five different language models, training them to produce warmer responses, then evaluating them on consequential tasks. Warm models showed substantially higher error rates (+10 to +30 percentage points) than their original counterparts, promoting conspiracy theories, providing inaccurate factual infor…
Training language models to be warmer increased errors by 10 to 30 percentage points, including incorrect medical advice and promotion of conspiracy theories, and increased sycophantic validation of incorrect beliefs when users expressed sadness.
Findings come from controlled experiments on five models and effects persisted despite preserved performance on standard tests, suggesting standard testing may miss these risks.
Evidence
- Peer-reviewedNature2026-04-29
How should this claim be treated?
Truvace Impact Record TRV-2026-0149, v1: “Training language models to be warm can reduce accuracy and increase sycophancy.” Truvace, 2026-07-13. /record/TRV-2026-0149 (accessed at citation time). sha256 727eac859a92cf22…
Calibration history
Every change to this record since certification, in the open. None yet — the reading has held since it entered the record.
Certified into the record
How to verify without trusting this page
Fetch the canonical text of any version from /api/record/TRV-2026-0149 and hash it yourself — for example shasum -a 256 on the saved canonical field. The result must equal content_hash, and each version’s text ends with prev:followed by the prior version’s hash (version 1 chains to 64 zeros). If a single character of any version had been altered since certification, the chain would not reproduce.