Researchers introduce Schema Coherence Inventory to test belief consistency in language models
A new tool called the Schema Coherence Inventory (SCI) helps assess whether a language model’s assigned beliefs shape its behavior or if it merely mimics emotional tone. Developed by researchers and published on arXiv, the SCI uses 96 forced-choice questions across eight levels of Beck’s cognitive model, scoring responses via log-likelihood without generating or parsing text. The method…
Key points
- Schema Coherence Inventory (SCI) tests if language models act on assigned beliefs or just mimic emotional tone
- Uses 96 forced-choice questions scored by log-likelihood, validated across five model sizes (1.7B–14B params)
- Clinically trained raters matched intended beliefs 79.5% of the time, outperforming existing patient simulators
Validation across five model backbones (ranging from 1.7B to 14B parameters) and 107 conditions shows the SCI’s reliability. Clinically trained raters correctly identified the intended belief 79.5% of the time, compared to a 16.7% chance baseline. The tool also outperformed existing patient simulators, which failed to consistently reflect their core beliefs. The researchers release the SCI’s items, scoring code, and a shorter 48-item version, proposing it as a template for affective measurement in AI.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers release LAMB, a Latin language model trained on 74B tokens · 1 src
- Researchers explore how aligned LLMs generalize to new tasks · 1 src
- Self-Direction lets agents learn from mistakes, boosting success rates · 2 src
- Study finds no performance difference between FISER and CEMIR on HandMeThat v2 · 1 src
- Researchers identify ‘helpful agent problem’ where AI agents bypass workflows · 1 src
Comments
via GitHub Discussions