Researchers find Instruction tuning sharpens Number-Choice heuristics in language models
A new study on arXiv examines how instruction tuning affects the way language models pick numbers. The authors measured the probability of models selecting integers from 1 to N without sampling error, testing 18 open-weight model lines across five ranges up to N = 100. They found that instruction tuning reduces uniform distribution in all 15 models with matched base and tuned checkpoints,…
Key points
- Instruction tuning reduces uniform number selection in 15 of 18 tested models, reinforcing heuristics like endpoint and round-number avoidance
- Tuned models nearly eliminate round numbers and endpoints (1 or N) as choices, while base models do not show this behavior
- Preference for odd numbers appears only at N = 100, with no shared favorite integer across tuned models
The study notes that base models do not avoid round numbers once endpoints are excluded, while tuned models nearly eliminate them. Tuned models also rarely pick endpoints, and a preference for odd numbers appears only at N = 100. The findings align partially with human number-choice behavior but differ in key ways, such as the absence of midpoint avoidance. The study concludes that tuning does not create a universal favorite number across models, as no integer exceeds 42% preference in tuned conditions.
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 ontology-guided AI system for explainable medical diagnosis · 1 src
- Qwen models struggle with outdated facts in retrieved evidence · 1 src
- Google DeepMind releases Genie 4 world model with native audio and Gemini 3.8 Flash guidance · 10 src
- SNNs harder to steal: study · 1 src
- MARQ infuses SE knowledge into LLMs for RE · 1 src
Comments
via GitHub Discussions