A new study published on arXiv investigates how instruction tuning affects the number-choice heuristics in language models. Researchers measured the probability distributions of integers assigned by 18 open-weight models, comparing base and instruction-tuned versions. The findings indicate that instruction tuning makes these distributions less uniform, specifically by installing a round-number avoidance heuristic and sharpening an endpoint avoidance heuristic. While base models showed less consistent behavior, tuned models almost never returned endpoints and significantly avoided multiples of ten. AI
IMPACT Provides insight into how instruction tuning shapes specific cognitive-like behaviors in LLMs, potentially influencing their reliability in tasks requiring numerical reasoning.
RANK_REASON Academic paper detailing research findings on language model behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
