Abstract
Large language models must satisfy hard orthographic constraints during controlled text generation, yet systematic cross-family evaluation remains limited. We evaluate 39 configurations spanning three model families (Qwen3, Claude Haiku 4.5, GPT-5-mini) on 58 word puzzles requiring character-level constraint satisfaction. Cross-family differences produce substantially larger performance gaps (2.0-2.2x, F1 = 0.761 vs. 0.343) than parameter scaling within families (83% gain from 4B to 32B scaling), and a partial-correlation analysis rules out tokenizer design as a confound for within-family scaling. Thinking budget sensitivity proves heterogeneous: high-capacity models show strong returns (+0.102 to +0.136 F1), while mid-sized variants saturate or degrade, showing inconsistent compute benefits. Using difficulty ratings from 10,000 human solvers per puzzle, we establish modest but consistent calibration (\r{ho} = 0.28-0.42) across all families, yet identify systematic failures on common words with unusual orthography ("data", "loll", "acai": 83-91% human success, 94-98% model miss rate). These failures point to over-reliance on distributional plausibility that penalizes orthographically atypical but constraint-valid patterns.
Abstract (translated)
URL
https://arxiv.org/abs/2511.21086
