IC-1374Counting difficulty in constrained generation increases with text level and constraint strictness, with exact sentence-level character counts being the hardest condition for all five models

Shunyu Yao, Howard Chen, Austin W. Hanjie, Runzhe Yang, Karthik R Narasimhan

SourceCOLLIE: Systematic Construction of Constrained Text Generation Tasks

The paper decomposes constraint difficulty by counting level and constraint type. Counting characters within a word is easier than within a sentence for all models. Exact equality constraints (e.g., exactly 82 characters in a sentence) are harder than range constraints (e.g., between 10 and 15 words per sentence), which are harder than simple upper/lower bounds. Adding a second counting level via logical composition (e.g., at least 20 words AND each word under 6 characters) uniformly degrades performance compared to the single-level version. Tasks word02, sent01, sent02, and para03 remain very challenging across all models.

Evidence
correlational
Key metric
Tasks word02, sent01, sent02, para03 remain very challenging across all models; sent03 uniformly trails word01 despite the added constraint being assumed manageable
Caveat
The paper reports these as qualitative trends illustrated in figures rather than giving per-task numeric tables in the main text; specific per-task rates are in figures 6-7.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo, PaLM 2, Vicuna Vicuna-7B
Concepts
Failure mode
Related findings
IC-1373, IC-1375
Extraction
automatic-extraction