Closed-source LLMs reach CSE Level 6 while most open-source baselines cluster at CSE Levels 3–4 on CSEBench¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q3
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q3 i? All closed-source models evaluated reach at least CSE-6 under both zero-shot and few-shot prompting. → Shangchao Min 2026 q3 i? Most open-source baselines cluster at CSE-3 to CSE-4, with Qwen2.5-7B-Instruct reaching CSE-5 in both prompting conditions. → Shangchao Min 2026 (2)

Evidence¶

Shangchao Min 2026¶

Shangchao Min, Shaonan Wang, Xinyu Gao, Hui Wang, Zhiling Jin, and 2 more. (2026). Standardized Assessment of LLM English Proficiency. https://doi.org/10.21203/rs.3.rs-8820245/v1

q3 · i? · design · r2

Standard-setting study mapping model scores to CSE levels via Bookmark cut scores on two linked diagnostic forms. The results section reports that "all closed-source models in our study reach at least CSE-6" under both prompting conditions (GPT-4o, Claude4, Deepseek-R1, Gemini2.5-Pro).

"The standard-setting results (see Table 1) indicate that all closed-source models in our study reach at least CSE-6 under both zero-shot and few-shot prompting."

Shangchao Min 2026 (2)¶

Shangchao Min, Shaonan Wang, Xinyu Gao, Hui Wang, Zhiling Jin, and 2 more. (2026). Standardized Assessment of LLM English Proficiency. https://doi.org/10.21203/rs.3.rs-8820245/v1

q3 · i? · design · r2

Same standard-setting mapping applied to open-source models (RWKV-v6-7b, LLaMA-3-8B, Mistral-7B, Qwen2.5-7B). The article reports most "open-source baselines cluster at CSE-3 to CSE-4" with Qwen2.5-7B-Instruct reaching CSE-5. No effect size is printed; level assignments are the reported outcome.

"In contrast, most open-source baselines cluster at CSE-3 to CSE-4, with Qwen2.5-7B-Instruct as a notable exception, reaching CSE-5 in both prompting conditions."

Discussion¶

Related Claims¶

• Targeted enhancements (retrieval-augmented inference, multimodal training) substantially raise open-source models' phonology mastery to near-ceiling levels — related
• Few-shot prompting does not uniformly improve LLM performance and can degrade subskill mastery estimates, most notably for phonology — related
• Open-source LLMs show persistent phonology deficits despite adequate mastery of other subskills — related
• Cohort members predominantly used off-the-shelf multimodal LLMs, mostly OpenAI's GPT, and nearly all combined closed and open source models — related