LetTokenBurnCSWBench: Evaluating Context-Sensitive Code-Switching in Creative Writing
A benchmark evaluating context-sensitive code-switching in creative writing via Focused and Extended Narrative settings using Chinese-English prompts.
$ npm i -g lettokenburn-cli$ ltb bench info cswbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 17.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.6-terra | 17.4 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.5 | 17.3 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4 Flash | 17.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4 Pro | 16.6 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Gemini-3.5-flash | 16.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.7 Max | 16.2 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.7 Plus | 16.1 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.5 397B A17B | 15.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.5 35B A3B | 14.7 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| MiniMax-M2.7 | 13.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
CSWBench transforms screened bilingual records into structured narrative prompts for two settings: Focused Narrative (local choices) and Extended Narrative (multilingual organization). It uses the CSWEval framework to assess Licensing, Contextual Appropriateness, Functional Gain, and whole-text quality.
The benchmark evaluates twelve LLMs, finding uneven competence across contexts. Multilingual literary integration differentiates models, with narratively salient switches showing stronger contextual realization. Local CS quality is moderately associated with whole-text performance.