Context Compression Fidelity on 10-Q MD&A
11Flip RateNo compression
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| No compressionDescription=Identity floor2026.06 | 11 | 0.078 | — | 100 | |
| ACCType=Multi-LLM, Model Framework=GPT-5.4-MINI + DEEPSEEK-V4-FLASH2026.06 | 20.3 | 0.102 | 9.07 | 4.3 | |
| ContextualizationBase Model=GPT-5.4-MINI2026.06 | 21.3 | 0.099 | 8.06 | 4.2 | |
| IntegratorType=Multi-LLM, Model Framework=GPT-5.4-MINI + DEEPSEEK-V4-FLASH2026.06 | 24.7 | 0.128 | 9.76 | 4.3 | |
| Naive PromptBase Model=GPT-5.4-MINI2026.06 | 33 | 0.172 | 6.69 | 4.1 | |
| ChunkingBase Model=GPT-5.4-MINI2026.06 | 37 | 0.155 | 6.03 | 4 | |
| LongLLMLinguaType=Token Pruning, Base Model=FINANCE-LLAMA3-8B2026.06 | 50 | 0.226 | — | 4 | |
| LLMLinguaType=Token Pruning, Base Model=FINANCE-LLAMA3-8B2026.06 | 53 | 0.231 | — | 4 |