Context Compression Fidelity on Earnings Call
50.8FlipLLMLingua
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| LLMLinguaType=Token Pruning, Base Model=FINANCE-LLAMA3-8B2026.06 | 50.8 | 0.247 | — | 4 | |
| LongLLMLinguaType=Token Pruning, Base Model=FINANCE-LLAMA3-8B2026.06 | 39.1 | 0.196 | — | 4 | |
| ChunkingBase Model=GPT-5.4-MINI2026.06 | 31.6 | 0.147 | 4.26 | 4 | |
| Naive PromptBase Model=GPT-5.4-MINI2026.06 | 23.9 | 0.137 | 3.83 | 4.6 | |
| IntegratorType=Multi-LLM, Model Framework=GPT-5.4-MINI + DEEPSEEK-V4-FLASH2026.06 | 22.9 | 0.128 | 6.52 | 4.7 | |
| ContextualizationBase Model=GPT-5.4-MINI2026.06 | 22.2 | 0.116 | 3.82 | 4.2 | |
| ACCType=Multi-LLM, Model Framework=GPT-5.4-MINI + DEEPSEEK-V4-FLASH2026.06 | 18.5 | 0.112 | 7.52 | 4.7 | |
| No compressionDescription=Identity floor2026.06 | 8.8 | 0.058 | — | 100 |