Question Answering on SpeechTriviaQA
42.2AccuracyAudioKV
Evaluation Results
| Method | Links | |
|---|---|---|
| AudioKVModel=Qwen2.5-Omni-7B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 42.2 | |
| AdaKVModel=Qwen2.5-Omni-7B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 40.8 | |
| PyramidKVModel=Qwen2.5-Omni-7B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 39.8 | |
| SnapKVModel=Qwen2.5-Omni-7B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 39.6 | |
| AudioKVModel=Qwen2.5-Omni-3B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 31.9 | |
| AdaKVModel=Qwen2.5-Omni-3B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 30.6 | |
| PyramidKVModel=Qwen2.5-Omni-3B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 28.3 | |
| SnapKVModel=Qwen2.5-Omni-3B, KV cache budget=40%, max_gen_tokens=64, Judge=Gemini 2.5 Flash2026.04 | 27.4 | |
| APinlin=✓, ldeep=-, FRR=93.56, Pre. GFLOPs=730.28, FLOPs Ratio=93.51, Setting=Conservative (τin=0.90, τdeep=0.80)2026.04 | 21.58 | |
| Vanillalin=-, ldeep=-, FRR=100.0, Pre. GFLOPs=780.94, FLOPs Ratio=100.0, Setting=N/A2026.04 | 21.29 | |
| APdeeplin=-, ldeep=✓, FRR=14.30, Pre. GFLOPs=718.12, FLOPs Ratio=91.96, Setting=Aggressive (τin=0.80, τdeep=0.70)2026.04 | 21.19 | |
| APdeeplin=-, ldeep=✓, FRR=33.29, Pre. GFLOPs=731.96, FLOPs Ratio=93.73, Setting=Conservative (τin=0.90, τdeep=0.80)2026.04 | 21.09 | |
| APinlin=✓, ldeep=-, FRR=78.64, Pre. GFLOPs=612.93, FLOPs Ratio=78.49, Setting=Aggressive (τin=0.80, τdeep=0.70)2026.04 | 20.9 | |
| DAPlin=✓, ldeep=✓, FRR=33.76, Pre. GFLOPs=686.39, FLOPs Ratio=87.89, Setting=Conservative (τin=0.90, τdeep=0.80)2026.04 | 20.9 | |
| DAPlin=✓, ldeep=✓, FRR=14.91, Pre. GFLOPs=566.30, FLOPs Ratio=72.52, Setting=Aggressive (τin=0.80, τdeep=0.70)2026.04 | 20.61 |