Top-k recovery on LLM Judge Comparison full-data
100Naive RecallDeepSeek-V4-flash
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DeepSeek-V4-flashn=3, cˆ (verbosity)=+0.37, κˆ (position)=−0.082026.07 | 100 | 100 | 0 | |
| DeepSeek-V4-pron=2, cˆ (verbosity)=+0.28, κˆ (position)=+0.012026.07 | 100 | 100 | 0 | |
| Claude-Sonnet-4.6n=3, cˆ (verbosity)=+0.28, κˆ (position)=+1.422026.07 | 100 | 100 | 0 | |
| GPT-4.1n=3, cˆ (verbosity)=+0.24, κˆ (position)=+1.182026.07 | 100 | 100 | 0 | |
| GPT-5.5n=3, cˆ (verbosity)=+0.16, κˆ (position)=+0.072026.07 | 100 | 100 | 0 | |
| Claude-Opus-4.8n=2, cˆ (verbosity)=+0.16, κˆ (position)=+1.372026.07 | 100 | 100 | 0 | |
| Gemini-2.5-pron=1, cˆ (verbosity)=+0.15, κˆ (position)=−0.312026.07 | 100 | 100 | 0 | |
| GPT-5.1n=3, cˆ (verbosity)=+0.13, κˆ (position)=+0.952026.07 | 100 | 100 | 0 | |
| Claude-Haiku-4.5n=3, cˆ (verbosity)=−0.16, κˆ (position)=+0.052026.07 | 93 | 93 | 0 | |
| Phi-4n=10, cˆ (verbosity)=+0.24, κˆ (position)=+0.322026.07 | 88 | 96 | 8 | |
| DeepSeek-chatn=5, cˆ (verbosity)=+0.59, κˆ (position)=−2.412026.07 | 84 | 100 | 16 | |
| Qwen-2.5 (7B)n=10, cˆ (verbosity)=+0.83, κˆ (position)=+1.092026.07 | 66 | 86 | 20 | |
| Gemini-2.5-flash-liten=3, cˆ (verbosity)=+0.75, κˆ (position)=−1.302026.07 | 60 | 93 | 33 | |
| GPT-4o-minin=5, cˆ (verbosity)=+0.61, κˆ (position)=−0.852026.07 | 56 | 84 | 28 | |
| Llama-3.1-8Bn=10, cˆ (verbosity)=+0.99, κˆ (position)=+0.352026.07 | 50 | 90 | 40 | |
| DeepSeek-Coder-6.7Bn=5, cˆ (verbosity)=−0.51, κˆ (position)=+1.752026.07 | 8 | 12 | 4 |