Long-context Reasoning on LongBench
73.8ScoreGHG-TDA
Evaluation Results
| Method | Links | |
|---|---|---|
| GHG-TDABase Model=GPT-4o2026.02 | 73.8 | |
| GHG-TDABase Model=Claude 3.5 Sonnet2026.02 | 73.5 | |
| AoTBase Model=GPT-4o2026.02 | 72.6 | |
| AoTBase Model=Claude 3.5 Sonnet2026.02 | 72.2 | |
| GHS-TDABackbone=DeepSeek-V32026.02 | 69.9 | |
| GHS-TDABackbone=GPT-4o-mini2026.02 | 69.5 | |
| AoTBackbone=DeepSeek-V3, Reasoning Paradigm=atomic reasoning2026.02 | 68.7 | |
| GHS-TDABackbone=Qwen-Turbo2026.02 | 68.6 | |
| AoTBackbone=GPT-4o-mini, Reasoning Paradigm=atomic reasoning2026.02 | 68.5 | |
| AoTBackbone=Qwen-Turbo, Reasoning Paradigm=atomic reasoning2026.02 | 68.1 | |
| GoTBase Model=Claude 3.5 Sonnet2026.02 | 67 | |
| GoTBase Model=GPT-4o2026.02 | 66.8 | |
| ToTBase Model=Claude 3.5 Sonnet2026.02 | 66.4 | |
| ToTBase Model=GPT-4o2026.02 | 66 | |
| GHS-TDABackbone=Llama 3-8B2026.02 | 64.64 | |
| AoTBackbone=Llama 3-8B2026.02 | 63.71 | |
| GoTBackbone=DeepSeek-V3, Reasoning Paradigm=graph-based2026.02 | 63.4 | |
| ToTBackbone=DeepSeek-V3, Reasoning Paradigm=tree-based2026.02 | 63.2 | |
| GoTBackbone=GPT-4o-mini, Reasoning Paradigm=graph-based2026.02 | 63.1 | |
| GHS-TDABackbone=Qwen 2-14B2026.02 | 62.8 | |
| ToTBackbone=GPT-4o-mini, Reasoning Paradigm=tree-based2026.02 | 62.8 | |
| GoTBackbone=Qwen-Turbo, Reasoning Paradigm=graph-based2026.02 | 62.8 | |
| ToTBackbone=Qwen-Turbo, Reasoning Paradigm=tree-based2026.02 | 62.4 | |
| CoTBase Model=GPT-4o2026.02 | 62 | |
| CoTBase Model=Claude 3.5 Sonnet2026.02 | 61.7 | |
| AFlowBackbone=DeepSeek-V3, Reasoning Paradigm=search-based2026.02 | 61.4 | |
| AoTBackbone=Qwen 2-14B2026.02 | 61.2 | |
| AFlowBackbone=GPT-4o-mini, Reasoning Paradigm=search-based2026.02 | 61 | |
| AFlowBackbone=Qwen-Turbo, Reasoning Paradigm=search-based2026.02 | 60.4 | |
| ToTBackbone=Qwen 2-14B2026.02 | 59.3 | |
| FoTBackbone=DeepSeek-V3, Reasoning Paradigm=forest-based, n=82026.02 | 59.3 | |
| FoTBackbone=GPT-4o-mini, Reasoning Paradigm=forest-based, n=82026.02 | 59.1 | |
| FoTBackbone=Qwen-Turbo, Reasoning Paradigm=forest-based, n=82026.02 | 59 | |
| CoT-SCBackbone=DeepSeek-V3, Reasoning Paradigm=chain-based, n=52026.02 | 58.9 | |
| GoTBackbone=Llama 3-8B2026.02 | 58.68 | |
| CoT-SCBackbone=GPT-4o-mini, Reasoning Paradigm=chain-based, n=52026.02 | 58.6 | |
| ToTBackbone=Llama 3-8B2026.02 | 58.4 | |
| CoT-SCBackbone=Qwen-Turbo, Reasoning Paradigm=chain-based, n=52026.02 | 58.4 | |
| Self-RefineBackbone=DeepSeek-V3, Reasoning Paradigm=chain-based2026.02 | 58.4 | |
| Self-RefineBackbone=GPT-4o-mini, Reasoning Paradigm=chain-based2026.02 | 58.2 | |
| GoTBackbone=Qwen 2-14B2026.02 | 58.1 | |
| Self-RefineBackbone=Qwen-Turbo, Reasoning Paradigm=chain-based2026.02 | 58 | |
| CoTBackbone=DeepSeek-V3, Reasoning Paradigm=chain-based2026.02 | 57.7 | |
| CoTBackbone=GPT-4o-mini, Reasoning Paradigm=chain-based2026.02 | 57.6 | |
| CoTBackbone=Qwen-Turbo, Reasoning Paradigm=chain-based2026.02 | 57.3 | |
| Full KVBackbone=Qwen2.5-14B, KV Cache Budget=Full2026.06 | 55.47 | |
| SEKVBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 54.71 | |
| CoTBackbone=Llama 3-8B2026.02 | 53.57 | |
| Analogical PromptingBackbone=DeepSeek-V3, Reasoning Paradigm=chain-based2026.02 | 53.4 | |
| Analogical PromptingBackbone=GPT-4o-mini, Reasoning Paradigm=chain-based2026.02 | 52.9 | |
| CoTBackbone=Qwen 2-14B2026.02 | 52.7 | |
| Analogical PromptingBackbone=Qwen-Turbo, Reasoning Paradigm=chain-based2026.02 | 52.7 | |
| SentenceKVBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 52.31 | |
| SemantiCacheBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 51.88 | |
| ChunkKVBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 51.42 | |
| Full KVBackbone=Llama-3.1-8B, KV Cache Budget=Full2026.06 | 51.18 | |
| SEKVBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 51.02 | |
| Full KVBackbone=Llama-3-8B, KV Cache Budget=Full2026.06 | 50.34 | |
| SEKVBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 50.18 | |
| PyramidKVBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 50.17 | |
| SentenceKVBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 49.93 | |
| SemantiCacheBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 49.61 | |
| SnapKVBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 49.34 | |
| ChunkKVBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 49.14 | |
| SentenceKVBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 49.12 | |
| SemantiCacheBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 48.73 | |
| ChunkKVBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 48.31 | |
| PyramidKVBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 48.02 | |
| H2OBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 47.18 | |
| PyramidKVBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 47.13 | |
| SnapKVBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 47.11 | |
| SnapKVBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 46.23 | |
| H2OBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 44.52 | |
| H2OBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 43.84 | |
| Full KVBackbone=Llama-3.2-3B, KV Cache Budget=Full2026.06 | 43.21 | |
| SEKVBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 42.63 | |
| StreamingLLMBackbone=Qwen2.5-14B, KV Cache Budget=10%2026.06 | 42.31 | |
| Full KVBackbone=Mistral-7B, KV Cache Budget=Full2026.06 | 41.83 | |
| SentenceKVBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 41.44 | |
| SemantiCacheBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 41.17 | |
| ChunkKVBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 40.83 | |
| PyramidKVBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 39.74 | |
| SEKVBackbone=Llama-3.2-3B, KV Cache Budget=10%2026.06 | 39.47 | |
| StreamingLLMBackbone=Llama-3.1-8B, KV Cache Budget=10%2026.06 | 39.17 | |
| SnapKVBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 38.91 | |
| StreamingLLMBackbone=Llama-3-8B, KV Cache Budget=10%2026.06 | 38.41 | |
| SentenceKVBackbone=Llama-3.2-3B, KV Cache Budget=10%2026.06 | 37.83 | |
| H2OBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 37.63 | |
| SemantiCacheBackbone=Llama-3.2-3B, KV Cache Budget=10%2026.06 | 37.14 | |
| ChunkKVBackbone=Llama-3.2-3B, KV Cache Budget=10%2026.06 | 36.82 | |
| PyramidKVBackbone=Llama-3.2-3B, KV Cache Budget=10%2026.06 | 35.44 | |
| BLASSTModel=Qwen3-8B, Sparsity=~50%, Deployment Phase=Prefill Phase2025.12 | 35.1 | |
| StreamingLLMBackbone=Mistral-7B, KV Cache Budget=10%2026.06 | 34.82 | |
| BLASSTModel=Qwen3-8B, Sparsity=~75%, Deployment Phase=Prefill Phase2025.12 | 34.4 | |
| SnapKVBackbone=Llama-3.2-3B, KV Cache Budget=10%2026.06 | 34.18 | |
| Dense AttentionModel=Qwen3-8B, Sparsity=Dense, Deployment Phase=Prefill Phase2025.12 | 33.6 | |
| Dense AttentionModel=Qwen3-8B, Sparsity=Dense, Deployment Phase=Prefill + Decode Phase2025.12 | 33.6 | |
| BLASSTModel=Qwen3-8B, Sparsity=~50%, Deployment Phase=Prefill + Decode Phase2025.12 | 33.3 | |
| BLASSTModel=Qwen3-8B, Sparsity=~75%, Deployment Phase=Prefill + Decode Phase2025.12 | 33.1 | |
| BLASSTModel=Llama-3.1-8B, Sparsity=~50%, Deployment Phase=Prefill + Decode Phase2025.12 | 32.4 |