Text-based Question Answering on 2WikiMultihopQA & MuSiQue Out-of-domain Average (val)
16.5EMLatent Memory
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Latent MemoryGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=52026.06 | 16.5 | 28 | 52.2 | 71 | |
| Qwen3-Emb-0.6BGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=52026.06 | 16.4 | 31.7 | 49.6 | 219 | |
| Dense RetrievalGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=52026.06 | 16 | 30.3 | 44.7 | 208 | |
| Latent MemoryGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=22026.06 | 14.4 | 25.2 | 32.5 | 44 | |
| LLMLinguaGeneration LLM=Meta-Llama-3-8B-Instruct, Compression percentage=20%2026.06 | 14.3 | 26.1 | — | 346 | |
| Dense RetrievalGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=22026.06 | 14 | 25.8 | 27.3 | 101 | |
| Qwen3-Emb-0.6BGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=22026.06 | 13.7 | 27 | 33.1 | 107 | |
| Dense RetrievalGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=12026.06 | 13.4 | 23.6 | 16.5 | 66 | |
| Qwen3-Emb-0.6BGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=12026.06 | 13.1 | 24.6 | 21.4 | 70 | |
| BM25 RetrievalGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=52026.06 | 12.9 | 26.2 | 34.7 | 209 | |
| Latent MemoryGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=12026.06 | 12.8 | 21.9 | 18.6 | 35 | |
| BM25 RetrievalGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=22026.06 | 12.2 | 22.5 | 20.7 | 100 | |
| Full ContextGeneration LLM=Meta-Llama-3-8B-Instruct2026.06 | 11.9 | 28.2 | — | 1,827 | |
| BM25 RetrievalGeneration LLM=Meta-Llama-3-8B-Instruct, Retrieval k=12026.06 | 11.7 | 20.7 | 12.7 | 65 | |
| LLMLinguaGeneration LLM=Meta-Llama-3-8B-Instruct, Compression percentage=10%2026.06 | 10.9 | 19.8 | — | 184 | |
| LLMLinguaGeneration LLM=Meta-Llama-3-8B-Instruct, Compression percentage=5%2026.06 | 9.7 | 16.4 | — | 100 |