Research map / Large language models / Retrieval and knowledge
Mitigating Error Accumulation in Knowledge Editing for Multi-Hop Question Answering
Jiaxin Guo, Hao Sun, Wenhao Zhang, Xuanbo Fan, Yan Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace tree/chain-based editing with tree-structured evidence expansion (ToE) | ToE · MQuAKE-CF reliability (Llama-3-8B) | accuracy: 48.3 (RAE-prune) -> 53.1 | Improved |
| use NLI observer instead of Llama-3-8B LLM as conflict detector (ablation) | ToE · MQuAKE-CF reliability | accuracy: 38.1 -> 53.1 | Improved |
| use NLI observer instead of GPT-4o-mini as conflict detector (ablation) | ToE · MQuAKE-CF reliability | accuracy: 51.8 -> 53.1 | Improved |
| use NLI observer instead of Qwen2.5-32B LLM observer (ablation) | ToE · MQuAKE-CF reliability | accuracy: 49.9 -> 53.1 | Improved |
| expand top-b scoring nodes rather than bottom-b (ablation) | ToE · MQuAKE-CF reliability | accuracy: 23.4 -> 53.1 | Improved |
| expand top-b scoring nodes rather than bottom-b (ablation) | ToE · DUNE New Info | accuracy: 81.0 -> 93.3 | Improved |
| increase beam width and branching from b=1,n=1 to b=2,n=3 (ablation) | ToE · MQuAKE-CF reliability | accuracy: 42.0 -> 53.1 | Improved |
| increase search further from b=2,n=3 to b=3,n=5 (ablation) | ToE · MQuAKE-CF reliability | accuracy: 53.1 -> 52.7 (AVG 76.2 -> 77.0) | No clear effect |
| apply ToE to GPT-4o-mini and GPT-4o versus RAE | ToE · MQuAKE-CF reliability (black-box LLMs) | accuracy: GPT-4o-mini 55.3 (RAE) vs 54.1; GPT-4o 58.7 (RAE) vs 60.8 | Mixed |
| use multi-GPU parallel branch exploration | ToE · MQuAKE-CF latency (300 examples) | latency | Improved |
| tree expansion versus chain-based reasoning on single GPU | ToE · MQuAKE-CF latency | latency: 2-3x higher latency | Hurt |
| ToE versus EditCoT on DUNE Science | ToE · DUNE Science (Llama-3-8B) | accuracy: 85.0 (EditCoT) -> 82.2 | Hurt |
Closest papers by meaning
- Knowledge Editing with Dynamic Knowledge Graphs for Multi-Hop Question Answering (AAAI 2025 · similarity 0.91)
- Uncovering Overfitting in Large Language Model Editing (ICLR 2025 · similarity 0.90)
- Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing (ICML 2025 · similarity 0.90)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.89)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.88)
- Neighboring Perturbations of Knowledge Editing on Large Language Models (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
