Research map / Large language models / Language, safety and interpretability
Harnessing the Unseen: The Hidden Influence of Intrinsic Knowledge in Long-Context Language Models
Yu Fu, Haz Sameen Shahgir, Hui Liu, Xianfeng Tang, Qi He, Yue Dong
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 |
|---|---|---|---|
| make the question require parametric recall before retrieval (Hybrid NIAH) instead of standard NIAH (ablation) | Llama-3.1-8B-Instruct · NIAH vs Hybrid NIAH (gen length 32, 3 random facts) | average accuracy: 98.83 -> 67.47 | Hurt |
| add 0 to 3 random fact distractors to Hybrid NIAH (ablation) | Llama-3.1-70B-Instruct · Hybrid NIAH (gen length 32) | average accuracy: 95.73 -> 71.81 | Hurt |
| add 0 to 3 random fact distractors to standard NIAH (ablation) | Qwen2.5-7B-Instruct · NIAH (gen length 32) | average accuracy: 99.78 -> 99.92 | No clear effect |
| scale Llama-3.1 from 8B to 70B | Llama-3.1 · Hybrid NIAH (gen length 32, 3 random facts) | average accuracy: 67.47 -> 71.81 | No clear effect |
| scale Qwen2.5 from 7B to 72B | Qwen2.5 · Hybrid NIAH (gen length 32, 3 random facts) | average accuracy: 73.14 -> 97.77 | Improved |
| increase generation length from 32 to 64 (ablation) | Qwen2.5-72B-Instruct · Hybrid NIAH (0 random facts) | average accuracy: 64.01 -> 94.16 | Improved |
| limit Qwen2.5-7B-Instruct-1M max input to 128k (ablation) | Qwen2.5-7B-Instruct-1M · Hybrid NIAH (gen length 32, 3 random facts) | average accuracy: 42.91 -> 79.31 | Improved |
| replace RoPE with STRING position remapping | Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.2 · HotpotQA-Context subset | accuracy | Improved |
| replace RoPE with STRING position remapping | Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.2 · HotpotQA-Parametric subset and I-WhoQA-Irrelevant | accuracy | Hurt |
| context that conflicts with parametric knowledge instead of aligning, at longer context lengths | Llama-3.1-8B-Instruct · I-WhoQA | accuracy gap: gap up to 10 points | Hurt |
Closest papers by meaning
- Training with “Paraphrasing the Original Text” Teaches LLM to Better Retrieve in Long-Context Tasks (AAAI 2025 · similarity 0.90)
- Taming Knowledge Conflicts in Language Models (ICML 2025 · similarity 0.89)
- A Controllable Examination for Long-Context Language Models (NeurIPS 2025 · similarity 0.89)
- Retrieval Head Mechanistically Explains Long-Context Factuality (ICLR 2025 · similarity 0.89)
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 2024 · similarity 0.89)
- ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.89)
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.
