Research map / Large language models / Retrieval and knowledge
The First Drop of Ink: Nonlinear Impact of Misleading Information in Long-Context Reasoning
Muhan Gao, Zih-Ching Chen, Kuan-Hao Huang
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 |
|---|---|---|---|
| add the first 10% of hard distractors to a fixed-length 128K context (ablation) | Llama-3.1-8B-Instruct · nq_easy 128K | accuracy (%): 87.0 -> 72.5 | Hurt |
| raise hard distractor share from 10% to 100% in 128K context (ablation) | Llama-3.1-8B-Instruct · nq_easy 128K | accuracy (%): 72.5 -> 62.0 | Hurt |
| add the first 10% of hard distractors to a fixed-length 128K context (ablation) | Qwen2.5-7B-Instruct · nq_easy 128K | drop ratio: drop ratio 0.58 vs 0.1 for linear | Hurt |
| lower softmax temperature at inference (tau < 1) to sharpen attention on the target (ablation) | Llama-3.1-8B-Instruct · nq_easy | accuracy | Hurt |
| filter hard distractors instead of weaker (random) distractors, from 131K down to 47K tokens (ablation) | Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct · multi-document QA, four datasets | accuracy | No clear effect |
| filter hard distractors until their share is near zero (3%) at 27K tokens, versus filtering random (ablation) | Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct · multi-document QA, four datasets | accuracy | Improved |
| reduce context length from 131K to 27K while keeping hard ratio fixed at 20%, 50% or 80% (ablation) | Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct · multi-document QA, four datasets | accuracy | Improved |
| vary the fixed hard distractor ratio among 20%, 50%, 80% during proportional context reduction (ablation) | Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct · multi-document QA, four datasets | accuracy | No clear effect |
| use LLM judge (gpt-4o-mini) instead of string containment matching | evaluation of QA accuracy · HotpotQA | false negatives: 36 of 200 false negatives under string matching vs 17 in total over 100 checked samples per dataset (4.25%) | Improved |
Closest papers by meaning
- Retrieval Head Mechanistically Explains Long-Context Factuality (ICLR 2025 · similarity 0.88)
- Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · similarity 0.87)
- Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads (NeurIPS 2026 · similarity 0.86)
- Adaptive Distraction: Probing LLM Contextual Robustness with Automated Tree Search (NeurIPS 2025 · similarity 0.86)
- Rethinking External Slow-Thinking: From Snowball Errors to Probability of Correct Reasoning (ICML 2025 · similarity 0.85)
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs (ICLR 2026 · similarity 0.85)
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.
