Research map / Large language models / Language, safety and interpretability
Stacked from One: Multi-Scale Self-Injection for Context Window Extension
Wei Han, Pan Zhou, Shuicheng YAN
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 |
|---|---|---|---|
| use the default right-branch policy instead of random branch selection (ablation) | SharedLLM · Arxiv-32K perplexity | perplexity: 2.61 -> 2.46 | Improved |
| use default policy instead of fixed-left selection (ablation) | SharedLLM · Arxiv-32K perplexity | perplexity: 2.49 -> 2.46 | Improved |
| use default policy instead of query-aware selection for language modeling (ablation) | SharedLLM · Arxiv-32K perplexity | perplexity: 2.48 -> 2.46 | Improved |
| use default policy instead of Lambda-shape policy (ablation) | SharedLLM · Arxiv-32K perplexity | perplexity: 2.52 -> 2.46 | Improved |
| use default policy instead of random selection on multi-document QA (ablation) | SharedLLM · MD-QA (LLaMA-2 SFT) | F1: 28.85 -> 30.93 | Improved |
| add query-aware information gathering, noise in node splitting, chunk-level positional indices (ablation) | SharedLLM · Design-choice ablation | performance | Improved |
| inject context into the bottom M layers rather than other layer choices (ablation) | SharedLLM · Injection layer ablation | performance | Improved |
| train with Books3 (data) vs without (ablation) | CEPE · Arxiv 4K perplexity | perplexity: 3.03 -> 2.86 | Improved |
| replace CEPE with SharedLLM | CEPE · Arxiv 4K perplexity | perplexity: 3.03 -> 2.99 | Improved |
| replace Activation Beacon with SharedLLM | Activation Beacon · MD-QA (LLaMA-2 SFT) | F1: 28.44 -> 30.93 | Improved |
| replace Activation Beacon with SharedLLM | Activation Beacon · InfBench Ret.N (LLaMA-2 SFT) | score: 80.58 -> 82.79 | Improved |
| add query-aware retrieval (ablation) | SharedLLM · Inference time (8K tokens) | time (s) / memory (GB): 11.81 -> 13.18 s; 22.39 -> 25.44 GB | Hurt |
Closest papers by meaning
- Retrieval meets Long Context Large Language Models (ICLR 2024 · similarity 0.89)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.88)
- Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.87)
- A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.87)
- AdmTree: Compressing Lengthy Context with Adaptive Semantic Trees (NeurIPS 2025 · similarity 0.87)
- Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks? (ICLR 2025 · similarity 0.87)
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.
