Research map / Large language models / Retrieval and knowledge
Attributing Response to Context: A Jensen–Shannon Divergence Driven Mechanistic Study of Context Attribution in Retrieval-Augmented Generation
Ruizhe Li, Chen Chen, Yuchen Hu, Yanjun Gao, Xi Wang, Emine Yilmaz
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 |
|---|---|---|---|
| attribute with ARC-JSD instead of ContextCite, ALTI-Logit and MIRAGE baselines | RAG LLMs (Qwen2, Gemma2) · MuSiQue | top-1 attribution accuracy: about 10.7% average improvement | Improved |
| use ARC-JSD instead of gradient or surrogate-based attribution | RAG attribution · compute-accuracy trade-off | compute: up to 3-fold speedup | Improved |
| gate top-5 JSD-located attention heads and MLPs with semantic gain mask | Qwen2-7B-IT · Hotpot QA (200 failed samples) | hallucination rate: 13.4% -> 8.2% | Improved |
| gate top-5 JSD-located components | Qwen2-7B-IT · Hotpot QA (200 failed samples) | Pass@1 Factual F1: 76.1 -> 75.9 | No clear effect |
| gate random 5 attention and MLP components instead of JSD-located ones (ablation) | Qwen2-7B-IT · Hotpot QA (200 failed samples) | Pass@1 Factual F1: 76.1 -> 69.4 | Hurt |
| mask top-10 JSD-located attention heads rather than random heads (ablation) | ARC-JSD · all RAG models, all datasets | average JSD: 1.53 -> 2.23 | Improved |
| use JSD instead of KL, Wasserstein, TV, MMD as divergence (ablation) | ARC-JSD · MuSiQue | compute-accuracy trade-off | Improved |
| ablate attention heads or MLP layers with higher JSD score (ablation) | Qwen2-1.5B-IT · TyDi QA | attribution accuracy drop | Hurt |
| compare ContextCite with 32 forward calls vs ARC-JSD | ContextCite-32 · MuSiQue | attribution accuracy | Hurt |
| apply ARC-JSD to LLaMA-3.1-8B-IT and Qwen3-Next-80B-A3B-IT | ARC-JSD · MuSiQue | compute-accuracy trade-off | Improved |
Closest papers by meaning
- Influence Guided Context Selection for Effective Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.89)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.88)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.88)
- Sufficient Context: A New Lens on Retrieval Augmented Generation Systems (ICLR 2025 · similarity 0.88)
- A Theory for Token-Level Harmonization in Retrieval-Augmented Generation (ICLR 2025 · 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.
