Research map / Large language models / Retrieval and knowledge
Resisting Contextual Interference in RAG via Parametric-Knowledge Reinforcement
Chenyu Lin, Yilin Wen, Du Su, Hexiang Tan, Fei Sun, Muhan Chen, Chenfu Bao, Zhonghou Lv
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 J_RPK objective (robust-PK re-evaluation under context) (ablation) | Knowledgeable-R1 · ConFiQA MR (TIFE: PK right, context wrong) | accuracy: 46.50% -> 75.80% (MC TIFE, with J_PK removed in both) | Improved |
| add the J_PK objective (ablation) | Knowledgeable-R1 · ConFiQA PC-MR & NC-MR avg | average accuracy: 64.42% -> 65.40% | Improved |
| add the global advantage for CK (ablation) | Knowledgeable-R1 · ConFiQA PC-MR & NC-MR avg | average accuracy: 60.95% -> 65.40% | Improved |
| add the local advantage for CK (ablation) | Knowledgeable-R1 · ConFiQA PC-MR & NC-MR avg | average accuracy: 64.69% -> 65.40% | Improved |
| use adaptive beta instead of a fixed beta (ablation) | Knowledgeable-R1 · ConFiQA PC-MR & NC-MR avg | average accuracy: 56.95% -> 65.40% | Improved |
| use adaptive beta instead of best fixed beta (0.01) (ablation) | Knowledgeable-R1 · ConFiQA MC-avg | MC-avg accuracy: 56.03% -> 56.43% | Improved |
| replace GRPO w/ RAG with Knowledgeable-R1 training | Qwen2.5-7B-Instruct · ConFiQA NC-MR | EM accuracy: 26.94% -> 43.94% | Improved |
| replace GRPO w/ RAG with Knowledgeable-R1 training | Qwen2.5-7B-Instruct · ConFiQA PC-MR (correct context) | EM accuracy: 77.56% -> 75.08% | Hurt |
| replace GRPO w/ RAG with Knowledgeable-R1 training | Qwen2.5-7B-Instruct · ConFiQA SC (self-conflicting) | EM accuracy: 75.33% -> 76.33% | No clear effect |
| raise CK objective weight to 1:2:1 (ablation) | Knowledgeable-R1 · ConFiQA-MC | S1 vs S2 accuracy: S1 75.51% -> 78.04%; S2 37.34% -> 32.73% | Mixed |
| use local+global CK advantage instead of local only (ablation) | Knowledgeable-R1 · ConFiQA avg score | average score: 60.95 -> 65.40 | Improved |
| train with only 1% bad context (ablation) | Knowledgeable-R1 · ConFiQA-MR S2 | accuracy: GRPO 14.31% -> 17.34% | Improved |
Closest papers by meaning
- ParamMute: Suppressing Knowledge-Critical FFNs for Faithful Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.91)
- KnowPO: Knowledge-Aware Preference Optimization for Controllable Knowledge Selection in Retrieval-Augmented Language Models (AAAI 2025 · similarity 0.90)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.89)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.88)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts (ICLR 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.
