AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

From Extraction to Deduction: Resolving Functional Misalignment in RAG via a Collaborative Critic-Reasoner Framework

Yufei Chen, Yao Wang, Haibin Zhang, Hualin zhou, Tao Gu

ICML 2026 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation

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 changedWhereMeasuredEffect
add DPO stage after SFT (ablation)Critic-Reasoner (Llama3.1-8B) · IR-QAanswer accuracy: 52.92% -> 70.58%Improved
add SFT stage before DPO (ablation)Critic-Reasoner (Llama3.1-8B) · IR-QAanswer accuracy: 55.58% -> 70.58%Improved
add SFT and DPO together vs either alone (ablation)Critic-Reasoner (Llama3.1-8B) · dCR (CR-QA)answer accuracy: 57.17% -> 80.14%Improved
add reasoning data to SFT mixture (ablation)Critic-Reasoner (Llama3.1-8B) · IR-QAanswer accuracy: 59.90% -> 70.58%Improved
add reasoning data to SFT mixture (ablation)Critic-Reasoner (Llama3.1-8B) · II-MCanswer accuracy: 30.67% -> 74.00%Improved
add [MASK] data to SFT mixture (ablation)Critic-Reasoner (Llama3.1-8B) · 2Wikianswer accuracy: 13.75% -> 67.25%Improved
add [MASK] data to SFT mixture (ablation)Critic-Reasoner (Llama3.1-8B) · IR-NQanswer accuracy: 31.05% -> 63.55%Improved
add [MASK] data to SFT mixture (ablation)Critic-Reasoner (Llama3.1-8B) · IM-QAanswer accuracy: 51.08% -> 67.58%Mixed
use Critic-Reasoner instead of standard RAG promptingLlama3.1-8B · IR-QAanswer accuracy: 38.00% -> 70.58%Improved
use Critic-Reasoner instead of standard RAG promptingLlama3.1-8B · II-MCanswer accuracy: 16.08% -> 74.00%Improved
use Critic-Reasoner instead of standard RAG promptingLlama3.1-8B · CR-QA (correct context)answer accuracy: 78.32% -> 80.14%No clear effect
use Critic-Reasoner on smaller 3B modelQwen2.5-3B · IR-NQanswer accuracy: 64.35% -> 60.85%Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy