Reproduce “CODE: A Contradiction-Based Deliberation Extension Framework for Overthinking Attacks on Retrieval-Augmented Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaolei Zhang, Xiaojun Jia, Liquan Chen, Songze Li
Researchers at Southeast University and Nanyang Technological University developed CODE, a multi-agent framework that induces 'overthinking' in retrieval-augmented generation (RAG) systems by poisoning their knowledge bases. The framework creates documents with structured contradictions and diverse styles, increasing reasoning token consumption by up to 43.451 times while maintaining task accuracy and demonstrating a stealthy, cost-inflating vulnerability in commercial reasoning models.
Key insights
- RAG systems integrating reasoning models inherit the "overthinking" susceptibility of standalone reasoning models, even when the external knowledge base is subtly poisoned.
- Systematically crafted cross-layer contradictions within retrieved documents are the primary driver for inflating reasoning chains, causing models to engage in extended self-correction and conflict reconciliation.
- Current defense mechanisms, including prompt-based efficiency constraints and trust-aware retrieval filtering, are insufficient to fully mitigate these contradiction-induced overthinking attacks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.