ReaLM-Retrieve：推理模型的自适应检索框架

ReaLM-Retrieve is an adaptive retrieval framework designed for large reasoning models. It addresses the fundamental mismatch between traditional RAG systems (context provided upfront) and reasoning models (needing dynamic evidence during multi-step reasoning). Key benefits include a 10.1% absolute performance boost on multiple benchmarks and a 47% reduction in retrieval calls compared to baselines.

Large reasoning models like DeepSeek-R1 and OpenAI o1 excel at multi-step reasoning with long chains. However, traditional RAG systems provide context before inference, while these models require dynamic evidence injection at specific reasoning steps. This timing mismatch limits their full potential.

ReaLM-Retrieve’s three key innovations:

1. Step-level Uncertainty Detector: Identifies knowledge gaps at individual reasoning steps (not token/sentence level) to pinpoint retrieval needs.
2. Retrieval Intervention Strategy: Intelligent decision mechanism to trigger retrieval only when beneficial (vs fixed-interval methods).
3. Efficiency-Optimized Integration: Reduces retrieval overhead by 3.2x, enabling real-time use.

Evaluated on MuSiQue, HotpotQA, and 2WikiMultiHopQA:

• F1 Score: 10.1% absolute improvement over standard RAG (9.0%-11.8% range).
• Retrieval Efficiency: 47% fewer calls than fixed-interval methods like IRCoT.
• MuSiQue Standout: 71.2% F1 with 1.8 retrievals per question.
• Evidence Quality: Recall@5 of 81.3% for supporting evidence, outperforming baselines in precision and MRR. All improvements are statistically significant (p<0.01).

Key technical aspects:

1. Uncertainty Modeling: Analyzes model internal states to detect uncertain reasoning steps (knowledge gaps).
2. Dynamic Retrieval Decision: Balances step uncertainty, evidence relevance, and reasoning benefits to decide when to retrieve.
3. Efficient Integration: Optimized evidence encoding, fast relevance scoring, and tight coupling with model generation reduce overhead.

For enterprises:

1. Cost Savings: 47% fewer retrieval calls cut costs (especially with commercial APIs).
2. Lower Latency: Fewer retrievals mean faster response times.
3. Higher Quality: 10.1% accuracy gain adds value in critical scenarios. This framework shifts RAG focus from 'what to retrieve' to 'when to retrieve'.

Current limitations and future work:

• Multilingual Support: Needs validation on non-English datasets.
• Domain Adaptation: Further research for vertical domains (healthcare/legal).
• Model Compatibility: Testing across different reasoning model architectures. ReaLM-Retrieve is a key step toward efficient, reliable AI reasoning systems.