Self-RAG

Unlike standard RAG which retrieves once per query, Self-RAG uses a retrieval critic to decide at each generation step whether retrieval is necessary. The model predicts a special [Retrieve] token to trigger a search, avoiding unnecessary retrieval for simple factual knowledge already stored in its parameters. This adaptive mechanism reduces latency and prevents irrelevant context from corrupting the output.

Self-RAG introduces reflection tokens—special vocabulary items the model generates to critique its own output and retrieved passages:

• [Relevant] / [Irrelevant]: Evaluates if a retrieved passage is useful
• [Supported] / [Partially Supported] / [No Support]: Checks if output is grounded in evidence
• [Utility: 1-5]: Rates the overall usefulness of the response These tokens are predicted alongside the main text, enabling real-time quality control.

Self-RAG is trained using a critic model that generates synthetic reflection token labels. The process involves:

• A critic (often GPT-4 or similar) evaluates retrieved passages and generated outputs
• The main LM is fine-tuned on data augmented with these reflection tokens
• The model learns to internalize the critic's judgment, enabling self-evaluation at inference time without external tools This creates a single model capable of both generation and quality assessment.

During inference, Self-RAG generates multiple output segments in parallel using beam search over reflection token predictions. For each segment, the model:

• Predicts whether to retrieve
• If yes, retrieves and evaluates relevance
• Generates text and predicts support/utility scores
• Selects the best segment based on a weighted score combining factuality and fluency This ensures the final output maximizes both truthfulness and readability.

Self-RAG demonstrates significant improvements in factual accuracy across benchmarks:

• FactScore on biography generation: outperforms standard RAG by explicitly rejecting irrelevant passages
• Citation precision: higher than retrieval-augmented baselines because the model critiques whether each claim is supported
• Reduced hallucination: the [No Support] token allows the model to acknowledge knowledge gaps rather than fabricating answers The framework is particularly effective for long-form generation where factual drift accumulates.

While both Self-RAG and Corrective RAG (CRAG) improve retrieval quality, they differ architecturally:

• Self-RAG: A single fine-tuned LM that internally decides when to retrieve and self-critiques using reflection tokens
• CRAG: An agentic system with separate retrieval evaluator, web search fallback, and knowledge refinement modules Self-RAG's unified approach reduces system complexity, while CRAG offers more explicit control over each retrieval decision. Both represent the shift toward agentic RAG architectures.