SAFE: An LLM-as-Verifier Framework for Evidence-Grounded Multi-Hop Reasoning

Researchers propose SAFE, an LLM-as-verifier framework that improves multi-hop question answering by validating reasoning steps against evidence during generation rather than only checking final answers. The approach uses Knowledge Graph triples to decompose reasoning into verifiable units and achieves 8.8 percentage point accuracy improvements across three benchmarks.

SAFE addresses a fundamental problem in LLM evaluation: models often produce correct answers through flawed reasoning paths, creating a gap between apparent performance and actual reasoning quality. This framework shifts verification from post-hoc answer judgment to real-time step-by-step validation, fundamentally changing how we assess LLM reasoning capabilities. By grounding intermediate steps in provided passages and Knowledge Graph representations, SAFE forces models to maintain logical coherence throughout their reasoning chain rather than allowing shortcuts to correct conclusions.

The technical innovation lies in decomposing complex reasoning into atomic, evidence-grounded units that can be independently verified. At training time, the framework filters benchmark data to identify reliable supervision signals, ensuring training data reflects genuine reasoning chains. During inference, an external verifier continuously checks each generated step and provides corrective feedback before errors compound—a meaningful departure from current approaches that tolerate intermediate mistakes if final answers prove correct.

This development has implications for AI reliability and interpretability in domains requiring traceable reasoning paths. Financial analysis, legal research, and scientific reasoning all depend on valid intermediate steps, not just correct answers. The 8.8 percentage point accuracy improvement across multiple benchmarks suggests the approach generalizes meaningfully. For AI developers building production systems, SAFE demonstrates that verification mechanisms operating during generation—rather than after—can significantly improve both accuracy and trustworthiness of reasoning-dependent tasks.

• →SAFE validates reasoning steps against evidence during generation, not after, improving accuracy by 8.8pp on average across multi-hop QA benchmarks
• →The framework decomposes reasoning into Knowledge Graph triples to create verifiable, atomic units of logic
• →Real-time verification prevents error propagation by catching invalid reasoning before it influences subsequent steps
• →Evidence-grounded verification shifts evaluation from rewarding spurious correctness to ensuring valid intermediate reasoning
• →The approach enables construction of higher-quality training data by filtering supervision under logical constraints