SAFE: Stepwise Atomic Feedback for Error correction in Multi-hop Reasoning

Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• SAFE identifies up to 14% of instances in standard multi-hop QA benchmarks as unanswerable during train-time verification.result 90%
• SAFE achieves an average accuracy gain of 8.4 percentage points over standard baselines.result 90%
• SAFE replaces ungrounded Chain-of-Thought with a strictly verifiable sequence of grounded entities.method 85%
• SAFE operates across two phases: train-time verification and inference-time verification.method 85%
• SAFE establishes an atomic error taxonomy and a Knowledge Graph-grounded verification pipeline for train-time verification.method 85%
• SAFE exposes critical flaws of existing benchmarks at train-time by eliminating noisy supervision.result 80%
• SAFE guarantees verifiable trajectories at inference-time through dynamic detection of ungrounded steps.technical 80%
• Multi-hop QA benchmarks frequently reward LLMs for spurious correctness, masking ungrounded or flawed reasoning steps.limitation 75%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card