Verb-ICL: Rethinking In-Context Learning for Structured Prediction
Verb-ICL uses token-level example selection plus generated error feedback to improve low-resource structured prediction with LLMs.
The paper targets tasks where outputs depend on fine-grained annotation patterns, not just sentence-level similarity. Its framework selects examples for token-level coverage, then adds feedback meant to encode task-specific labeling rules. The authors evaluate it on six information extraction and semantic parsing datasets and report consistent gains over selective annotation baselines. They also say the feedback helps beyond one-off corrections, acting more like reusable task guidance. ArXiv · AI/CL/LG's note
The paper targets tasks where outputs depend on fine-grained annotation patterns, not just sentence-level similarity. Its framework selects examples for token-level coverage, then adds feedback meant to encode task-specific labeling rules. The authors evaluate it on six information extraction and semantic parsing datasets and report consistent gains over selective annotation baselines. They also say the feedback helps beyond one-off corrections, acting more like reusable task guidance. ArXiv · AI/CL/LG's note
score 4