Detecting Edit Failures In Large Language Models: An Improved Specificity Benchmark
This work addresses the issue of ensuring safe and reliable model editing for AI practitioners, but it is incremental as it builds upon existing benchmarks and metrics.
The paper tackled the problem of large unwanted side effects in model editing techniques for large language models, which are not detected by existing specificity benchmarks, and found that these techniques suffer from low specificity when evaluated with their improved benchmark.
Recent model editing techniques promise to mitigate the problem of memorizing false or outdated associations during LLM training. However, we show that these techniques can introduce large unwanted side effects which are not detected by existing specificity benchmarks. We extend the existing CounterFact benchmark to include a dynamic component and dub our benchmark CounterFact+. Additionally, we extend the metrics used for measuring specificity by a principled KL divergence-based metric. We use this improved benchmark to evaluate recent model editing techniques and find that they suffer from low specificity. Our findings highlight the need for improved specificity benchmarks that identify and prevent unwanted side effects.