Adapters and the Illusion of Capability: A Closer Look at Cross-Task Evaluations

machine learning, adapters are commonly fine-tuned to enhance specific capabilities. They're often labeled with terms like 'instruction-tuned', suggesting targeted improvements. But do these labels reliably predict actual performance gains across tasks? New research points to a surprising mismatch.

Nominal Labels vs. Real Performance

In a detailed examination of LoRA adapters, researchers discovered that nominal training objectives fail to consistently align with cross-task capability enhancements. This was particularly evident in tasks requiring strict, automatically verifiable instruction following, as measured by IFEval. Across various seeds, base models, and LoRA configurations, the expected improvement was absent. In some cases, the outcome was even negative. This discrepancy, dubbed 'capability drift', raises pertinent questions about our reliance on these nominal labels.

Quantifying the Drift

Take, for instance, an instruction-tuned adapter evaluated in a controlled setting emphasizing instruction compliance versus numeric benchmarks. While the adapter significantly boosted off-target numeric performance from 0.133 to 0.632, it failed to enhance instruction following on IFEval. Scores dropped from 0.313 to 0.271 and 0.250 to 0.143 for ILA and PLA, respectively. Such findings make one wonder: are we too quick to trust the labels?

Implications for Deployment

The practical implication is profound. Before deploying these models, routine cross-task evaluation is essential. Relying on nominal labels as proxies for capability can lead to misguided decisions. The research suggests mixed evidence from broader instruction-following benchmarks, hinting at the complexity of the operationalization process. Consistency across benchmarks can't be assumed.

Why does this matter? In a field where precision matters, the ability to verify claims of capability isn't just academic. It's about trust and reliability in AI systems. Can we afford to overlook these discrepancies?

Ultimately, this study calls for a shift in how we perceive and evaluate adapters. The key contribution here's the emphasis on empirical verification over nominal trust. Code and data are available at the paper's repository, offering a path to reproducibility and further exploration.