The Decodability Criterion for Model Selection
The paper introduces a novel metric—the decodability criterion—to address the inefficiency of majority voting in LLM ensembles. While majority voting is a standard technique for improving reliability, it is computationally expensive as it requires multiple full-model forward passes. The authors propose that by analyzing the internal hidden states of a model, one can predict which output is more likely to be correct without needing to generate multiple full responses.
Hidden-State Selection vs. Majority Voting
The core argument is that the internal representation (hidden state) of an LLM contains latent information about the model's confidence and the correctness of its output. By evaluating the 'decodability' of these states—essentially measuring how easily the model's internal representation can be mapped to a correct token prediction—builders can select the most accurate output from a set of candidates. This approach often outperforms majority voting because it leverages the model's internal 'certainty' rather than relying on the frequency of output tokens, which can be misleading in cases of systematic bias or hallucination.
Practical Implications for Inference
This research suggests a shift in how we handle multi-agent or ensemble-based inference. Instead of running multiple full generations and performing a simple vote, developers can potentially:
- Generate a smaller set of candidate outputs.
- Use the decodability criterion to score the hidden states associated with those outputs.
- Select the output with the highest decodability score.
This method reduces the overhead associated with redundant generation while maintaining, or in many cases exceeding, the accuracy gains typically associated with majority voting.