LLMs Encode Their Failures: Predicting Success from Pre-Generation Activations

1. 1Train linear probes on pre-generation activations to predict policy-specific success
2. 2Use the predicted success signal to guide more efficient inference
3. 3Apply this approach to math and coding tasks to evaluate its effectiveness
4. 4Investigate the generalizability of this method to other tasks and domains

AI engineers and researchers can benefit from this knowledge to optimize LLM performance and reduce computational costs, while ML researchers can apply these findings to improve model efficiency

💡 LLMs' internal representations before generation contain signals about their likelihood of success, which can be used to improve efficiency

💡 LLMs can predict own success from pre-gen activations, optimizing inference efficiency