representational similarity analysis

**Representational similarity analysis** is the **method that compares geometric relationships among activations to evaluate similarity between model representations** - it abstracts away from individual units to compare representational structure at scale. **What Is Representational similarity analysis?** - **Definition**: Builds similarity matrices over stimuli and compares matrix structure across layers or models. - **Input**: Uses activation vectors from selected tokens, prompts, or tasks. - **Comparison**: Similarity can be measured with correlation, cosine, or distance-based metrics. - **Output**: Reveals whether two systems encode relationships among inputs in similar ways. **Why Representational similarity analysis Matters** - **Cross-Model Insight**: Supports architecture and checkpoint comparison without unit matching. - **Layer Mapping**: Shows where representational transformations become task-aligned. - **Interpretability**: Helps identify convergent or divergent encoding strategies. - **Neuroscience Link**: Enables shared analysis framework across biological and artificial systems. - **Limitations**: Similarity does not by itself establish causal equivalence. **How It Is Used in Practice** - **Stimulus Design**: Use balanced prompt sets that isolate target phenomena. - **Metric Sensitivity**: Evaluate robustness across multiple similarity metrics. - **Complementary Tests**: Combine RSA with intervention methods for causal interpretation. Representational similarity analysis is **a geometric framework for comparing internal representations** - representational similarity analysis is most useful when geometric findings are tied to task and causal evidence.