SPEX and ProxySPEX: Methods for Identifying Influential Interactions in LLMs at Scale
OpenAI
Researchers from BAIR (Berkeley AI) introduced SPEX and ProxySPEX — algorithms capable of identifying critical interactions among features, data, and components of AI models, including LLMs, orders of magnitude faster than existing methods. The methods are based on the idea of sparsity of influential interactions and use techniques from coding theory and signal processing. Application of SPEX showed that standard attribution methods (e.g., SHAP) can miss complex synergies, as in the case of the trolley problem. ProxySPEX, which exploits the property of hierarchy, achieves the same accuracy with roughly 10 times fewer ablations.
Researchers from BAIR (Berkeley Artificial Intelligence Research) have developed algorithms SPEX and ProxySPEX to identify influential interactions among components of complex machine learning systems, including large language models (LLMs). The challenge is that the number of potential interactions grows exponentially with the number of features, data points, or internal components, making exhaustive enumeration computationally infeasible. SPEX leverages sparsity and low-degree assumptions: the number of truly influential interactions is small, and they involve only a small subset of elements. By using strategically selected ablations and sparse recovery algorithms from coding theory, SPEX efficiently identifies these interactions. ProxySPEX further exploits hierarchy: if a high-order interaction is important, its lower-order subsets are also important, reducing the number of ablations by about tenfold compared to SPEX. In feature attribution tasks, SPEX outperformed existing methods on long contexts (thousands of features). For instance, in a modified 'trolley problem,' the standard SHAP method attributed GPT-4o mini's incorrect answer to individual words like 'trolley,' whereas SPEX identified synergy between two occurrences of 'trolley' with 'pulling' and 'lever'; replacing these four words with synonyms reduced the model's error rate to near zero. In data attribution, ProxySPEX successfully highlighted synergistic and redundant interactions among training samples for ResNet on CIFAR-10. In mechanistic interpretability (component attribution), the method revealed how attention heads in different LLM layers interact and improved target model performance after pruning compared to other methods. The code for SPEX and ProxySPEX is available in the SHAP-IQ repository.
Source: BAIR (Berkeley AI) —
original