Intel / article

Identifying Interactions at Scale for LLMs

Source: bair.berkeley.edu
Date: Published Mar 13, 2026

Context

An explanation of an AI decision can miss the cause if it scores each input independently. Two phrases, training examples, or internal components may matter chiefly through their interaction. A research post from Berkeley AI Research describes SPEX and ProxySPEX, which look for these combinations by measuring how outputs change when selected parts are removed.

Testing every combination would become prohibitively expensive. SPEX instead assumes that relatively few interactions dominate and usually involve small groups. ProxySPEX adds the assumption that important larger combinations tend to contain important smaller ones. The authors report that ProxySPEX matches SPEX's performance with around ten times fewer removal tests, making interaction analysis more feasible under those structural assumptions.

Key takeaways · AI-distilled

ProxySPEX exploits a 'hierarchy' property, where a higher-order feature interaction being important implies its lower-order subsets are too, to match SPEX's accuracy with about 10x fewer ablations.
On a modified trolley problem where GPT-4o mini answered correctly only 8% of the time, SHAP blamed the word 'trolley' alone, but SPEX found the real driver was a synergy of both 'trolley' instances plus 'pulling' and 'lever'; replacing all four cut the failure rate to near zero.
Applied to a ResNet on CIFAR-10, ProxySPEX distinguished synergistic training-data interactions (different classes jointly explaining a prediction, e.g. sports-car and truck images explaining an automobile prediction) from redundant near-duplicates reinforcing one concept.
Analyzing -head interactions by depth on an MMLU task, early layers act in a largely linear/independent regime while later layers show heads interacting heavily within the same layer; a ProxySPEX-informed pruning strategy actually improved target-task performance.

Terms in this piece · Glossary

attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.

Read the source bair.berkeley.edu