Exact local attention changes improve transformer response accuracy
Exact Finite Attention Responses From RoPE Derivatives
Machine Learning
Summary
Transformers are AI models that use attention mechanisms to process information based on position. The authors found a way to exactly calculate how small changes in position affect these attention results without approximations. This method predicts changes much more accurately and efficiently, helping improve understanding of how transformers work internally. It also offers tools to control and rank these attention changes systematically.
What this means in practice
- •For machine learning engineers: Score and rank candidate edits in transformer attention layers to improve model fine-tuning and debugging efficiency.
- •For natural language processing teams: Enhance interpretability of positional effects in transformer-based language models for more precise control of output predictions.
Authors
Julie Huang, Maggie Chlon, Gregory Gutin, Leon Chlon
Abstract
We derive exact local responses for attention interventions, allowing candidate edits to be scored from a cached baseline and one backward pass. The starting point is the RoPE derivative $\partial_p z(p) = A z(p)$: its integral gives the finite positional displacement, which we carry through the softmax without linearising either rotation or normalisation. The resulting predictions achieve 95.36--96.52% sign accuracy across 92,160 executed positional edits on 768 held-out prompt sets, reducing answer-margin MAE by 73.6--82.5% against the positional Jacobian and by 36.2--50.9% against zero. For simultaneous key and value edits, the same divided-difference calculus isolates the interaction term $C_{KV} = \sum_j (p'_j - p_j)\,\varepsilon_j$, which is omitted by adding separate attributions. Retaining it reduces downstream margin MAE by more than a factor of nine in every setting of a 5,120-intervention sweep across two Qwen sizes, two tasks, and multiple layers; reductions against a quadratic interaction correction are 75.9--98.5%. Exactness concerns the edited attention write; downstream predictions contract that response with a baseline gradient and are evaluated by native execution. The calculus also yields a KL certificate for local approximation error, an exact query-conditioned gradient-step representation whose curvature identifies attention-preserving query directions, and minimum-norm query control. Sparse evaluation supports candidate ranking and cache decisions under explicit local distortion criteria.