# PRISM

Molecular properties depend on chemical structure at different ranges, yet graph neural networks usually fix one propagation depth for every atom. PRISM turns each atom's own encoded features into a bounded scalar that positions a unimodal distribution over retained message-passing states, so the allocated depth is analytically tractable and moves monotonically with that scalar. Under matched scaffold partitions and tuning budgets, PRISM trained from scratch reaches the lowest or tied-lowest mean error on ESOL and FreeSolv, and on a pre-trained GIN it improves on last-layer readout on four of six endpoints.

Molecular properties depend on chemical structure at different ranges, yet graph neural networks usually fix one propagation depth for every atom. PRISM turns each atom's own encoded features into a bounded scalar that positions a unimodal distribution over retained message-passing states, so the allocated depth is analytically tractable and moves monotonically with that scalar. Under matched scaffold partitions and tuning budgets, PRISM trained from scratch reaches the lowest or tied-lowest mean error on ESOL and FreeSolv, and on a pre-trained GIN it improves on last-layer readout on four of six endpoints.

The work proposes using the variance of specialists' task vectors across specialists as a measure of interference, yielding a pre-merge score that predicts collapse, and designs PRISM: an operator that averages task vectors first and then soft-thresholds each layer at a level set by that layer's interference; across twenty-two merge configurations from four model families only destructive merges exceed the score threshold, twelve of fourteen pre-evaluation predictions were correct, and PRISM keeps all five destructive merges above the threshold within evaluation noise of the base model without data or tuning, where plain averaging falls at least 14.4 points below it or collapses entirely.

The work proposes using the variance of specialists' task vectors across specialists as a measure of interference, yielding a pre-merge score that predicts collapse, and designs PRISM: an operator that averages task vectors first and then soft-thresholds each layer at a level set by that layer's interference; across twenty-two merge configurations from four model families only destructive merges exceed the score threshold, twelve of fourteen pre-evaluation predictions were correct, and PRISM keeps all five destructive merges above the threshold within evaluation noise of the base model without data or tuning, where plain averaging falls at least 14.4 points below it or collapses entirely.

Through empirical analysis, this work identifies a "similarity bias" in dot-product self-attention that systematically overlooks heterogeneous item relations carrying causal signal for target prediction, and proposes PRISM, a module that calibrates attention through K Perspective Lenses operating in two complementary views, an Affinity View and a Contrast View, consistently outperforming state-of-the-art baselines on seven real-world benchmarks.

Through empirical analysis, this work identifies a "similarity bias" in dot-product self-attention that systematically overlooks heterogeneous item relations carrying causal signal for target prediction, and proposes PRISM, a module that calibrates attention through K Perspective Lenses operating in two complementary views, an Affinity View and a Contrast View, consistently outperforming state-of-the-art baselines on seven real-world benchmarks.