With in-context learning (ICL), transformer models adapt to new tasks from illustrative examples at inference time without updating their parameters. A common interpretation is that transformers implement implicit learning algorithms in their forward pass. We test this interpretation and ICL capability under distribution shift in noiseless polynomial regression, where the context uniquely determines the target function. Models that are near-perfect in distribution exhibit sharp performance degradation under shifts in the input or target-function distribution. This behavior persists across polynomial degrees and multiple training interventions, and in higher-dimensional nonlinear regression. We identify two architectural mechanisms that explain this breakdown: output saturation at finite boundary values linked to final normalization, and attention concentration on a small subset of the context, which can limit context-dependent adaptation required for ICL. Controlled experiments, architectural ablations, targeted interventions, and theoretical analysis provide strong evidence for both mechanisms and their consequence: strong in-distribution ICL does not imply distribution-independent ICL, which calls for caution when using ICL instead of explicit retraining or adaptation under distribution shift.
Ample empirical evidence in deep neural network training suggests that a variety of optimizers tend to find nearly global optima. In this article, we adopt the reversed perspective that convergence to an arbitrary point is assumed rather than proven, focusing on the consequences of this assumption. From this viewpoint, in line with recent advances on the edge-of-stability phenomenon, we argue that different optimizers effectively act as eigenvalue filters determined by their hyperparameters. Specifically, the standard gradient descent method inherently avoids the sharpest minima, whereas Sharpness-Aware Minimization (SAM) algorithms go even further by actively favoring wider basins. Inspired by these insights, we propose two novel algorithms that exhibit enhanced eigenvalue filtering, effectively promoting wider minima. Our theoretical analysis leverages a generalized Hadamard--Perron stable manifold theorem and applies to general semialgebraic $C^2$ functions, without requiring additional non-degeneracy conditions or global Lipschitz bound assumptions. We support our conclusions with numerical experiments on feed-forward neural networks.
Omar Naim, Swarnadeep Bhar, Jerome Bolte, Nicholas Asher CNRS
原文摘要
While transformer models exhibit strong in-context learning (ICL) abilities, they often fail to generalize under simple distribution shifts. We analyze these failures and identify Softmax, the scoring function in the attention mechanism, as a contributing factor. We propose Scaled Signed Averaging (SSA), a novel attention scoring function that mitigates these failures. SSA significantly improves performance on our ICL tasks and outperforms transformer models with Softmax on several NLP benchmarks and linguistic probing tasks, in both decoder-only and encoder-only architectures.
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表