Abstention and Noise Filtering: Two Missing Primitives of Softmax Attention
A batch of arXiv papers treated abstention, calibration and routing as first-class model capabilities.
TL;DR
- One arXiv paper argued that abstention and noise filtering are two missing primitives of softmax attention.
- A second proposed an interpretable memory decision controller for LLM agents based on three-signal complementarity, decoupling confidence and consistency.
- Two more addressed reliability: RACER for role-aligned competence estimation in human-AI routing, and 'Available Guardrails' for certifying selective prediction.
One arXiv paper argued that abstention and noise filtering are two missing primitives of softmax attention. [1]
A second proposed an interpretable memory decision controller for LLM agents based on three-signal complementarity, decoupling confidence and consistency. [2]
Two more addressed reliability: RACER, a role-aligned competence estimator for human-AI routing, and 'Available Guardrails', which certifies selective prediction across ML systems. [3] [4]
On Hacker News, a Show HN said an attention kernel (MoA) was proven minimal before the code was written. [5]
Why it matters
Abstention, calibration and routing are becoming first-class research primitives, which is what turns raw model capability into deployable reliability.
Editor's note
All arXiv items are cited at the abstract level and were not peer-reviewed here; the MoA minimality claim is from its authors.