system2-attention

Name: system2-attention
Availability: InStock
Author: plurigrid

Apply System2-attention to regenerate context and refine transformer attention weights.

60|13|Updated Dec 22, 2025

Search Tags:#grounding #attention #transformers #deliberate #two-pass #system2

Ready to run

npx skills add https://github.com/plurigrid/asi --skill system2-attention

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: system2-attention
Source: https://github.com/plurigrid/asi/tree/main/ies/music-topos/.ruler/skills/system2-attention
Command: npx skills add https://github.com/plurigrid/asi --skill system2-attention

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

System 2 Attention (S2A) validates and filters transformer attention by regenerating context and applying a deliberate second pass to improve factual grounding and reduce noise.

Core Features & Use Cases

Context filtering to remove irrelevant information.
Two-pass attention strategy: fast pass followed by deliberate re-attention.
Grounding validation to measure factual alignment.

Quick Start

Apply S2A filtering to a given query and context and compare the results of the two passes.

Frequently Asked Questions about system2-attention

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How does two-pass attention improve transformer reliability?▼

Two-pass attention regenerates context and applies a deliberate second pass to re-attend over refined information, reducing noisy or sycophantic outputs and improving factual grounding compared to single-pass approaches.

When should I use attention filtering for long-context reasoning?▼

Attention filtering removes irrelevant information from long contexts, making it essential when transformers struggle with factual accuracy across extended sequences or when irrelevant details distract from core reasoning.

Can I reduce hallucinations in transformer outputs with context regeneration?▼

Yes. System 2 Attention validates factual grounding by regenerating context and applying uncertainty-driven re-attention, filtering opinionated or unfounded content to produce more reliable, grounded outputs.

How does grounding validation measure factual alignment?▼

Grounding validation compares regenerated context against original attention weights to measure factual alignment, identifying and filtering attention directed at irrelevant or contradictory information.

What's the difference between fast-pass and deliberate re-attention?▼

The fast pass captures initial attention patterns; deliberate re-attention applies System 2 reasoning to reconsider and refine those weights based on grounding validation, reducing noise and improving accuracy.