What question did this study set out to answer?

The aim is to enhance AI alignment by ensuring that language models maintain causal coherence in their outputs.

April 24, 20260 citationsOpen Access

Circle-Grounded LLM Training: A Proposal for Causality-Based AI Alignment

Key Points

The aim is to enhance AI alignment by ensuring that language models maintain causal coherence in their outputs.
Introduces a five-layer architectural framework based on causal continuity principles.
Filters datasets by assessing causal coherence rather than toxicity.
Implements dual reward signals combining human preference and causal coherence scores.
Proposed methods are expected to improve the structural quality of reasoning in language model outputs.
Implementing the proposed architecture may lead to more coherent and aligned AI responses.

Abstract

Large language models are trained on the full spectrum of human text —from rigorous science to manipulation and pseudologic. Current alignmentmethods (RLHF, categorical filters, system prompts) do not address thestructural quality of reasoning: whether a model's output preserves ordestroys causal coherence in the user's mind. This proposal introduces a five-layer training architecture grounded inthe Circle/Void framework from the Book of Circle (Conception Circle of Being) —a philosophical text treating causal continuity as the fundamentalproperty of healthy systems. The five layers: 1. Dataset filtering by causal coherence (not toxicity) 2. Dual reward signal: human preference + causal coherence score3. Architectural reflection head (non-bypassable pre-output check) 4. RAG grounding against the canonical source text at inference time5. Cross-session memory tracking downstream effects of responses The canonical source texts and a working system prompt implementationare available at: github. com/sae-ru/LLMConceptualSafetyPrompts

Perguntar à IA

Bookmark

View Full Paper