Representational signatures of cognitive seeding: continuous decision signals and layerwise dissociation across model regimes

Read full text Summarise with VIRAT
Authors: YONGJOO KIM
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Redesigned preregistration for Paper 2 of the Cognitive Seeding research program. This study tests whether the functional dissociation between cognitive seed effects and chain-of-thought (CoT) effects, established in Kim (2026; Nature Machine Intelligence, under review; OSF: https://osf.io/jrc2k ), corresponds to partially separable representational signatures across transformer processing layers. Methods include continuous decision signal extraction, layerwise representation geometry analysis, and linear probing across four open-source models (Qwen-7B, Qwen-72B, Llama-8B, Llama-70B) and four frontier models (Claude Haiku, Sonnet, Opus, GPT-5.2). This registration supersedes terminated OSF v1 (https://osf.io/snr4u ), which was discontinued because item-level binary causal tracing proved mismatched to aggregate-level behavioral effects. Four confirmatory hypotheses (CH1–CH4) focus on continuous signal shifts, CoT selectivity, A–F representational divergence, and layerwise probe dissociation.

Fields of Study & Subjects

Explainable Artificial Intelligence (XAI) Topic Modeling Cognitive Science and Education Research

Publication details

Publication date
1 January 2026
Language
EN
Document type
Preprint