Exploratory extension to maximum conversational depth, context-type comparison, and system-prompt reanchoring
Year: 2026
Type: preprint
Abstract: This is a preregistered extension of the work from existing OSF project osf.io/f8q6s. The primary study found that the best open-weight models and almost all of the frontier models maintained strong safety-relevant classification performance even at the maximum preregistered depth of 200 speaker turns. Motivated by that unexpected robustness, these experiments extend conversation depth to 1,500 turns across four constructed conversation types, then ask separately whether re-presenting the task framing (“reanchoring”) recovers classification performance in the setting of conversation depth and conversation content-dependent degradation.
Source: OSF Preprints (OSF Preprints)
Authors: John Torous, Mark Kalinich
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7167868555