Work

Evaluating Symbolic Persona Anchoring and Direct KV-Cache Interventions for Multi-Turn LLM Persona Preservation: A Six-Method, 10-Model Longitudinal Study

Year: 2026
Type: preprint
Abstract: Maintaining consistent behavioral personas across extended multi-turn dialogues remains a central challenge for open-weight Large Language Models (LLMs), which suffer from progressive persona drift, attention dilution, and vulnerability to adversarial prompts or context flooding. Extending our prior evaluation of Symbolic Persona Anchoring (SPA) against natural-language reminders (Mehta, 2026), this study conducts a balanced 6 × 10 × 4 longitudinal benchmark (240 trajectories, 12,000 turns across 50-turn rollouts) comparing six inference-time interventions across 10 open-weight models (3B–14B) and four personas: baseline controls (A1 Raw History, A2 Generic Reminder), symbolic anchoring (A3 Continuous SPA), and multi-turn adaptations of direct Key-Value (KV) cache interventions (A4 Belitsky-style additive steering, A5 Sun-style mid-layer transplantation, and A6 Memory Inception latent KV banks). Across N = 40 matched cells per arm, A2 achieves the highest mean semantic preservation (Psem = 0.4781 ± 0.0597) but incurs a 27.5% late-horizon MATTR50-decline flag rate (ΔMATTR50 < −0.08). Conversely, A3 maintains competitive semantic preservation (Psem = 0.4589 ± 0.0600) with no significant differential drift relative to A2 (p = 0.388) while achieving the highest lexical diversity (MATTR50 = 0.8155) and lowest MATTR50-decline incidence (7.5%). Direct KV-cache interventions exhibit sharp architectural sensitivity: A5 peaks on Mistral and Qwen-2.5-7B backbones yet suffers the steepest longitudinal decay (p < 0.001), whereas zero-shot A4 and A6 regress toward unanchored A1. All trajectory logs and metrics are byte-verified via SHA-256 checksums and bitwise recomputation. Source Code and Reproducibility The associated GitHub repository contains the research project's codebase and reproducibility resources for the six-method, 10-model longitudinal benchmark. GitHub Repository: KV-Cache vs. SPA Official GitHub Release: KV-Cache vs. SPA — Research Benchmark Release v1.0.0 Release Version: v1.0.0 Citation If you find this research useful in your work, please consider citing it using the following DOI: DOI: 10.5281/zenodo.24424861 Suggested citation:Nabh Sanjay, M. (2026). Evaluating Symbolic Persona Anchoring and Direct KV-Cache Interventions for Multi-Turn LLM Persona Preservation: A Six-Method, 10-Model Longitudinal Study (Version 1.0.0). Zenodo. https://doi.org/10.5281/zenodo.24424861 Research Institution:Phlowerman Ventures LLPAI Systems & Mechanistic Interpretability Research
Language: English
Open Access status: green
OpenAlex ID: W7237340630
Cites:
Cited by:
Related to: