Nabh Mehta (MS-MBA) Principal Investigator | Director and Partner, Phlowerman Ventures LLP phlowermanventuresllp@gmail.com Abstract Large Language Models (LLMs) configured as specialized agents rely heavily on extensive system prompts to enforce behavioral, lexical, and stylistic personas. However, during extended multi-turn interactions, these models experience Persona Drift due to attention dilution across widening context horizons. While brute-force re-prompting introduces severe token overhead, this paper evaluates Symbolic Pseudonym Anchoring (SPA)—a context engineering framework that maps dense behavioral schemas to compact, model-safe symbolic identifier tokens—as a low-overhead intervention. Rather than claiming universal efficacy, we present a rigorous cross-model evaluation of SPA against natural-language reminders and structural baselines across ten distinct open-weight model architectures using a 7-arm ablation matrix. Empirical results show that SPA’s effectiveness is highly model-dependent: while Continuous SPA improved aggregate Psem relative to the baseline for specific architectures like Qwen2.5:3B and Gemma2:9B, it introduced reduced persona-preservation performance in others, such as the Llama-3 and Mistral variants. In cross-model aggregate performance, a standard Generic Reminder achieved the highest overall persona preservation score (0.686), whereas Continuous SPA and Drift-Triggered SPA achieved 0.673 and 0.677 respectively. These findings suggest that symbolic token anchoring does not universally outperform natural-language constraints but serves as a highly conditional, model-specific intervention.
Nabh Sanjay Mehta (Thu,) studied this question.