SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation
| Authors | Han Luo & Guy Laban |
| Year | 2026 |
| HF Upvotes | 3 |
| arXiv | 2604.09212 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Large language models are increasingly deployed in multi-turn settings such as tutoring, support, and counseling, where reliability depends on preserving consistent roles, personas, and goals across long horizons. This requirement becomes critical when LLMs are used to generate synthetic dialogues for training and evaluation, since LLM--LLM conversations can accumulate identity-related failures such as persona drift, role confusion, and "echoing", where one agent gradually mirrors its partner. We introduce SPASM (Stable Persona-driven Agent Simulation for Multi-turn dialogue generation), a modular, stability-first framework that decomposes simulation into (i) persona creation via schema sampling, plausibility validation, and natural-language persona crafting, (ii) Client--Responder dialogue generation, and (iii) termination detection for coherent stopping. To improve long-horizon stability without changing model weights, we propose Egocentric Context Projection (ECP): dialogue history is stored in a perspective-agnostic representation and deterministically projected into each agent's egocentric view before generation. Across three LLM backbones (GPT-4o-mini, DeepSeek-V3.2, Qwen-Plus) and nine Client--Responder pairings, we construct a dataset of 4,500 personas and 45,000 conversations (500 personas X 10 conversations per pairing). Ablations show ECP substantially reduces persona drift and, under human validation, eliminates echoing; embedding analyses recover persona structure and reveal strong responder-driven interaction geometry. Our code is available at https://github.com/lhannnn/SPASM.
Engineering Breakdown
Plain English
This paper addresses a critical problem in LLM deployment: when large language models generate synthetic dialogue data by talking to each other, they suffer from consistency failures like persona drift, role confusion, and 'echoing' where agents start mirroring each other. The authors introduce SPASM (Stable Persona-driven Agent Simulation for Multi-turn dialogue), a framework that decomposes the dialogue generation process into modular components: persona creation through schema sampling and validation, and Client-Responder interaction with explicit role preservation. The approach aims to produce high-fidelity synthetic training and evaluation data for multi-turn dialogue tasks like tutoring, support, and counseling by maintaining stable, consistent character identities across long conversation horizons.
Core Technical Contribution
The core innovation is decomposing LLM-LLM dialogue generation into a stability-first pipeline with explicit persona engineering rather than leaving it to emergent behavior. Most prior work treats dialogue generation as a black-box end-to-end problem, but SPASM introduces a modular architecture that separates concerns: persona definition (schema sampling for diversity, plausibility validation for coherence, natural-language persona crafting for embedding into prompts), then Client-Responder simulation with role-locking mechanisms. This addresses a failure mode specific to synthetic data generation—persona drift and echoing—that doesn't appear in human-supervised dialogue but emerges predictably in uncontrolled LLM interactions. The contribution is both architectural (the modular decomposition) and methodological (the specific validation and role-preservation techniques).
How It Works
SPASM operates in three main phases. First, persona creation: the system generates candidate personas using schema sampling (likely structured templates for traits, goals, background), validates them for plausibility (checking for internal contradictions or implausible combinations), and renders them into natural-language descriptions that get embedded in system prompts. Second, it instantiates a Client-Responder dialogue structure where roles are explicitly defined and maintained—the Client maintains a consistent goal or persona, the Responder maintains a consistent professional or support role. Third, during multi-turn generation, the framework enforces role consistency through prompt engineering and likely monitoring mechanisms that detect and prevent persona drift (when a character's behavior diverges from their definition) and echoing (when conversational partners begin to converge on similar behavior or vocabulary). The output is a corpus of synthetic dialogues where identity and role boundaries remain stable over many turns, suitable for training dialogue models that need consistent agent behavior.
Production Impact
For teams building dialogue systems (chatbots, tutoring agents, customer support), SPASM directly reduces the need for expensive human annotation of multi-turn conversations. Instead of hiring annotators to write or validate long dialogues, you can generate synthetic data at scale while maintaining consistency guarantees that standard LLM generation lacks. In production pipelines, this means faster iteration on dialogue policies and persona design: you can sample different personas from schemas, validate them automatically, and generate training data without manual quality control overhead. The trade-offs are non-trivial: the modular approach requires upfront engineering of persona schemas, plausibility validators, and role-preservation mechanisms (not a drop-in replacement), and the generated data is only as good as your schema design—poorly designed personas won't produce diverse or realistic dialogues. Integration cost is moderate: you need to define your domain's persona dimensions, but once done, the sampling and generation are automated. For large-scale training, the compute cost is lower than human annotation but higher than naive LLM generation.
Limitations and When Not to Use This
The paper assumes that stable persona definitions can be meaningfully formalized via schemas, which may not hold for complex, evolving characters or nuanced roleplay scenarios where contradiction and growth are features, not bugs. The plausibility validation mechanism is not fully described in the abstract, leaving unclear how it avoids becoming a bottleneck or overfitting to specific cultural or contextual assumptions baked into validators. The approach is evaluated only on synthetic data quality (persona consistency, avoiding echoing), but the abstract doesn't indicate whether synthetic dialogues generated by SPASM actually improve downstream task performance when used for training real dialogue models—this is a critical missing piece. The framework also likely requires significant tuning per domain and application; personas for medical counseling, technical support, and tutoring may have entirely different schema requirements, limiting generalizability. Finally, there's no discussion of how SPASM handles adversarial or edge-case scenarios where users deliberately try to break character consistency.
Research Context
This work builds on a growing body of research on synthetic data generation and data augmentation for NLP, particularly multi-turn dialogue tasks (prior work on dialogue evaluation, persona-based dialogue generation, and LLM-LLM interaction stability). It directly addresses failure modes documented in recent work on LLM hallucination and consistency in long-context generation, where unconstrained LLMs are known to drift in identity and goals. The paper fits into the broader category of LLM reliability and alignment research—ensuring that deployed models maintain their intended behavior—but focuses on the data generation angle rather than fine-tuning or RLHF. It opens research directions in automatic persona validation, formal methods for role consistency, and scalable synthetic dialogue generation, which are increasingly important as dialogue systems move into safety-critical domains like mental health and education.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::