💬 NLP

SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation

The paper introduces SPASM, a modular framework featuring Egocentric Context Projection (ECP) to enhance long-horizon stability in multi-turn dialogue generation by effectively preventing persona drift, role confusion, and echoing across diverse LLM backbones.

Original authors: Han Luo, Guy Laban

Published 2026-04-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Han Luo, Guy Laban

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are directing a play where two actors, both played by advanced AI robots, are improvising a long conversation. One actor is supposed to be a stressed customer, and the other is a helpful support agent.

In the past, when these AI actors talked to each other for a long time, they would start to forget their scripts. The stressed customer would suddenly start giving advice like a therapist, and the therapist would start complaining like a customer. They would eventually start sounding exactly the same, mirroring each other's words and losing their unique personalities. This is called "Persona Drift" and "Echoing." It's like two actors on stage forgetting who they are supposed to be and just repeating whatever the other one says.

This paper introduces a new system called SPASM (Stable Persona-driven Agent Simulation) to fix this. Think of SPASM as a super-organized Stage Manager and a clever Costume Designer that keeps the actors in character, no matter how long the play goes on.

Here is how it works, broken down into simple parts:

1. The "Character Sheet" (Persona Creation)

Before the play starts, SPASM doesn't just guess who the actors are. It uses a strict checklist (a schema) to create a detailed character profile.

The Validator: Imagine a strict casting director who checks the character sheet. If the sheet says "18-year-old student" but the plot is about "planning a retirement," the director says, "No, that doesn't make sense!" and asks for a new character.
The Crafter: Once the character is approved, this tool writes a natural, flowing biography for the actor to memorize, ensuring they have a clear voice and backstory.

2. The "Magic Glasses" (Egocentric Context Projection)

This is the paper's biggest innovation.

The Old Way (The Problem): Imagine the two actors are sitting in a room with a single, giant whiteboard in the middle. They both write everything they say on it. The problem is, the whiteboard just lists "Speaker A said X, Speaker B said Y." As the conversation gets long, the actors get confused. They look at the board and think, "Wait, did I say that? Or did they? Oh, I'll just say what they said." This causes the "Echoing."
The SPASM Way (The Solution): SPASM gives each actor a pair of Magic Glasses.
- When the Customer looks at the whiteboard through their glasses, the text automatically transforms. Every time the other person speaks, the glasses label it "THEM." Every time the customer speaks, it labels it "ME."
- When the Responder looks through their glasses, the labels flip. The customer's words become "THEM," and their own words become "ME."
- Why this works: It forces the AI to see the conversation from its own perspective. It can't accidentally copy the other person because the "Magic Glasses" constantly remind it, "That wasn't you; that was them." This stops the actors from drifting into each other's roles.

3. The "Exit Sign" (Termination Detection)

Sometimes, AI conversations go on forever and get repetitive or nonsensical. SPASM has a smart Exit Sign that watches the conversation. It knows when the customer has said "Thank you" or "I'm all set," and it gently signals the actors to stop the scene naturally, rather than letting them ramble on until they break character.

The Results: A Perfect Play

The researchers tested this system with 4,500 different characters and 45,000 conversations.

Without SPASM: The actors often forgot who they were, and the conversation became a boring loop of them repeating each other.
With SPASM: The actors stayed in character perfectly. The "Magic Glasses" (Egocentric Context Projection) almost completely eliminated the "Echoing" problem.

Why Does This Matter?

We need these AI conversations to train other AIs, test safety, or study human behavior. If the training data is full of confused robots that don't know who they are, the new AIs we build will also be confused. SPASM ensures that the "training data" is high-quality, stable, and realistic, acting like a reliable factory for creating perfect dialogue scenarios.

In short: SPASM is a toolkit that gives AI actors a clear script, a strict director, and a pair of glasses that help them remember who they are, so they can have long, meaningful conversations without losing their minds.

1. Problem Statement

Large Language Models (LLMs) are increasingly used to generate synthetic multi-turn dialogues for training, evaluation, and research in domains like counseling, tutoring, and emotional support. However, existing LLM-to-LLM simulation frameworks suffer from long-horizon instability, characterized by three primary failure modes:

Persona Drift: Agents gradually deviate from their assigned roles, goals, or personality traits as the conversation lengthens.
Role Confusion: Agents lose track of who they are versus who their partner is.
Echoing: One agent begins to mirror the language, stance, and objectives of the other, leading to a collapse of distinct identities and a loss of dialogue diversity.

Current methods typically rely on naive concatenation of dialogue history, which fails to maintain role consistency over extended interactions, rendering synthetic datasets unreliable for downstream tasks.

2. Methodology: The SPASM Framework

The authors propose SPASM (Stable Persona-driven Agent Simulation for Multi-turn dialogue generation), a modular framework designed to ensure stability without modifying model weights. The framework consists of three core stages and a novel history construction mechanism.

A. Modular Pipeline

Persona Generation:
- Schema Sampling: Generates initial profiles by sampling from predefined fields (demographics, emotional state, interaction context, behavior patterns).
- Persona Validator: An LLM-based module that checks the logical coherence and plausibility of sampled profiles (e.g., ensuring an 18-year-old isn't planning for retirement).
- Persona Crafter: Converts validated structured fields into a coherent natural-language persona description.
Dialogue Simulation:
- Involves a Client (enacting the generated persona) and a Responder (playing a specific role like a listener or expert).
Termination Detection:
- A dedicated module monitors the dialogue to identify natural stopping points (e.g., expressions of gratitude or closure) to prevent degenerate or runaway conversations.

B. Core Innovation: Egocentric Context Projection (ECP)

To address the root cause of drift and echoing, SPASM introduces Egocentric Context Projection (ECP).

Perspective-Agnostic History: Instead of storing dialogue as a static text buffer with absolute roles (e.g., "User/Assistant"), the system stores a global history $H_t$ where speaker identities are metadata ( $S = \{Client, Responder\}$ ) but the content is neutral.
Role-Relativization: Before an agent generates a response, the global history is deterministically projected into an egocentric view specific to that agent.
- The Client sees: SELF (Client) vs. PARTNER (Responder).
- The Responder sees: SELF (Responder) vs. PARTNER (Client).
Mechanism: This projection operator $\Psi_i$ maps absolute speaker IDs to relative descriptors. By conditioning generation on this normalized view, the model is prevented from misinterpreting the partner's utterances as its own continuation target, thereby reducing role ambiguity and feedback loops.

3. Key Contributions

SPASM Framework: A stability-first, modular pipeline for generating high-quality, controllable multi-turn dialogues that integrates persona validation and natural termination detection.
Egocentric Context Projection (ECP): A novel, lightweight history construction mechanism that eliminates the need for fine-tuning. It reduces role confusion and echoing to near-zero by normalizing the dialogue context from each agent's perspective.
Large-Scale Dataset & Analysis: The authors constructed a dataset of 4,500 personas and 45,000 conversations across nine combinations of three LLM backbones (GPT-4o-mini, DeepSeek-V3.2, Qwen-Plus). They provided a comprehensive geometric and behavioral analysis of these interactions.

4. Experimental Results

The paper evaluates SPASM across three dimensions:

Persona Drift Reduction:
- Using periodic probe questions (Concerns, Emotion, Motivation), the authors measured semantic drift via embedding distance.
- Result: ECP significantly reduced drift compared to standard concatenation (CONCAT). For example, on GPT-4o-mini, emotion-related drift was reduced with a Cohen's $d$ of $-0.75$.
Elimination of Echoing:
- Human Validation: Under full human annotation, ECP resulted in 0% echoing across all tested configurations. In contrast, the CONCAT baseline exhibited high echoing rates (ranging from 7% to 24% depending on the model pairing).
- LLM Judges: Automated judges also confirmed the near-total elimination of echoing with ECP.
Interaction Geometry & Retrieval:
- Clustering: Same-backbone interactions (e.g., GPT-Client + GPT-Responder) produced more compact and separable persona clusters in embedding space than cross-backbone interactions.
- Responder Dominance: The Responder model's backbone was found to dominate the emergent interaction geometry; changing the Responder had a larger impact on clustering quality than changing the Client.
- Retrieval: Persona retrieval accuracy (Acc@K) was high for same-backbone pairs (e.g., 0.96 for Top-1 in GPT/GPT) but dropped for cross-backbone pairs, indicating that cross-model interactions introduce intra-persona variance without collapsing separability.

5. Significance and Implications

Data Infrastructure: SPASM establishes a robust infrastructure for generating synthetic dialogue data, addressing a critical bottleneck in the scalability of high-quality, role-consistent datasets.
Stability without Fine-tuning: The ECP mechanism demonstrates that structural changes to context representation (rather than model weights) can solve complex stability issues like echoing and drift.
Theoretical Insight: The paper provides mechanistic hypotheses for why drift occurs (role-label ambiguity, alignment priors, and closed-loop feedback loops) and offers a practical solution to mitigate them.
Resource Release: The authors release the 45k-conversation dataset and the codebase, facilitating future research in controllable dialogue synthesis, robust evaluation, and stable agent simulation.

In summary, SPASM provides a proven method to generate long, coherent, and role-consistent dialogues between LLM agents, solving the "identity collapse" problem that has hindered the reliability of synthetic dialogue generation.