At a Glance
A small change in how conversation history is presented to each simulated participant—showing a neutral history from each agent’s own viewpoint—keeps role-based chatbots stable over long dialogs and eliminates the common echoing failure.
ON THIS PAGE
Key Findings
A persona-first simulation pipeline (SPASM) that validates and crafts personas, runs two-agent conversations, and detects natural stops produces high-quality multi-turn dialogue. The core trick, called Egocentric Context Projection, stores a shared, neutral history and projects it into each agent’s personal view before generation. That projection prevents agents from mirroring each other or drifting away from their assigned persona, producing role-consistent conversations even over long exchanges.
Explore evaluation patternsSee how to apply these findings
Data Highlights
145,000 total simulated conversations across 9 model pairings (9 backbones × 500 personas × 10 conversations each).
2Max conversation length capped at 50 utterances (25 turns per agent) to avoid runaway interactions.
3Egocentric Context Projection eliminated echoing in the human-validated assessments (effectively 0 echoing cases) and sharply reduced persona drift versus standard history concatenation.
Implications
Engineers building multi-turn conversational agents and data teams creating synthetic dialogue should use this approach to produce cleaner, role-consistent training data. Evaluation and safety teams can adopt the projection method to get more reliable agent-to-agent evaluation and clearer failure-mode signals. For teams focused on practical applications, consider Multi-Agent Customer Support as a use-case example to guide deployment decisions.
Key Figures
Fig 1: Figure 1: SPASM pipeline for stable persona-driven dialogue generation, consisting of (i) modular persona generation (schema sampling, validation, and crafting), (ii) dialogue simulation with egocentric context projection over a perspective-agnostic history, and (iii) a termination detector for natural and coherent stopping.
Fig 2: (a) Concerns
Fig 3: (a) GPT-4o-mini / GPT-4o-mini.
Fig 4: (a) GPT-4o-mini / GPT-4o-mini.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Results are demonstrated on three largely English, instruction-tuned model families; effectiveness on other languages, architectures, or much smaller models is untested. SPASM focuses on two-party interactions with fixed roles and may not directly extend to group chats or dynamic role-switching. Persona descriptions are schema-driven and may miss the full richness of real human behavior, and human validation used to confirm echoing is subjective and hard to scale. When thinking about potential limitations, organizations may want to monitor for Context Drift.
The Details
SPASM is a modular pipeline for generating persona-driven, multi-turn conversations without letting simulated agents lose their identity over time. Rather than concatenating raw chat history, SPASM builds a single, perspective-agnostic history and then projects that history into an agent-specific view (the agent’s “self” vs. the partner). For example, instead of replaying the full transcript to both participants, the system rewrites the shared history from the client’s perspective when prompting the client and from the responder’s perspective when prompting the responder. The pipeline also includes persona sampling from structured fields (age, occupation, emotion, behavior style), a plausibility validator to filter unrealistic personas, and a termination detector to stop conversations at natural endpoints. Across nine client–responder model pairings (three backbone families), the team generated 45,000 conversations (500 personas × 10 conversations per pairing) with a 50-utterance cap to avoid runaway threads. The egocentric projection method (ECP) dramatically improved long-horizon stability: echoing—where one agent simply mirrors the other—was eliminated in human-validated checks and persona drift was greatly reduced versus standard history concatenation. Practically, that means synthetic datasets and agent-to-agent evaluations derived from SPASM are less contaminated by role confusion, making them more useful for training, auditing, and safety testing. The dataset and code are released to help teams produce more reliable multi-turn dialogue data and to improve agent-to-agent evaluation practices. Safety testing
Need expert guidance?We can help implement this
Credibility Assessment:
One author (Guy Laban) has moderate h-index (~14), suggesting a recognized researcher; still an arXiv preprint with no citations, so mid-level credibility.