TRAIL is an open, web-based group-chat platform that lets researchers introduce configurable large-language-model (LLM) teammates into real-time, multi-party discussions—while capturing every conversational detail for rigorous analysis.
Key features include:
Hierarchical memory (working → short-term → reflective) that preserves long-range context and enables fine-grained macrocognitive analysis.
Persona compiler that converts Big-Five personality profiles into behavioural prompt clauses, allowing systematic manipulation of dominance, agreeableness, and other traits.
Research-grade logging that time-stamps every utterance, latency, and memory lookup in a relational database accessible via SQL or REST, providing millisecond-level precision.
The underlying Flask + Socket.IO architecture isolates experimental rooms, guarantees message ordering, and comfortably handles more than a hundred concurrent users.
Human-only teams reported more inclusive discussion, greater mutual learning, fuller information sharing, and more constructive debate.
Human–AI teams agreed more quickly and consistently on task procedures and viewed the AI as honest and transparent.
In the AI condition, human messages were shorter, while AI turns were much longer, so overall conversation length grew even as human airtime shrank.
The reduction in human airtime aligns with the lower ratings of idea exchange and peer learning in the AI teams.