LinkedIn publication text
Single-turn or multi-turn?
Single-turn or multi-turn? It is not just UX: it is an architecture decision that defines cost, reliability, risk and governance of any LLM product. 👇
🔵 APIs keep no state: the whole history is resent on every turn, and cumulative input grows ~T² (quadratic, not exponential).
🟢 Prompt caching cuts input cost; reasoning (output) then dominates the bill. The effort level is the biggest cost lever.
🟡 Laban et al. (2025): 15 LLMs lose 39% of performance on average when the same task is revealed over several turns (unreliability +112%, aptitude −16%). More reasoning does not fix it.
🟠 Multi-turn widens the risk surface: gradual escalation (Crescendo), indirect prompt injection and persistent malicious state.
🔴 What helps: consolidated requirements, typed immutable state, deterministic authorization outside the LLM and an audit trail.
🟣 Source rigor: I audited the material against primary sources and fixed 8 points (e.g., the NIST AI RMF is AI 100-1, not "SP 1270"; the EU AI Act high-risk calendar moved to 2 Dec 2027 under Reg. (EU) 2026/1744).
The carousel has 28 slides: foundations, a 2017–2026 timeline, evidence, cost by provider, security, SGAEIA, glossary and an audited bibliography. Cost figures are my own calculations, with explicit assumptions.
What is your rule for choosing between single-turn and multi-turn?
© Aridio Silva @aridiosilva — November 2026 · CC BY 4.0
Aridio Silva — Independent Researcher