We introduce State-Grounded Conditioning (SGC) as a design principle for user‑facing LLM agents that must condition on live user state (e.g., game progress, session history, live inventory). SGC externalises state‑dependent control into rule kernels over structured inputs and wraps them with three primary layers—Perception, Grounding, and Interaction—each exposing explicit conditioning dependencies. We also define a distinct failure mode called direction drift, where a task‑complete response chooses a direction misaligned with the current state.
Our evaluation uses an anonymised benchmark of 200 sessions (≈ $1,000$ assistant model turns) from an in‑game conversational coaching agent that guides players through consecutive competitive matches. We report mean first‑token latency and five human‑annotated dialogue‑quality metrics that jointly cover factual grounding and coach‑like guidance progression.
The Perception wrapper keeps mean first‑token latency at $1.5\text{s}$, compared with $6.1\text{s}$ for a PE‑Agent inside a production tool‑use harness. Enabling all three wrappers lifts turn‑level grounded accuracy from $61.1\%$ / $69.8\%$ (Prompting / PE‑Agent) to $96.7\%$, and session‑level grounded accuracy from $20.0\%$ / $26.5\%$ to $83.5\%$. Session‑level grounding‑failure incidents drop by roughly $78\%$ relative to the strongest baseline.
A cumulative ablation shows complementary incremental gains as the wrappers are added, suggesting approximate orthogonality among state slices, though independent per‑wrapper effects remain unestablished.
Review