State-Grounded Conditioning is introduced, a design principle for user-facing LLM agents that must condition on live user state (game state, session history, live inventory), and a distinct failure class the authors call direction drift: task-complete responses whose chosen direction misaligns with the current state.
Abstract
We introduce State-Grounded Conditioning (SGC), a design principle for user-facing LLM agents that must condition on live user state (game state, session history, live inventory), and a distinct failure class we call direction drift: task-complete responses whose chosen direction misaligns with the current state. SGC externalises state-dependent control into rule kernels over structured inputs and three primary state slices, via Perception, Grounding, and Interaction wrappers with explicit conditioning dependencies. We evaluate SGC on a 200-session anonymised benchmark ($\approx$1,000 assistant model turns) from an in-game conversational coaching agent that guides players through consecutive competitive matches, reporting mean first-token latency and five human-annotated dialogue-quality metrics that jointly cover factual grounding and coach-like guidance progression. The Perception wrapper holds mean first-token latency at 1.5s (vs. 6.1s for 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 $\approx$78% relative to the strongest baseline. A cumulative ablation shows complementary incremental gains as the wrappers are added. These results inform approximate state-slice orthogonality, without establishing independent per-wrapper effects.
SAGE (State-Grounded Abstention-Aware Evaluation), which compiles a workflow specification and per-turn state diff into atomic, schema-grounded criteria and routes each through a cascade of symbolic and encoder/NLI verifiers that abstain rather than guess, aggregating criterion verdicts into a turn-level decision with...
Rayan Khoury, Shih-Yao Lin, P. Mishra· 0 citations
The results suggest that broad SFT brings most of the model's capability improvement; turn-local supervision can be effective when failure detection is precise, with observed transfer concentrated primarily within-family.
These findings motivate a state-action controller that starts from the base model action and selectively edits it using structural readouts, without requiring a complete predicted belief state as an intermediate representation.
Large language models (LLMs) often solve a fully specified instruction but fail when the same evidence arrives over multiple turns, a failure known as Lost-in-Conversation (LiC). We present CM-USR, a multi-agent inference-time systems that combines Critical-Missing (CM) calibration with User State and Regeneration (USR...
Ya-Qian Xu, Xiong Li· 2026 6th International Confe...· 0 citations
Intent drift is established as a measurable multi-turn failure mode and explicit state maintenance as a partial mitigation and IntentFlux is introduced, an executable benchmark that converts verifiable tasks into dialogues with controlled intent changes while preserving their original graders.
General turn-taking behavior in real-time dialogue systems requires deciding whether to keep listening or start responding while listening, and whether to continue or stop while speaking. Existing turn detectors use heterogeneous, task-specific label spaces and are often trained on limited annotations or evaluated on i...
Zhan-Xun Liu, Yi-Fan Duan, Heng-Tao Wu et al.· 0 citations