State-Grounded Conditioning: Wrapping User-Facing LLM Agents Where Direction Depends on Live State

Qi Liu; Xiaoyang Yuan; Yubin Ruan; Zhuomeng Zhang; Wenjin Wang; Di Wu; Mingye Xu; Xinyi Mou

cs.AISep 23, 2026

State-Grounded Conditioning: Wrapping User-Facing LLM Agents Where Direction Depends on Live State

Authors: Qi Liu, Xiaoyang Yuan, Yubin Ruan, Zhuomeng Zhang, Wenjin Wang, Di Wu, Mingye Xu, Xinyi Mou,

Organizations: Tencent

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.

Figures & tables

Figure 1: SGC: primary state slices and explicit conditioning dependencies. Dashed arrow carries $c_{1}$ into $\mathcal{W}_{2}$ ; separate responsibilities do not imply independent computations. Reply LLM $\mathcal{L}$ additionally conditions on shared input $x$ and static state $s_{\text{static}}$ .

Figure 2: $\mathcal{W}_{3}$ scheduler decision tree with two paths (fire / abstain) and the meta writeback loop.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Failure notion	Per- user?	Per- moment?	State-var dependent?
Hallucination	$\times$	$\times$	$\times$
Personalization failure	✓	$\times$	$\times$
Context-adherence failure	$\times$	partial	$\times$
Commonsense-constr. viol.	$\times$	$\times$	partial
Classical DST error	partial	✓	✓
Direction drift (ours)	✓	✓	✓

Appendix

Table 1: Direction drift vs. five neighbouring failure notions along three orthogonal axes; only direction drift is the conjunction of all three.

Wrapper	Reads (state fields)	Writes / emits ( $c_{i}$ )
$\mathcal{W}_{1}$	intent, sub_intent,	$\hat{T}$ (tool set),
	snapshot game state,	<NARRATION> tokens,
	device / session opener	early-decision flag
$\mathcal{W}_{2}$	QuerySlots (from $x$ ),	Conflict set + tier,
	$\sigma(s_{\text{live}}^{\text{fs}})$ inventory,	eligible candidates $\mathcal{A}$
	catalog / retrieval	required paragraphs $\mathcal{R}$

Appendix

Table 2: Primary state and control-output summary. Parsed slots, candidate pools, and weights are derived inputs or configuration, not additional disjoint raw-state fields. Both

\mathcal{W}_{2}

and

\mathcal{W}_{3}

also consume

c_{1}

(signatures below).

Figure 3: Fact-talk separation in $\mathcal{W}_{2}$ . The FACT layer performs slot extraction, deterministic repair, slot diff against live facts $\sigma(s_{\text{live}}^{\text{fs}})$ , three-tier grading (hard / soft / hint), and paragraph mutex $\Psi$ ; the TALK layer performs constrained generation over the enabled paragraphs. Structural selection does not guarantee semantic correctness of the final text.

Metric	Prompting	PE-Agent	SGC w/o $\mathcal{W}_{2},\mathcal{W}_{3}$	SGC w/o $\mathcal{W}_{3}$	Full SGC
$M_{0}$ · First-token latency (mean, s, $\downarrow$ )
Latency	2.3	6.1	1.5	1.5	1.5
$M_{1}$ · TGA (%, $\uparrow$ )
Turn-level grounded acc.	61.1	69.8	89.1	89.9	96.7
$M_{2}$ · SGA (%, $\uparrow$ )
Session-level grounded acc.	20.0	26.5	46.5	56.5	83.5

Appendix

Table 3: Aggregate benchmark metrics across all five configurations on the 200-session anonymised benchmark (

\approx\!1{,}000

assistant model turns; Section 7.1 ).

M_{0}

is a system-level cost budget;

M_{1}

–

M_{5}

are downstream dialogue-quality metrics defined in Section 7.2 . PE-Agent is a Plan-and-Execute agent ( Wang et al., 2023 ) deployed within a production tool-use harness ( Xu et al., 2026 ) . The two intermediate SGC rows disable

\mathcal{W}_{3}

(and also

\mathcal{W}_{2}

) to isolate incremental wrapper contributions. Bold indicates the best result in each row (including ties). All five configurations are scored under the same dual human-annotation protocol (Appendix G ), so between-group deltas are directly comparable; no LLM-judge label enters any reported number (Limitations L2).

Contrast	$\Delta M_{0}$	$\Delta M_{1}$	$\Delta M_{2}$	$\Delta M_{3}$	$\Delta M_{4}$	$\Delta M_{5}$
	(s)	(pp)	(pp)	(pts)	(pts)	(pts)
$\mathcal{W}_{1}$ vs. PE	$-4.6$	$+19.3$	$+20.0$	$+1.9$	$+2.2$	$+0.9$
Add $\mathcal{W}_{2}$	$0.0$	$+0.8$	$+10.0$	$0.0$	$0.0$	$-0.1$
Add $\mathcal{W}_{3}$	$0.0$	$+6.8$	$+27.0$	$+0.2$	$+0.8$	$+0.3$

Appendix

Table 4: Cumulative differences calculated from the displayed values in Table 3 . PE denotes PE-Agent; successive rows compare

\mathcal{W}_{1}

with PE-Agent,

\mathcal{W}_{1}{+}\mathcal{W}_{2}

with

\mathcal{W}_{1}

, and Full SGC with

\mathcal{W}_{1}{+}\mathcal{W}_{2}

. Lower

M_{0}

and higher quality scores are better. These conditional increments are not independent leave-one-out effects;

0.0

denotes equality at reported precision, not statistical equivalence.

Work	Optimised for	Method level	Direction drift?	State-slice approx. orthogonal?
Plan-and-Execute ( Wang et al., 2023 ; Xu et al., 2023 )	Task completion	Loop (planner+executor)	$\times$	$\times$
LIFE-HARNESS ( Xu et al., 2026 )	Task completion	Runtime	$\times$	$\times$
Workflow Graphs ( Park et al., 2025 )	Flow control	Flow structure	Partial	$\times$
Talker-Reasoner ( Christakopoulou et al., 2024 )	Latency	Dual-system	$\times$	$\times$
SLOT ( Shen et al., 2025 ) / GCD ( Raspanti et al., 2025 )	Output structure	Decoding	Partial	$\times$
FACTS Grounding ( Jacovi et al., 2025 )	Factuality	Evaluation	$\times$	$\times$

Appendix

Table 5: Comparison to prior work. SGC is orthogonal to and composable with each prior work: it targets a distinct failure class (direction drift) that prior work does not address. The Plan-and-Execute and LIFE-HARNESS rows jointly represent the method family of the empirical PE-Agent baseline used in Section 7 .

Figure 4: Schematic event sequences, not measured durations. (a) Naive 3-stage cascade: intent $\to$ planner $\to$ tools $\to$ reply, with the planner conditioning on stale state and first screen at $\approx\!3{,}000$ ms. (b) $\mathcal{W}_{1}$ streams NARR, INTENT, SUB_INTENT, then PLAN (fallback only); when $\Phi$ fires on sub-intent closure, tools and retrieval dispatch in parallel under TTS cover, and the early path leads PLAN by $\approx\!1.5$ s. Referenced from Section 4 .

Intent	Sub-intent	Frozen tool set $\hat{T}$
recommend	counter-pick	{lineup_search, counter_rank, dex_lookup}
recommend	lineup-build	{team_search, synergy_rank}
teaching	mechanics-qa	{doc_retrieve, rule_lookup}
teaching	skill-explain	{skill_dex, prereq_check}
knowledge-qa	entity-info	{dex_lookup}
knowledge-qa	meta-brief	{meta_report_retrieve}

Appendix

Table 6: Illustrative excerpt of the

\Phi

table. The full table (approximately 50 rows) will be released with the benchmark.

Rule id	Behaviour
R1 · hard-conflict eligibility	Exclude praise affirming a violated constraint; require constraint-account for that constraint.
R2 · selection mutex	Select at most one of the eligible alternatives that affirm and deny satisfaction of the same constraint.
R3 · soft-only eligibility	praise and apology may both be eligible candidates; R2 resolves incompatible alternatives at selection time.
R4 · hint-only eligibility	Enable scene-flavor ; an advisory hint alone does not require an apology about a missing entity.
R5 · intent eligibility	For knowledge-qa intent, guidance is disabled regardless of conflicts.
R6 · empty-inventory requirement	If $\sigma(s_{\text{live}}^{\text{fs}})=\emptyset$ , exclude praise asserting availability; require an apology explaining the empty inventory.

Appendix

Table 7: Paragraph eligibility and selection constraints. R2 applies after candidate eligibility; eligibility does not mean a paragraph has been selected.

Guidance type	Cooldown	$p_{c}$	$w_{c}$
lineup-recommend	3 turns	0.60	1.0
counter-tip	5 turns	0.45	0.8
mechanics-hint	2 turns	0.70	0.6
encourage	4 turns	0.30	0.4
empty-state-nudge	6 turns	0.20	0.3

Appendix

Table 8: Illustrative

\mathcal{W}_{3}

cooldown configuration. Values are representative production defaults, tuned quarterly against the guidance-repetition rate metric.

LLM Grounding

Large language model (LLM) grounding studies how models anchor answers and actions in evidence, structured knowledge, and real-world context.

See all 320 papers

Jun 15, 2026

ACCORD: Action-Conditioned Contextual Grounding for Language Agents

User instructions are often underspecified because humans rely on implicit assumptions about the surrounding environment. For large language model (LLM) agents operating in information-rich digital and physical environments, these assumptions cannot be inferred from the instruction alone; they must be recovered from the current state of tools, data, interfaces, and observations. Effective execution therefore requires agents to identify missing context, ground it in observed evidence, and carry it forward into subsequent actions. We show that current agents often fail to do so. They act from assumed rather than observed specifics, overlook information they could have gathered, and fail to incorporate evidence that has already been returned. Building on this insight, we propose ACCORD (Action-Conditioned Contextual Grounding), a simple and effective agent framework for adaptive grounding. Before each action, ACCORD actively probes the environment for missing information and integrates relevant context from the agent's trajectory that would otherwise be overlooked. Requiring no additional training or task-success signals, ACCORD improves task-goal completion on AppWorld by up to +20.6 points with GPT-5-mini, from 42.0% to 62.6%, compared to strong baselines. These gains persist with a substantially stronger base model (+10.8 with Claude-4.5-sonnet), an open-weight model (+10.1 with Qwen3.5-27B-FP8), and on the embodied AlfWorld benchmark (+7.4 success rate with GPT-5-mini).

Lai Jiang, Cheng Qian, Zhenhailong Wang +3

University of Illinois Urbana-Champaign · Stanford University

LLM Grounding Tool-Augmented Language Model Agents Active Information Gathering LLM Agents LLM Context Management

Ask PDF

Sep 26, 2026

When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents

LLM agents often operate over multi-turn interactions in which user intent changes before execution. We study intent drift: the failure mode in which superseded parts of the user's intent continue to influence the final answer or tool action. We introduce IntentFlux, an executable benchmark that converts verifiable tasks into dialogues with controlled intent changes while preserving their original graders. In a 627-case calibration, mean task score falls from 0.476 to 0.384 as dialogues contain more superseded and withdrawn information. Across eight models, the rate of fully correct solutions is significantly lower when the same final task must be recovered from an evolving dialogue rather than given directly in a single turn. We further introduce StateForge, which explicitly maintains the active requirements before generation. On General-Test, it improves mean task score from 0.367 to 0.467. Providing the ground-truth final state improves performance further but still does not recover single-turn performance, indicating that state-estimation errors explain only part of the gap. These results establish intent drift as a measurable multi-turn failure mode and explicit state maintenance as a partial mitigation.

Yanjie Zhang, Bowen Cao, Zixin Chen +1

HKUST, Hong Kong SAR, China · CUHK, Hong Kong SAR, China · LIGHTSPEED, Shenzhen, China

LLM Agents LLM Agent Evaluation LLM Agent Reliability AI Agent Benchmarks

Ask PDF

Jun 15, 2026

State-Grounded Multi-Agent Synthetic Data Generation for Tool-Augmented LLMs

Training tool-augmented LLM agents requires large corpora of multi-turn, tool-grounded conversational data that is expensive to annotate, privacy-constrained in production settings, and largely absent from public datasets. We present StateGen, a synthetic data generation platform that produces scored, reasoning-trace-rich training conversations by orchestrating a four-role LLM loop: a persona-conditioned user simulator, an agent under test, a state-grounded tool simulator, and a multi-axis LLM judge. The key architectural contribution is an authoritative state manager that maintains a structured world-state object across turns, enforcing a backend-is-truth invariant that eliminates the dominant class of tool-call hallucinations by construction. StateGen extends naturally to hierarchical multi-agent settings by declaring sub-agents as tools, all sharing a single state object. We report results on 64,698 evaluated conversations across three production corpora: tool-call hallucination scores reach 9.66/10, the system supports persona-driven variation via a 23-dimensional trait vector, and a cleanly separated train and golden evaluation set split confirms the data is not memorization bait (per-criterion gap analysis). Comparison with eight external systems shows that no single publicly available platform combines multi-turn generation, state-grounded tool simulation, hierarchical multi-agent support, and built-in judge scoring.

Rahul Khedar, Eshita, Sneha Teja Sree Reddy Thondapu +10

PayPal AI

Multi-Agent LLM Systems Tool-Augmented Language Model Agents LLM Tool Use LLM Agent Training Synthetic Data Generation for Language Models Synthetic Data Generation for LLM Agents

Ask PDF

Abstract

LLM Grounding

Explore similar work

ACCORD: Action-Conditioned Contextual Grounding for Language Agents

When Users Change Their Minds: Measuring and Repairing Intent Drift in LLM Agents

State-Grounded Multi-Agent Synthetic Data Generation for Tool-Augmented LLMs