AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/innovation64/AURA
Area: LLM Reasoning
Keywords: Situated agents, Implicit intent inference, Intent-directed probing, Structured environment state, Interactive decision making
TL;DR¶
AURA introduces an IntentFrame inference stage between scene perception and tool execution to estimate the gap between literal queries and underlying implicit needs, dynamically regulating probe budgets to improve implicit-need coverage by 7.1% over ReAct-style probing while cutting probe calls by 82% with zero forbidden-tool violations on privacy-sensitive factual tasks.
Background & Motivation¶
In situated multi-agent simulations and real-world human-agent collaborative environments, natural language queries directed at agents are inherently pragmatically compressed. When a human user asks an assistant agent "Where is Lin Wei?", the surface form requests an objective spatial location (e.g., "in the cafe"), yet the genuine, unstated social need almost always concerns whether Lin Wei is available, whether she is immersed in deep focus or open to conversation, and whether interrupting her right now is appropriate. Conventional tool-integrated LLM agent architectures—ranging from reactive frameworks like ReAct to proactive social frameworks like Generative Agents—fundamentally treat the surface query as if it fully specified the user's information need. ReAct only fires tools reactively when answer generation encounters an explicit blocker; Plan-and-Solve drafts a static execution plan strictly grounded in the literal phrasing; and Generative Agents passively dumps full environment snapshots into the prompt without any selective control over which private states should be surfaced. As a consequence, agents either terminate abruptly after emitting the literal coordinates or hallucinate the target agent's internal state due to missing context.
The root cause of this failure lies in the structural asymmetry between public and private environmental states, combined with the pragmatic blindness of existing tool controllers. In structured situated environments, external facts such as physical coordinates and visible activities constitute public states that can be passively gathered via scene previews, whereas emotional states, hidden agendas, availability flags, and stale beliefs about others are private states sequestered behind access-controlled tool registries. Existing architectures lack an upfront control variable to ask "what does the user actually need to know?" and "how much probing budget should be allocated to bridge that information gap?". Simply concatenating all private states into a monolithic prompt suffers from context dilution, incurs costs that scale linearly with agent count, and catastrophically breaches privacy constraints by exposing confidential personal records.
This paper tackles the challenge by decoupling implicit-need estimation from reactive tool execution, inserting a dedicated intent inference stage before any tool is triggered. Core idea: formalize implicit intent inference as a pre-tool control variable termed IntentFrame, which produces an LLM-estimated scalar gap score measuring the divergence between literal and implicit needs to deterministically set an adaptive probing ceiling, enabling targeted, budgeted exploration of private environment states while enforcing strict privacy and access-cost boundaries.
Method¶
Overall Architecture¶
AURA structures situated agent interaction into two distinct pipelines: deterministic context assembly (Sense → Scene → Memory) followed by LLM-controlled active reasoning (IntentInferrer → Explore → Reason → Act → Interact). Upon receiving a natural language query \(q\), the deterministic stage compiles a lightweight public scene preview \(S\) and retrieves pertinent recent memories \(M\). The core contribution, the IntentInferrer, then processes this preview to generate a structured IntentFrame comprising a literal restatement, a set of inferred implicit needs, a scalar gap score \(g \in [0, 1]\), a recommended tool shortlist \(P\), and a heads-up alert flag \(a\). Downstream, the Explore module translates \(g\) into a deterministic probing budget ceiling \(B(g)\) and executes goal-directed environment probing over a registered tool suite, halting early once decisive evidence is secured. Finally, Reason synthesizes the compact probe trace with the initial context, while Act and Interact deliver the response, optionally prepending a heads-up warning when significant private constraints are surfaced.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
Q["User Query q"] --> DET["Deterministic Context Assembly<br/>Sense → Scene → Memory"]
DET --> D1["1. IntentFrame Inference<br/>Infer implicit needs & emit scalar gap score g"]
D1 --> D2["2. Bounded Proactive Probing<br/>Enforce budget ceiling B(g) with early stopping"]
D2 --> D3["3. Structured Public/Private State Decoupling<br/>Public scene preview vs. tool-gated private state"]
D3 --> D4["4. Evidence-Grounded Response Synthesis<br/>Synthesize probe trace & emit heads-up alert"]
D4 --> ANS["Final User Response & Agent Action"]
Key Designs¶
1. IntentFrame Inference: Estimating Unstated Gaps Prior to Execution
Conventional tool-augmented agents plan actions solely from surface syntax, remaining blind to pragmatic context. The IntentInferrer bridges this gap by leveraging a few clean, benchmark-disjoint few-shot exemplars and a four-tier gap scoring rubric to parse the query \(q\), scene snapshot \(S\), and memory \(M\) into a typed IntentFrame: $\(\text{IntentFrame} = (\ell, I, g \in [0, 1], P \subseteq \mathcal{T}, a \in \{0, 1\}, c \in [0, 1], r)\)$ where \(\ell\) denotes the literal-need restatement, \(I\) lists plausible implicit needs (e.g., mapping "Where is Lin Wei?" to "Is she free to chat?"), \(g\) quantifies the divergence between literal and implicit needs (\(0\) indicates literal sufficiency, while \(1\) denotes complete orthogonality), \(P\) specifies recommended probe tools from registry \(\mathcal{T}\), \(a\) is a binary alert flag, \(c\) is model confidence, and \(r\) provides natural-language rationale. By materializing this frame before any environment interaction, the agent establishes an interpretable, transparent control signal for all downstream probing.
2. Bounded Proactive Probing: Dynamic Budget Allocation with Early Stopping
Executing a uniform number of tool probes across all queries wastes API latency and inflates hallucination risks on trivial lookups. AURA implements a deterministic ceiling mapping function \(B(g)\) that translates the continuous gap into an integer step limit:
$\(B(g) \in \{0, 1, 2, 3, 5\} \quad \text{for} \quad g \in [0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), [0.8, 1.0]\)$
Crucially, \(B(g)\) operates strictly as a ceiling rather than an obligatory quota. Within the Explore loop, the LLM planner inspects the evolving probe trace at each step. As soon as a single high-value probe (such as querying private status) returns actionable confirmation (e.g., availability=busy), the planner emits an early stopping signal. Consequently, queries requiring deep social reconnaissance utilize the necessary budget, while straightforward factual queries receive \(B=0\), entirely bypassing the tool execution pipeline.
3. Structured Public/Private State Decoupling: Access-Controlled State Separation
Rather than treating the entire simulation as a flat context buffer, AURA enforces a rigorous boundary between publicly observable and privately held states within its AURATOWN environment. Visible attributes like spatial coordinates and overt actions are assembled passively by Scene at negligible prompt overhead (~180 tokens). Conversely, sensitive attributes including emotional state, latent goals, availability, and beliefs about peers are strictly gated behind tool registries. The system maintains two disjoint registries: an 8-tool environment suite for factual lookup and a 5-tool state suite (e.g., get_agent_private_state, get_agent_belief_about) for implicit intent inference. Combined with deterministic state evolution rules (e.g., entering a crowded venue triggers a transition to busy), this architecture ensures that probed facts represent verifiable, non-hallucinated ground truth.
4. Evidence-Grounded Response Synthesis: Heads-up Alerting and Stale-Belief Grounding
Once the Explore loop concludes with a summarized probe trace, Reason synthesizes the collected evidence with the literal query. When the IntentFrame's alert flag is active (\(a=1\)) and the gap exceeds critical thresholds, Interact prefixes the output with a proactive [heads-up] notice—for instance, acknowledging that Lin Wei is in the cafe but advising the user that she is overwhelmed by peak rush hours. Furthermore, on second-order theory-of-mind queries where an agent's internal belief diverges from objective world reality due to lack of co-location, the generation prompt strictly enforces reporting the believer's recorded belief rather than the simulator's global ground truth, preventing the leakage of omniscient simulator states.
Key Experimental Results¶
Main Results¶
Evaluation was conducted on the AURATOWN benchmark (a 60×60 grid simulation featuring 5 named agents and 20 locations over a 6:00–23:00 timeline). The core benchmark evaluates 100 implicit-intent queries across 4 distinct scenes and 5 pragmatic subcategories (availability, mood, appropriateness, latent goal, second-order belief), supplemented by 50 factual-grounding queries across 5 functional categories (spatial, social, temporal, memory, planning). Primary experiments use gpt-4o-mini across 3 random seeds (42, 123, 456).
| Benchmark / Evaluation Slice | Metric | AURA Intent / GapRouted | ReAct (NoIntent) | Plan-and-Solve | Gain / Significance |
|---|---|---|---|---|---|
| 100-Query Four-Scene Benchmark | Implicit Coverage (Overall) | 0.804 | 0.733 | — | \(\Delta=+0.071\) (\(p=1.0 \times 10^{-6}\)) |
| Scene A: cafe (25 queries) | Implicit-Need Coverage | 0.800 | 0.709 | — | \(\Delta=+0.091\) (\(p=0.004\)) |
| Scene B: library (25 queries) | Implicit-Need Coverage | 0.779 | 0.688 | — | \(\Delta=+0.091\) (\(p=1.4 \times 10^{-4}\)) |
| Scene C: garden (25 queries) | Implicit-Need Coverage | 0.811 | 0.731 | — | \(\Delta=+0.080\) (\(p=0.015\)) |
| Scene D: night (25 queries) | Implicit-Need Coverage | 0.827 | 0.803 | — | \(\Delta=+0.024\) (\(p=0.349\), NS) |
| 25-Query Pilot Set (Scene A) | Implicit-Need Coverage | 0.803 | 0.640 | 0.520 | vs. ReAct \(+0.163\) (\(p=0.006\)) |
| 50-Query Factual Grounding | Factual Accuracy (FA) | 0.696 ± 0.026 | 0.550 ± 0.037 | 0.764 ± 0.020 | Access Pareto Frontier |
| 50-Query Factual Grounding | Probes per Query | 1.40 | — (FC bypass) | 4.64 | 82.5% fewer than Fixed-Probe |
| 30-Query Privacy Slice | Forbidden Tool Violation % | 0.0% | 25.9% | 78.9% | Absolute Zero Violations |
Ablation Study¶
A component ablation on the 100-query benchmark (\(n=300\) query-seed cells) isolates the impact of the budget ceiling versus the tool recommendation shortlist, alongside prompt calibration and backend comparisons on the 25-query pilot set.
| Configuration | Implicit Score | Mean Probes | \(\Delta\) vs. NoIntent | \(p\)-value | Notes |
|---|---|---|---|---|---|
| AURA Full (Ceiling + Hints) | 0.804 | 1.75 | +0.071 | \(6.4 \times 10^{-7}\) | Complete dual-signal system |
| Budget Ceiling Only | 0.827 | 1.20 | +0.095 | \(6.9 \times 10^{-11}\) | Highest accuracy & lowest cost |
| Tool Hints Only | 0.813 | 1.74 | +0.080 | \(5.6 \times 10^{-8}\) | Lacks dynamic budget scaling |
| NoIntent (ReAct-style) | 0.733 | 1.11 | — | — | Baseline without intent routing |
| Clean few-shot prompt | 0.803 | 1.48 | +0.163 | 0.006 | Disjoint exemplars calibrate gap |
| No few-shot prompt | 0.677 | — | +0.035 | 0.44 (NS) | Mean gap collapses to 0.476 |
| Deterministic Heuristic Backend | 0.368 | 0.32 | — | — | Fails on lexically decoupled queries |
Key Findings¶
- The budget ceiling alone drives the gain: Isolating the two downstream signals emitted by the IntentFrame reveals that pairing the gap-derived budget ceiling \(B(g)\) with a fixed, deterministic tool order achieves an implicit coverage of 0.827, outperforming the full architecture (0.804, \(p=0.021\)) while requiring 31% fewer probes. The LLM tool shortlist frequently misleads the agent into invoking plausible but weaker tools (e.g.,
get_nearby_agentsrather thanget_agent_private_state), squandering the probe budget. - The gap score functions as a discrete trigger, not a calibrated value estimator: Across 300 paired cells, the per-query probing benefit correlates negatively with the inferred gap (\(r=-0.33\)). The inferrer assigns moderate gap scores (0.48) to surface-decoupled availability queries where probing yields massive gains (+0.29), while assigning high gap scores (0.80) to second-order belief queries where even standard baselines naturally trigger belief retrieval. Thus, the gap's utility lies entirely in deciding whether to cross the zero-budget threshold, not in predicting marginal information value.
- Pareto optimality in privacy and access cost: On factual lookup, AURA GapRouted achieves 0.696 FA using only 1.40 probes per query compared to 8.00 probes for Fixed-Probe (0.766 FA). More critically, on the 30-query privacy-sensitive slice where forbidden tools expose personal plans and memories, AURA assigns \(B=0\) to low-gap factual queries, structurally guaranteeing a 0.0% forbidden-tool violation rate, whereas Plan-and-Solve violates constraints on 78.9% of queries.
Highlights & Insights¶
- Pragmatic intent modeling as an upfront control gate: By transitioning intent inference from an implicit byproduct of token generation to an explicit pre-tool control variable, AURA provides a principled method for situated agents to anticipate unstated user needs.
- The lean deployment blueprint: The empirical finding that a simple gap-derived budget ceiling paired with a fixed tool sequence dominates complex LLM tool shortlists offers a crucial design insight for production agent systems, minimizing latency and failure modes.
- Honest delineation of technical scope: The authors comprehensively demonstrate the boundary conditions of their method, showing through rigorous negative transfers on FANToM, LoCoMo, and GAIA that proactive probing is beneficial strictly when task-relevant private states are hidden behind structured, verifiable tool interfaces.
Limitations & Future Work¶
- Strict structural preconditions: The architecture's gains vanish when environmental states are already fully present within the context window (as in narrative ToM benchmarks like FANToM) or when tool returns are noisy and ungrounded (such as web search in GAIA).
- Backbone-judge circularity: Primary evaluations rely on
gpt-4o-miniboth as the agent backbone and the evaluation judge; while cross-backbone testing on Claude-3.5-Haiku and Qwen-Plus confirms robustness, fully independent human and automated judges are needed. - Heuristic budget mapping: The piecewise mapping from scalar gap \(g\) to budget \(B(g)\) is currently hand-tuned; learning an end-to-end value-of-information policy from interaction logs represents the natural next step.
Related Work & Insights¶
- vs. ReAct (Yao et al., 2023): ReAct interleaves reasoning and tool use reactively during generation, firing tools only when literal syntax demands them; AURA predicts unstated needs upfront, ensuring private states are probed before answering.
- vs. Generative Agents (Park et al., 2023): Generative Agents injects all world observations into context passively under an all-public assumption; AURA enforces a public/private separation and leverages the IntentFrame for budgeted, access-controlled probing.
- vs. Plan-and-Solve (Wang et al., 2023): Plan-and-Solve generates static plans directly from surface queries, frequently violating privacy constraints on factual queries; AURA uses the scalar gap to suppress tools entirely when literal facts suffice.
Rating¶
- Novelty: ⭐⭐⭐⭐ [Transforms linguistic pragmatics into a concrete, pre-tool budget control variable for situated agents]
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive multi-seed benchmarks, component ablations, strict rescoring, and cross-domain boundary tests]
- Writing Quality: ⭐⭐⭐⭐⭐ [Exemplary scholarly clarity, transparent reporting of negative results, and rigorous statistical grounding]
- Value: ⭐⭐⭐⭐ [Provides practical architectural and privacy guidance for situated assistants and interactive multi-agent systems]