💬 NLP

AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents

AURA introduces an intent-directed probing framework for situated LLM agents that inserts an inference step to generate structured IntentFrames with gap scores, significantly improving implicit-need coverage while reducing tool probes and eliminating privacy violations compared to standard ReAct-style approaches.

Original authors: Yang Li, Jiaxiang Liu, Jiang Cai, Mingkun Xu

Published 2026-06-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yang Li, Jiaxiang Liu, Jiang Cai, Mingkun Xu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking into a busy coffee shop and you ask a friend, "Where is Lin Wei?"

A standard AI assistant (like the ones we use today) acts like a very literal GPS. It looks at the map, sees Lin Wei is in the coffee shop, and says, "She is at the coffee shop." Then it stops. It assumes that's all you wanted to know.

But in real life, you probably didn't just want the location. You might have been wondering:

"Is she free to talk right now, or is she in a meeting?"
"Is she in a good mood, or should I wait?"
"Is it a good time to interrupt her?"

The paper introduces a new AI system called AURA (Intent-Directed Probing for Implicit-Need Surfacing). Think of AURA not as a GPS, but as a socially aware concierge.

The Problem: The "Literal" Trap

Most AI agents today are stuck in a loop where they only answer the exact words you type. If you ask "Where is Lin Wei?", they look for the location and stop. They miss the hidden reason you asked. They don't know that you actually need to know if she's available before you walk over.

The Solution: The "Concierge" Step

AURA inserts a special thinking step between "seeing the world" and "answering you."

The Scan (Perception): Like other agents, AURA sees the coffee shop and knows Lin Wei is there.
The "Gap" Check (The Magic Step): Before answering, AURA pauses and asks itself: "What is the gap between what the user literally asked and what they probably actually need?"
- It creates a score (a "Gap Score") from 0 to 1.
- Score 0: You just want the location. No extra work needed.
- Score 0.8: You asked for the location, but you likely need to know if she's busy, stressed, or in a meeting.
The Budget (The Wallet): Based on that score, AURA decides how much "detective work" to do.
- If the score is low, it saves money (computing power) and just gives the location.
- If the score is high, it gets a "budget" to ask specific questions behind the scenes, like checking Lin Wei's private calendar or mood, without you having to ask for them explicitly.
The Answer: It then gives you a complete answer: "Lin Wei is at the coffee shop, but she looks busy and tired, so maybe wait a bit."

The Analogy: The Detective vs. The Librarian

Old Agents (ReAct style): Are like a Librarian. You ask for a book title, they find the book, and hand it to you. They don't ask if you actually need the book for a school project or just to read for fun.
AURA: Is like a Detective. You ask, "Where is the suspect?" The Detective doesn't just give the address. They think, "Why is the client asking? Do they need to arrest him, or just talk to him?" If the client seems urgent, the Detective checks the suspect's alibi and mood before giving the address.

What the Paper Actually Found

The researchers tested this in a simulated town with 5 virtual people (AURATown). They asked 100 questions designed to hide the real need (e.g., asking "Where is X?" when the real need was "Is X free?").

Better at Reading Minds: AURA successfully figured out the hidden needs 7% more often than the standard "literal" agents. This is a statistically significant win.
Smart Spending: AURA didn't just ask everything all the time. It learned to ask fewer questions when the answer was obvious, saving time and resources.
Privacy: Because AURA only asks for private information (like mood or availability) when it thinks it's necessary, it accidentally became better at privacy. It didn't peek into people's private lives unless the "Gap Score" said it was needed.
Not a Magic Bullet for Facts: If you just ask a simple factual question like "What time is it?", AURA doesn't necessarily give a more accurate answer than a standard agent. Its superpower is specifically for social situations where you need to understand why someone is asking.

The Bottom Line

AURA teaches AI agents to stop being literal robots and start being intuitive partners. It adds a tiny "thinking pause" to figure out what you really need, so it can give you a helpful answer instead of just a technically correct one.

Note: The paper does not claim this works for medical diagnosis, legal advice, or real-world navigation yet. It was tested strictly in a simulated social environment.

Technical Summary: AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents

1. Problem Definition

The paper addresses a critical limitation in current LLM-based situated agents: the tendency to answer only the literal content of a user query while missing the implicit information need. In structured environments where private state (e.g., an agent's mood, availability, or unspoken goals) is hidden behind tool-mediated access, standard agents often fail to infer that a query like "Where is Lin Wei?" might actually mean "Is Lin Wei free to chat?" or "Is it appropriate to interrupt her?"

Existing approaches like ReAct interleave reasoning and tool use but only trigger tools when the surface query explicitly demands them. Plan-and-Solve pre-plans tools based on literal queries, and Generative Agents inject all environment state as passive context without control over which private states to surface. Consequently, agents make decisions with incomplete context, invent details instead of grounding responses in current state, and fail to translate environment state into the specific context a user needs for their next action.

2. Methodology: The AURA Framework

AURA introduces a control mechanism that inserts an inference step between scene perception and tool use. It operates in a "situated" regime where private state is partitioned and accessible only via active probing.

Core Architecture

The framework consists of two phases:

Deterministic Context Assembly: A passive phase (Sense → Scene → Memory) that aggregates public context, similar to Generative Agents.
LLM-Controlled Reasoning: The novel contribution, which includes:
- IntentInferrer: An LLM-mediated stage that generates an IntentFrame.
- Explore: A bounded proactive probing loop that executes tool calls based on the IntentFrame's directives.
- Reason/Act/Interact: The final response generation and interaction.

The IntentFrame

The core innovation is the IntentFrame, a structured output generated by the IntentInferrer before any tool is called. It contains:

$\ell$ (Literal Need): Restatement of the surface query.
$I$ (Implicit Needs): A list of plausible hidden needs (e.g., availability, mood).
$g$ (Gap Score): A scalar value in $[0, 1]$ representing the divergence between the literal answer and the implicit need.
$P$ (Recommended Probes): A subset of candidate tools to investigate.
$a$ (Alert Flag): Indicates if a heads-up alert is needed.
$c$ (Confidence): Self-reported confidence.
$r$ (Rationale): Explanation for the inference.

Bounded Proactive Probing

The Explore stage uses the gap score $g$ to determine a probe budget ceiling $B(g)$ .

The mapping is deterministic: $g \in [0, 0.2) \to B=0$ , up to $g \in [0.8, 1] \to B=5$ .
Crucially, $B(g)$ is a ceiling, not a target. The LLM planner decides when to stop probing once actionable information is gathered, often issuing fewer probes than the ceiling allows.
This loop is goal-directed, bounded (to control API costs), and composable (results update context for subsequent decisions).

3. Key Contributions

IntentFrame as a Control Variable: The paper introduces the IntentFrame as a pre-tool routing decision that factors out gap estimation. This allows the agent to determine which private-state probes to issue and how many before generating a response.
Empirical Validation on Implicit-Need Benchmark: On a 100-query, four-scene benchmark, AURA significantly improves implicit-need coverage compared to ReAct-style probing ( $\Delta = +0.07$ , $p < 10^{-6}$ ). Three of the four scenes showed individual significance.
Mechanism Characterization: The authors demonstrate that the performance gain is driven by gap calibration via the LLM, not answer memorization. Ablation studies show that removing the few-shot examples collapses the gain, and replacing the LLM IntentInferrer with a heuristic drops performance from 0.803 to 0.368.
Access-Cost Pareto Frontier: In factual grounding tasks, AURA does not necessarily maximize raw accuracy but achieves a Pareto optimal point: it reduces probe counts by 82% compared to fixed-probe baselines and achieves 0% forbidden-tool violations on privacy-sensitive slices, whereas other methods frequently violate privacy constraints.
Open Resources: The release of the AURATown simulator, the 100-query benchmark with subcategory labels (inter-annotator agreement $\kappa=0.61$ ), and per-seed run records.

4. Experimental Results

The evaluation uses gpt-4o-mini as the backbone and judge, with cross-backbone robustness checks on claude-haiku-4.5 and qwen-plus.

Implicit-Need Surfacing: AURA-Intent achieved an implicit-need coverage score of 0.804, significantly outperforming NoIntent (ReAct-style) at 0.733 and Literal at 0.223. The largest gains were observed in "availability" queries where the surface form fully masks the implicit need.
Factual Grounding: In factual lookup tasks, AURA (GapRouted) achieved a factual accuracy (FA) of 0.696 with an average of 1.40 probes per query. This is lower than the Fixed-Probe baseline (FA 0.766, 8.00 probes) but represents a massive efficiency gain.
Privacy: On a privacy-sensitive slice, GapRouted reduced forbidden-tool violations to 0%, compared to 78.9% for Plan-and-Solve and 25.6% for ReAct.
Human Evaluation: Independent raters scored AURA significantly higher than Vanilla LLMs across environmental awareness, response helpfulness, agent believability, and factual accuracy.

5. Significance and Claims

The paper positions AURA not as a universal accuracy booster, but as a specialized mechanism for situational agents operating in environments with hidden private states.

Scope: The authors explicitly state that AURA targets "situated regimes with tool-mediated hidden state." It is not designed for open-world exploration or embodied navigation.
Trade-off: The mechanism is an access-cost Pareto point. It trades raw factual accuracy for significant reductions in probe costs and privacy violations in factual scenarios, while providing the primary accuracy lift in implicit-need scenarios.
Control Point: The paper argues that intent-directed probing is an LLM-prompted operation at a specific control point (the gap estimation), rather than an emergent property of the entire pipeline.
Limitations: The authors acknowledge that the few-shot examples are "load-bearing" for gap calibration. They also note that the current gap-to-budget map is a hand-tuned step function, suggesting that learning this mapping from interaction logs could be a future improvement. The framework does not automatically transfer to external belief benchmarks (e.g., FANToM) without specific adaptation.

In summary, AURA provides a structured approach to bridging the gap between what a user asks and what they actually need to know in complex, multi-agent environments, enabling agents to proactively surface hidden state without indiscriminate or privacy-violating probing.