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

# Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents
Source: [https://arxiv.org/html/2606.05557](https://arxiv.org/html/2606.05557)
Yang Li Jiaxiang Liu Jiang Cai Mingkun Xu Guangdong Institute of Intelligence Science and Technology \{liyang, liujiaxiang, caijiang, xumingkun\}@gdiist\.cn

###### Abstract

A situated query like “where is Lin Wei?” often encodes more than its literal content: the user may also want to know whether Lin Wei is free, in a good mood, or worth interrupting now\. Standard tool\-use agents answer the literal question and stop\.AURAinserts an inference step between scene perception and tool use that produces anIntentFrame: a structured estimate of the implicit need with a scalar*gap*score that controls per\-query probe budget and tool selection\. On a 100\-query four\-scene implicit\-intent benchmark,AURAimproves implicit\-need coverage over ReAct\-style probing \(𝚫=\+0\.07\\boldsymbol\{\\Delta\{=\}\{\+\}0\.07\},𝒑<𝟏𝟎−𝟔\\boldsymbol\{p\{<\}10^\{\-6\}\}\); three of four scenes are individually significant, the gain reproduces on a second backbone, and a prompt ablation attributes the lift to gap calibration rather than answer memorisation\. On factual lookup the controller trades raw accuracy for𝟖𝟐%\\mathbf\{82\\%\}fewer probes and zero forbidden\-tool violations on a privacy\-sensitive slice; scope conditions are detailed in Limitations\. Code, simulator, and benchmark are released at[https://github\.com/innovation64/AURA](https://github.com/innovation64/AURA)\.

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

Yang Li Jiaxiang Liu Jiang Cai Mingkun Xu††thanks:Corresponding author\.Guangdong Institute of Intelligence Science and Technology\{liyang, liujiaxiang, caijiang, xumingkun\}@gdiist\.cn

## 1Introduction

LLM\-based agents\(Xiet al\.,[2025](https://arxiv.org/html/2606.05557#bib.bib12); Wanget al\.,[2023a](https://arxiv.org/html/2606.05557#bib.bib18)\)have been deployed in social simulations\(Parket al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib1)\)and software teams\(Honget al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib13)\), but they often answer the user’s*literal*question about an environment while missing the*implicit*information need behind it\. Asking “where is Lin Wei?” may be a request for location, but it may also mean “is she free to chat?”\. Three downstream problems follow: agents make decisions with incomplete context; user\-facing responses invent details instead of grounding in current state; and environment state is rarely translated into the specific context a user needs for the next action\.

Prior approaches answer this only partially\.ReAct\(Yaoet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib4)\)interleaves reasoning and tool calls during answer generation, but tools fire only when the surface query explicitly demands them — the loop has no step that asks “what does the user actually want to know?”\.Plan\-and\-Solve\(Wanget al\.,[2023b](https://arxiv.org/html/2606.05557#bib.bib58)\)pre\-plans tool calls from the literal query, with no mechanism to bridge to an implicit need that the literal query does not name\.Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib1)\)inject all environment state as passive context but expose no control over which private state to surface for a given query\. Each treats the user’s surface query as if its literal form fully specified what the user wants\.AURA’s contribution is the missing step: making implicit\-need inference an*independent control variable*for tool use\. Rather than appending another LLM call to a ReAct loop,AURAfactors out gap estimation as a pre\-tool routing decision that determines which private\-state probes to issue and how many\.

By*situated*we mean structured environments where private state is partitioned behind tool\-mediated access—the agent can observe public context passively but must actively probe for hidden state\. This is narrower than the general “situated agent” umbrella, which includes embodied navigation, open\-world exploration, and real\-time sensor fusion; we do not address those settings\. We study this problem as implicit\-intent inference over structured environment state\.AURAintroduces theEnvironment Agent: a pipeline \(Perceive→\\toScene→\\toMemory→\\toReason\) plus an LLM\-generatedIntentFramethat estimates the gap between a user’s literal query and plausible implicit need, then uses that estimate to direct per\-query probing, tool selection, and optional heads\-up alerts\.

Contributions\.\(i\)We introduce theIntentFrameas a*pre\-tool control variable*: before any tool fires, the agent infers the user’s implicit need and emits a scalar gap score that determines the probe budget and shortlists candidate tools\.\(ii\)On a 100\-query four\-scene implicit\-intent benchmark, gap\-routed probing improves implicit\-need coverage over ReAct\-style NoIntent \(Δ=\+0\.07\\Delta\{=\}\{\+\}0\.07,p<10−6p\{<\}10^\{\-6\}; three of four scenes significant\), reproducing on a 25\-query pilot, a second backbone, and under disjoint\-example ablation\.\(iii\)We characterise the mechanism’s regime boundary: the controller is an access–cost Pareto point on factual grounding, not a universal accuracy win \(Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1); Limitations\)\.\(iv\)We release theAURATownsimulator, all 100 implicit\-intent queries with subcategory labels \(κ=0\.61\\kappa\{=\}0\.61inter\-annotator agreement\), and per\-seed run records\.

## 2Related Work

#### Agent architectures and tool use\.

ReAct, Reflexion, Toolformer, AgentBench, ToolLLM, T\-Eval\(Yaoet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib4); Shinnet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib7); Schicket al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib5); Liuet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib11); Qinet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib36); Chenet al\.,[2024b](https://arxiv.org/html/2606.05557#bib.bib35)\)treat tool use as reactive: tools fire*during*reasoning\.AURA’s Explore runs*before*reasoning with a bounded budget routed by the inferred literal\-vs\.\-implicit gap\. Plan\-and\-Solve\(Wanget al\.,[2023b](https://arxiv.org/html/2606.05557#bib.bib58)\)plans tool calls from the surface query without that gap routing\.

#### Multi\-agent simulation and memory\.

Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib1)\)inject environmental observations into LLM prompts;AURAadds \(i\) per\-query access*control*by an LLM\-producedIntentFrame, vs\. passive top\-kkretrieval; \(ii\) a public/private state split that supplies the substrate for implicit\-intent and belief\-state evaluations Park’s all\-public state cannot\. SOTOPIA\(Zhouet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib10)\)provides the 7\-dimension framework we adopt; broader multi\-agent benchmarks\(Zhouet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib10); Zhuet al\.,[2025](https://arxiv.org/html/2606.05557#bib.bib33); Xuet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib34); Vezhnevetset al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib32); Chenet al\.,[2024a](https://arxiv.org/html/2606.05557#bib.bib23); Liet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib20); Honget al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib13)\)treat the environment as serving agents on request\. Memory frameworks\(Packeret al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib2); Zhonget al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib27); Lewiset al\.,[2020](https://arxiv.org/html/2606.05557#bib.bib16); Gaoet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib15); Liet al\.,[2026](https://arxiv.org/html/2606.05557#bib.bib59)\)target long\-term storage;AURA’s memory is populated by proactive probing, not passive accumulation\. Proactive context agents \(ContextAgent, ProAgent, PROBE, ProAgentBench\(Yanget al\.,[2026](https://arxiv.org/html/2606.05557#bib.bib43),[2025](https://arxiv.org/html/2606.05557#bib.bib44); Pasternaket al\.,[2025](https://arxiv.org/html/2606.05557#bib.bib49); Tanget al\.,[2026](https://arxiv.org/html/2606.05557#bib.bib46)\); full comparison in Appendix[A7](https://arxiv.org/html/2606.05557#A7.SS0.SSS0.Px4)\) target*when to assist*; concurrent industry work\(Thinking Machines Lab,[2026](https://arxiv.org/html/2606.05557#bib.bib61)\)argues for treating interactivity as a native multimodal capability with continuous micro\-turn perception rather than turn\-bounded prompting;AURAis orthogonal to both axes, targeting*what private\-state context to surface*for an already\-received query\.

#### Pragmatic and conditional QA\.

A parallel line of work targets the same motivation in text\-only QA: users omit context they assume the model already shares\.Liet al\.\([2025](https://arxiv.org/html/2606.05557#bib.bib62)\)introduceCondAmbigQA, a 2,000\-query benchmark with condition\-aware reasoning that improves QA accuracy by11\.75%11\.75\\%\.AURAaddresses the same gap in a different regime: the implicit need lives in another agent’s hidden private state rather than in textual context, and the resolution mechanism is a budgeted probe over a structured environment registry rather than a textual condition rewrite\.

#### Belief\-state evaluations\.

Theory of mind in LLMs is contested:Kosinski \([2023](https://arxiv.org/html/2606.05557#bib.bib52)\)argues spontaneous emergence;Sapet al\.\([2022](https://arxiv.org/html/2606.05557#bib.bib53)\); Ullman \([2023](https://arxiv.org/html/2606.05557#bib.bib55)\)show failure on trivially\-altered Sally–Anne tasks;Sclaret al\.\([2023](https://arxiv.org/html/2606.05557#bib.bib54)\)build an explicit symbolic tracker\.AURAis*not*a ToM benchmark: of our five implicit\-intent subcategories, only*second\_order*probes belief tracking \(stale\-belief templates adapted fromUllman \([2023](https://arxiv.org/html/2606.05557#bib.bib55)\)\); the other four \(*availability, mood, appropriateness, latent\_goal*\) target surfacing of non\-belief private state through a budgeted controller\. We positionAURAas evidence about*when intent\-directed probing helps surface hidden state*; belief queries appear in our 5\-subcategory taxonomy as one substrate \(*second\_order*\)\. Transfer to external belief benchmarks does not hold automatically: Appendix[A20](https://arxiv.org/html/2606.05557#A20)reports a FANToM transfer null\.

## 3TheAURAFramework

AURAfactors a situated agent into two phases \(Fig\.[1](https://arxiv.org/html/2606.05557#S3.F1)\):*deterministic context assembly*\(Sense→\\toScene→\\toMemory\) followed by*LLM\-controlled reasoning*\(IntentInferrer→\\toExplore→\\toReason→\\toAct→\\toInteract\)\. The deterministic phase establishes a passive context preview using the same passive\-perception pattern as Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib1)\)and is not the contribution\. The contribution is theIntentInferrer, which produces anIntentFramewhose scalar gap field controls per\-query probe budget and tool selection in Explore \(Section[3\.3](https://arxiv.org/html/2606.05557#S3.SS3)\); Reason, Act, and Interact then plan and emit the response, with an optional heads\-up alert when the inferrer’s gap value crosses a threshold\. The remainder of this section formalises the problem \(Section[3\.1](https://arxiv.org/html/2606.05557#S3.SS1)\), describes the bounded\-probing loop \(Section[3\.2](https://arxiv.org/html/2606.05557#S3.SS2)\), and details theIntentFramestage \(Section[3\.3](https://arxiv.org/html/2606.05557#S3.SS3)\)\.

![Refer to caption](https://arxiv.org/html/2606.05557v1/figures/aura-architecture.png)Figure 1:AURApipeline\. Sense receives the user query; Scene and Memory assemble a deterministic context preview\. TheIntentInferrer\(⋆\\star\) outputs anIntentFramewhose scalargapfield routes probe budget and tool selection in Explore \(Section[3\.3](https://arxiv.org/html/2606.05557#S3.SS3)\)\. Reason plans over the enriched context; Act and Interact emit the response\.### 3\.1Problem Formulation and Architecture

We consider a situated multi\-agent system whereNNagents operate in a shared environmentℰ\\mathcal\{E\}over discrete time steps\. A human userhhmay interact with any agent via natural\-language queriesqq\. We seek to maximise*environmental grounding*of agent actions and user\-facing responses:

maxπE⁡𝔼​\[∑tG​\(αit,ℰt\)\+∑qR​\(rq,ℰqt\)\],\\max\_\{\\pi\_\{E\}\}\\mathbb\{E\}\\\!\\left\[\\sum\_\{t\}G\(\\alpha\_\{i\}^\{t\},\\mathcal\{E\}^\{t\}\)\+\\sum\_\{q\}R\(r\_\{q\},\\mathcal\{E\}^\{t\}\_\{q\}\)\\right\],whereGGmeasures action\-state consistency,RRmeasures response factual accuracy, andπE\\pi\_\{E\}is the Environment Agent’s probing policy\.

The Explore stage’s useful range is bounded by the residual uncertainty after passive perception: when Sense/Scene/Memory cover the facts the query needs, extra probes add little; when the query depends on a remote agent’s private state, one targeted probe can change the answer\.

### 3\.2Bounded Proactive Probing

TheExplorestage implements bounded environment probing\. Unlike ReAct\(Yaoet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib4)\)where tool calls are interleaved*during*reasoning, our probing occurs*before*answer generation, so the final response is conditioned on a compact probe trace\. The probe loop iterates up to a budgetBBsteps; at each step an LLM plannerϕ\\phiexamines the current context and the tool registry𝒯=\{t1,…,tK\}\\mathcal\{T\}\{=\}\\\{t\_\{1\},\\dots,t\_\{K\}\\\}and either issues a tool call \(whose result is appended to a probe trace and used to update the context\) or signalsstop; the trace is summarised at the end and consumed by Reason\. Pseudocode in Algorithm[1](https://arxiv.org/html/2606.05557#alg1)\(Appendix[A1](https://arxiv.org/html/2606.05557#A1)\)\.

#### Design\.

The loop is goal\-directed \(LLM\-driven tool selection given current context, not exhaustively\), bounded \(step budgetBBcaps API cost\), composable \(each tool result updates context for the next decision, enabling multi\-hop gathering\), and separable \(probing runs independently from reasoning, so results can be cached or skipped\)\. TheAURATowninstantiation registers eight base environment tools \(Table[8](https://arxiv.org/html/2606.05557#A5.T8)\) for the factual\-grounding benchmark, and a separate five\-tool scripted registry of agent\-state probes for the implicit\-intent benchmark \(Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\); both register under a pattern\-based allow/deny policy\.

### 3\.3Intent Inference: Modeling the User’s Implicit Need

Proactive probing decides*what to ask the environment*; intent inference estimates*what useful answer the user may need*\.AURAfactors the second decision out as an explicit LLM\-mediated stage that sits between Scene/Memory and the Explore budget selector\.

Given the user’s surface queryqq, the preview scene𝒮\\mathcal\{S\}, and recent memoriesMM, theIntentInferrerproduces anIntentFrame:

IntentFrame=\\displaystyle\\texttt\{IntentFrame\}=\\,\(ℓ,I,g∈\[0,1\],P⊆𝒯,\\displaystyle\(\\ell,\\;I,\\;g\\in\[0,1\],\\;P\\subseteq\\mathcal\{T\},a∈\{0,1\},c∈\[0,1\],r\)\\displaystyle\\;a\\in\\\{0,1\\\},\\;c\\in\[0,1\],\\;r\)whereℓ\\ellis the literal\-need restatement,IIis the list of plausible implicit needs,ggis the*gap*between literal and implicit \(0 if the literal answer suffices, 1 if the user’s real need is orthogonal\),PPis a set of recommended probe tools drawn from the registry𝒯\\mathcal\{T\},aais an alert flag,ccis the inferrer’s self\-reported confidence, andrra rationale\.

The gapggis the control input: downstream stages use a deterministic mapB​\(g\)=0,1,2,3,5B\(g\)=0,1,2,3,5forg∈\[0,0\.2\),\[0\.2,0\.4\),\[0\.4,0\.6\),\[0\.6,0\.8\),\[0\.8,1\]g\\in\[0,0\.2\),\[0\.2,0\.4\),\[0\.4,0\.6\),\[0\.6,0\.8\),\[0\.8,1\]respectively, truncated by the globalexplore\_max\_stepsbudget\. Importantly,B​\(g\)B\(g\)is a*ceiling*, not a target: the downstream Explore loop receives both the probe budget and therecommended\_probeshint, and we observe in Section[A2\.1](https://arxiv.org/html/2606.05557#A2.SS1)that the LLM typically stops short of the ceiling when one well\-targeted probe has already returned actionable information\.

#### Worked example\.

For the query “where is Lin Wei?” at 14:30 with Lin Wei present in the cafe scene, the inferrer outputsℓ=\\ell\{=\}“locate Lin Wei”,I=I\{=\}\{“is she free to chat?”, “is she taking a break?”\},g=0\.6g\{=\}0\.6,P=P\{=\}\{get\_agent\_private\_state,get\_agent\_plan\},a=1a\{=\}1,c=0\.7c\{=\}0\.7\. Theg=0\.6g\{=\}0\.6maps to a budget ceiling ofB=3B\{=\}3, but Explore stops after two probes whenget\_agent\_private\_state\(‘‘Lin Wei’’\)returnsavailability=busy, which suffices to answer the implicit need\.

#### Two backends and control boundary\.

A heuristic backend \(deterministic surface\-cue matching, for tests and offline execution\) and an LLM backend \(structured\-JSON output, calibrated with a four\-tier gap rubric and clean benchmark\-disjoint few\-shot exemplars; recommended\-probes whitelisted against the live tool registry\) are both provided\. The architectural pipeline up to this point \(Sense–Scene–Memory\) is deterministic; theIntentFrameis the first stage at which the LLM affects control flow \(probe budget, tool priority, alert flag\)\. Section[A2\.1](https://arxiv.org/html/2606.05557#A2.SS1)quantifies the resulting per\-query adaptation, and Appendix[A12](https://arxiv.org/html/2606.05557#A12)shows that the examples act as gap calibration rather than answer templates\. The heuristic backend is provided for offline tests and air\-gapped deployments; all reported results use the LLM backend\.

#### Memory architecture\.

AURA’s memory scores each itemmmagainst queryqqat timettby a weighted combination of recency, importance, and lexical similarity \(Eq\.[1](https://arxiv.org/html/2606.05557#A23.E1), Appendix[A23](https://arxiv.org/html/2606.05557#A23), with weights0\.3:0\.4:0\.30\.3\{:\}0\.4\{:\}0\.3and decay0\.010\.01\)\. Memory types followTulving and others \([1972](https://arxiv.org/html/2606.05557#bib.bib24)\)\(observation, conversation, reflection, plan\), with reflections triggered everyθ=10\\theta\{=\}10observations\. Keyword\-based similarity is a deliberate reproducibility trade\-off; embedding retrieval is a natural extension\. A three\-stage enrichment protocol routes each query through context gathering, single\-step probe verification, and enriched generation \(Appendix[A23](https://arxiv.org/html/2606.05557#A23)\)\.

#### Relationship to long\-context approaches\.

A natural alternative is to pack the full environment state into the prompt and rely on the LLM’s long\-context attention\.AURA’s selective probing pays one extra IntentInferrer call but issues 0–3 targeted probes whose count tracks the inferred gap, rather than the world size\. We do not run a head\-to\-head against a stuff\-everything baseline \(such a baseline would also need to decide which slice of memory to include\); the Static\-Context baseline \(§[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)\) \(a packed scene snapshot\) is the closest in\-suite analogue and reaches FA0\.4500\.450vs\.AURAFull’s0\.6400\.640\. Appendix[A7](https://arxiv.org/html/2606.05557#A7.SS0.SSS0.Px3)sketches the token\-cost and prompt\-content differences\.

## 4AURATown: A Multi\-Agent Social Simulation

AURATownis a 60×\\times60 grid\-based social simulation with 5 named agents and 20 named locations over a 6:00–23:00 day \(full setup, map, and a tick\-18:00 mechanism snapshot in Appendix[A4](https://arxiv.org/html/2606.05557#A4)\)\. Two decisions depart fromParket al\.\([2023](https://arxiv.org/html/2606.05557#bib.bib1)\)’s 25\-agent Smallville\.*\(i\) Scale\.*5 agents yields tractable per\-query ground truth; we test a per\-agent mechanism, not multi\-agent emergence\.*\(ii\) Public/private state split\.*locationandactionare visible in the scene snapshot, butavailability,emotional\_state,unspoken\_goal, andbeliefs\_about\_othersare hidden and only retrievable via probe tools\. Each agent’s private state updates deterministically each tick via a seven\-rule priority table \(Appendix[A3](https://arxiv.org/html/2606.05557#A3)\): e\.g\., an agent at a loaded workplace becomes busy and tired\-focused; one at an empty workplace becomes available and lonely\. Beliefs about other agents refresh only on co\-location, producing the staleness that*second\_order*queries probe\. The released codebase additionally includes a chunk\-based procedural world used by the demo deployment \(demo/town/chunks\.py\); experiments here use only the fixed 60×\\times60 / 5\-agent / 20\-location subset\.

## 5Experiments

We evaluateAURAalong three axes: environment access as a boundary condition \(§[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)\), intent\-directed probing \(§[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\), and adaptive probe allocation\. Primary experiments usegpt\-4o\-minias both agent backbone and LLM\-as\-judge \(full hyperparameters in Appendix[A6](https://arxiv.org/html/2606.05557#A6); cross\-backbone robustness onclaude\-haiku\-4\-5,qwen\-plus,gemini\-2\.5\-flashin Appendix[A13](https://arxiv.org/html/2606.05557#A13)\)\. The same\-family judge–policy setup limits independence; we address this with a strict precision rescore in Appendix[A9](https://arxiv.org/html/2606.05557#A9)\. Baselines, metrics, and additional diagnostic checks \(routine grounding, component ablation, SOTOPIA, human eval, budget sweep\) are in Appendices[A8](https://arxiv.org/html/2606.05557#A8)–[A17](https://arxiv.org/html/2606.05557#A17)\.

#### Regime characterisation\.

AURAIntent leads the implicit\-need regime \(§[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\); on factual grounding \(§[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)\), gap\-routed probing is an access\-cost Pareto point rather than an accuracy winner\. Cross\-domain sanity checks \(FANToM, LoCoMo, GAIA\) are reported in Appendices[A20](https://arxiv.org/html/2606.05557#A20)–[A22](https://arxiv.org/html/2606.05557#A22)\.

#### Routine grounding: metric saturates\.

Under 100 simulation steps per condition×\\times3 seeds, all five architectures \(Vanilla, Static Context, ReAct,AURANo\-Probe,AURAFull\) fall within0\.0240\.024absolute GA spread, every pairedtt\-test vs\. vanilla givesp\>0\.5p\>0\.5\. The metric is saturated: memory\-utilisation≈1\.0\\approx 1\.0for every method, most actions are trivially grounded \(“sleeping at home at 6 AM”\)\. This is evidence about the metric, not the mechanism \(Appendix[A16](https://arxiv.org/html/2606.05557#A16)\)\.

### 5\.1Factual Grounding: Environment Access as a Boundary Condition

This subsection tests whether gap\-routed probing is merely a generic accuracy booster \(it is not\)\. We collect 50 user queries about the environment spanning 5 categories \(spatial, social, temporal, memory, planning\)\. Responses are scored against ground\-truth environment state by agpt\-4o\-miniLLM\-as\-judge \(same family as the agent backbone — see Appendix[A9](https://arxiv.org/html/2606.05557#A9)for a strict precision rescore that reduces sensitivity to judge softness\)\. We repeat across three random seeds\{42,123,456\}\\\{42,123,456\\\}and report mean±\\pmstd, together with pairedtt\-tests against two reference conditions: the vanilla LLM \(to measure total architectural effect\) andAURA\(No Probe\) \(to isolate the marginal contribution of proactive probing on top of the Perceive/Scene/Memory pipeline\)\.

Table 1:Factual Accuracy \(FA\) on environment\-grounded queries \(N=50N\{=\}50, 3 seeds\)\.Boldmarks within\-column best by criterion: Fixed\-Probe / Plan\-and\-Solve on raw FA; GapRouted infers a per\-query gap and executes only the recommended probes;AURAGapRouted \(⋆\\star, Pareto frontier in Fig\.[2](https://arxiv.org/html/2606.05557#S5.F2)\) on access cost \(Probes\); ReAct on contradicted\-claim rate\. Contrad\. % is diagnostic;AURAmakes no hallucination claim in this regime\. Strict precision rescore in Appendix[A9](https://arxiv.org/html/2606.05557#A9)\.†ReAct calls tools via OpenAI function\-calling, bypassing the AURATown probe registry; its probe count is not directly comparable\.#### Scope evidence: factual lookup is an accuracy–access tradeoff regime\.

The 50 factual queries ask for environment facts — “Where is X?”, “What time is it?” — so saturated access is a strong control\. Fixed\-Probe invokes all eight environment tools on every query and reaches0\.7660\.766FA, statistically above GapRouted \(0\.6960\.696; pairedΔ=\+0\.070\\Delta\{=\}\{\+\}0\.070,p=0\.031p\{=\}0\.031\)\. Plan\-and\-Solve is similar in raw FA \(0\.7640\.764; paired vs\. GapRoutedΔ=\+0\.068\\Delta\{=\}\{\+\}0\.068,p=0\.056p\{=\}0\.056\)\. The gap\-routed controller therefore is*not*the raw\-accuracy winner on factual lookup\. Its contribution is an access Pareto point \(Figure[2](https://arxiv.org/html/2606.05557#S5.F2)\):1\.401\.40probes/query vs\. Fixed\-Probe’s8\.008\.00\(𝟖𝟐%\\mathbf\{82\\%\}fewer;p=4\.9×10−52p\{=\}4\.9\\times 10^\{\-52\}\), and disclosure0\.920\.92vs\.5\.005\.00\(p=6\.7×10−38p\{=\}6\.7\\times 10^\{\-38\}\)\. On a 30\-query privacy\-sensitive factual slice \(Appendix[A10](https://arxiv.org/html/2606.05557#A10)\), GapRouted ties Plan\-and\-Solve and ReAct in FA \(p=0\.86p\{=\}0\.86and0\.650\.65\) while reducing forbidden\-tool violations to𝟎%\\mathbf\{0\\%\}\(Plan\-and\-Solve78\.9%78\.9\\%, ReAct25\.6%25\.6\\%, Fixed\-Probe100%100\\%\)\. Per\-query wall\-clock latency tells a more cautious story: GapRouted pays the IntentInferrer’s extra LLM round trip and is slower at the median than Fixed\-Probe \(4\.084\.08vs\.2\.372\.37s; Appendix[A11](https://arxiv.org/html/2606.05557#A11)\), so the cost\-of\-selectivity claim holds on probe count and disclosure rather than on latency\. The mechanism’s primary accuracy contribution remains the implicit\-need setting in Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2); the factual\-grounding regime bounds the cost of selective access\.

#### Reflexion failure mode\.

Reflexion \(Table[1](https://arxiv.org/html/2606.05557#S5.T1)\) collapses to0\.1740\.174FA because the reflection step withdraws probe\-supported claims \(Appendix[A15](https://arxiv.org/html/2606.05557#A15)\); the Full\-vs\-No\-Probe contrast is near\-null at gap≈0\{\\approx\}0\(p=0\.060p\{=\}0\.060\)\.

![Refer to caption](https://arxiv.org/html/2606.05557v1/figures/rq2-pareto.png)Figure 2:Factual\-grounding access Pareto\. Left: factual lookup \(N=50×3N\{=\}50\\times 3seeds\) — factual accuracy vs\. mean probes per query\.AURAGapRouted \(⋆\\star\) sits on the frontier with1\.401\.40probes; Fixed\-Probe wins raw FA at8×8\\timesthe access cost\. Right: privacy\-sensitive distractor slice \(N=30×3N\{=\}30\\times 3seeds\) — factual accuracy vs\. forbidden\-tool violation rate\.AURAGapRouted holds the Pareto vertex with0%0\\%violations while approaching Fixed\-Probe \(100%100\\%\) and Plan\-and\-Solve \(78\.9%78\.9\\%\) in FA\.
#### Privacy as a design benefit\.

Gap routing’s zero forbidden\-tool violation rate is not a post\-hoc observation: low\-gap factual queries receiveB=0B\{=\}0budget and skip the probe registry entirely, so they structurally cannot invoke a forbidden tool\. This makes the access–cost Pareto a privacy Pareto as well\.

#### Strict\-precision robustness\.

A strict\-precision rescore \(Appendix[A9](https://arxiv.org/html/2606.05557#A9)\) confirms the access\-cost picture: per\-category gains concentrate on temporal and memory; the overall architectural margin above ReAct narrows \(p=0\.064p\{=\}0\.064\)\. The IntentFrame’s primary accuracy contribution is the implicit\-need regime \(§[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\)\.

### 5\.2Implicit\-Need Surfacing via Intent Inference

This subsection tests whether intent inference itself surfaces the user’s implicit information need on queries specifically designed to hide that need behind a literal surface form\.

#### Query set\.

We construct 25 primary implicit\-intent queries stratified across five subcategories that vary the relationship between surface form and implicit need:*availability*\(e\.g\., “where is Lin Wei?”→\\to“is she free?”\),*mood*,*appropriateness*,*latent\_goal*, and*second\_order*\(representative queries for the remaining four and the second\-order belief\-vs\-truth judging rule in Appendix[A14](https://arxiv.org/html/2606.05557#A14)\)\. The scene exposes only public state \(location, action\); private\-state fields \(availability, emotional state, unspoken goal, beliefs about others\) are only retrievable via probe tools\.

#### Conditions\.

Four answer strategies share the same backbone \(gpt\-4o\-mini, temperature 0\.1\) and a 5\-tool agent\-state registry \(Appendix[A5](https://arxiv.org/html/2606.05557#A5); disjoint from the factual\-grounding registry\):Literal\(scene only, no tools\);NoIntent\(ReAct\-style, up toB=3B\{=\}3tool calls\);Plan\-and\-Solve\(Wanget al\.,[2023b](https://arxiv.org/html/2606.05557#bib.bib58)\)\(plan/execute/synthesise, budget 3\);Intent\(IntentInferrer→\\togap\-derived probe ceiling 0–3→\\todirected tool loop, with heads\-up prefix on high\-gap queries\)\. Each response is scored by agpt\-4o\-minijudge on two dimensions in\[0,1\]\[0,1\]:literal\_scoreandimplicit\_score\(does it surface the implicit need?\)\. We run 3 seeds \(42, 123, 456\) over the 25 queries and report query\-level pairedtt\-tests as primary, with seed\-level paired tests andquery\_idcluster bootstraps as secondary checks \(Appendix[A14](https://arxiv.org/html/2606.05557#A14)\)\.

Table 2:Implicit\-intent comparison \(N=25N\{=\}25, 3 seeds; scores in\[0,1\]\[0,1\]\)\.Boldmarks the best deployable system per column; Fixed\-private and Oracle\-intent are diagnostic upper bounds \(require unconditional private\-state access or gold\-tool knowledge\) and not bolded\. Tool\-using rows share the same 5\-tool registry\. Prompt\-example and component ablations in Appendices[A12](https://arxiv.org/html/2606.05557#A12),[A17](https://arxiv.org/html/2606.05557#A17)\.Table 3:100\-query four\-scene implicit\-intent benchmark \(3 seeds,gpt\-4o\-mini\)\. Implicit\-need coverage scores in\[0,1\]\[0,1\]\.Boldmarks best per column\.ppvalues are paired tests on query\-seed cells\.
#### Overall\.

On the 100\-query four\-scene benchmark \(Table[3](https://arxiv.org/html/2606.05557#S5.T3)\),AURA\-Intent reaches0\.804\\mathbf\{0\.804\}implicit\-need coverage versus0\.7330\.733for NoIntent \(ReAct\-style\) and0\.2230\.223for Literal: paired𝚫=\+0\.071\\boldsymbol\{\\Delta\{=\}\{\+\}0\.071\},𝒑=1\.0×𝟏𝟎−𝟔\\boldsymbol\{p\{=\}1\.0\\times 10^\{\-6\}\}\. Three of four scenes are individually significant; the post\-event night scene D ties because public state already telegraphs availability\. A 25\-query pilot \(Scene A of the four\-scene set, Table[2](https://arxiv.org/html/2606.05557#S5.T2)\) confirms the direction at higher absolute gain \(Δ=\+0\.16\\Delta\{=\}\{\+\}0\.16,p=0\.006p\{=\}0\.006\) and additionally tests Plan\-and\-Solve \(Δ=\+0\.28\\Delta\{=\}\{\+\}0\.28,p=8\.3×10−5p\{=\}8\.3\\times 10^\{\-5\}\) and two diagnostic upper bounds \(fixed\-private0\.8510\.851, oracle\-intent0\.8610\.861; underpowered to separate fromAURAatN=25N\{=\}25\)\.

#### Per\-subcategory structure\.

The 100\-query v2 subcategory breakdown reveals where intent inference helps most\.*Availability*shows the largest gain \(Δ=\+0\.29\\Delta\{=\}\{\+\}0\.29,p=2\.7×10−11p\{=\}2\.7\\times 10^\{\-11\}\): surface queries like “where is X?” fully mask the implicit need, so the IntentFrame’s gap score routes a probe that would not otherwise fire\.*Appropriateness*\(\+0\.11\+0\.11,p=8\.2×10−4p\{=\}8\.2\\times 10^\{\-4\}\) and*mood*\(\+0\.07\+0\.07,p=1\.7×10−3p\{=\}1\.7\\times 10^\{\-3\}\) follow at smaller magnitudes\.*Second\_order*ties \(−0\.02\-0\.02, NS\) because “does X think Y…” already cues belief\-state probing without gap inference\.*Latent\_goal*shows a residual deficit \(−0\.09\-0\.09,p=6\.2×10−4p\{=\}6\.2\\times 10^\{\-4\}\)\. The pattern tracks the controller’s design: the gap score adds value precisely when the surface form is maximally decoupled from the implicit need\.

#### Ablation: prompt examples\.

A three\-way prompt ablation \(Appendix[A12](https://arxiv.org/html/2606.05557#A12)\) shows the gain is not example memorisation: disjoint examples reduce Intent by only0\.037\\mathbf\{0\.037\}\(contrast remains significant\)\.Removing examples entirelycollapses the gap calibration and the Intent\-vs\.\-NoIntent contrastbecomes non\-significant\. The pattern is consistent with load\-bearing gap calibration, not answer\-template memorisation\.

#### Cross\-backbone\.

The Intent\-vs\.\-NoIntent gain reproduces onclaude\-haiku\-4\.5\(Δ=\+0\.086\\Delta\{=\}\{\+\}0\.086on v2,p=3\.7×10−3p\{=\}3\.7\\times 10^\{\-3\}\) and onqwen\-plus\(Δ=\+0\.25\\Delta\{=\}\{\+\}0\.25, CI\[\+0\.14,\+0\.37\]\[\+0\.14,\+0\.37\]\)\. Per\-backbone breakdown and a Gemini JSON\-parse failure case in Appendix[A13](https://arxiv.org/html/2606.05557#A13)\.

#### Backend ablation: gap inference is load\-bearing\.

Replacing theLLMIntentInferrerwith a deterministicHeuristicIntentInferrer\(rule\-based gap estimation, identical downstream plumbing\) drops overall implicit score from0\.803\\mathbf\{0\.803\}to0\.3680\.368\(Δ=−0\.44\\Delta\{=\}\{\-\}0\.44; Appendix[A18](https://arxiv.org/html/2606.05557#A18)\), with the largest drops on lexically\-decoupled subcategories\. The lift attaches to LLM\-mediated gap inference, not the surrounding scaffolding\.

#### Human evaluation\.

Eight independent raters scored 50 paired \(AURAvs\. Vanilla\) scenarios on four dimensions \(Appendix[A17\.3](https://arxiv.org/html/2606.05557#A17.SS3)\)\.AURAreceives significantly higher ratings on all four: environmental awarenessΔ=\+1\.86\\Delta\{=\}\{\+\}1\.86\(p=0\.017p\{=\}0\.017, rater\-aggregated Wilcoxon\), response helpfulness\+1\.58\+1\.58, agent believability\+1\.59\+1\.59, factual accuracy\+1\.39\+1\.39; all four cluster\-bootstrap CIs exclude zero\. At the cell level,74%74\\%of \(scenario, dimension\) cells show≥6/8\{\\geq\}6/8rater consensus forAURA;0%0\\%show Vanilla consensus\.

#### Adaptive budget: ceiling, not target\.

Under fixedexplore\_max\_steps=3, mean probes per query range0\.800\.80–2\.202\.20across subcategories and all four values\{0,1,2,3\}\\\{0,1,2,3\\\}appear in the 75 runs; the Pearson correlation between gap and realised probe count is only𝐫=0\.19\\mathbf\{r\{=\}0\.19\}, so the gap routes a ceiling rather than determining a target; in practice the agent issues fewer probes than a fixed\-budget system on most queries while retaining full budget for high\-gap ones \(per\-subcategory distribution in Appendix[A2\.1](https://arxiv.org/html/2606.05557#A2.SS1)\)\.

## 6Conclusion

AURAinserts a small inference step between scene perception and tool use: anIntentFramewhose gap score routes private\-state probes before the agent answers\. On a 100\-query four\-scene benchmark the controller significantly improves implicit\-need coverage over ReAct\-style probing \(Δ=\+0\.07\\Delta\{=\}\{\+\}0\.07,p<10−6p\{<\}10^\{\-6\}\), with three of four scenes individually significant and the gain reproducing across a 25\-query pilot, a second backbone, and a prompt ablation\. A backend ablation \(LLM→\\toheuristic gap inference:0\.803→0\.3680\.803\\to 0\.368\) localises the lift to LLM\-mediated gap calibration, suggesting that intent\-direction is an LLM\-prompted operation at a specific control point rather than an emergent property of the pipeline\. The mechanism’s scope is bounded to situated regimes with tool\-mediated hidden state \(Limitations\); whether this control\-point view generalises to multi\-turn, multi\-user, or open\-ended planning settings is an open question\. Two extensions follow naturally\. First, the IntentFrame currently operates on a single user query; in multi\-turn dialogue the gap score could be updated incrementally as the conversation reveals more of the user’s intent, potentially reducing probe cost on follow\-up queries\. Second, the current gap\-to\-budget map is a hand\-tuned step function; learning the mapping from interaction logs could improve calibration beyond what few\-shot exemplars provide\.

## Limitations

#### Regime scope\.

The IntentFrame controller targets situated queries with tool\-mediated hidden state\. On factual grounding \(§[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)\) it acts as an access–cost Pareto point rather than an accuracy winner; cross\-domain checks on FANToM, LoCoMo, and GAIA show no measurable lift when private state is already in\-context or structurally inaccessible \(Appendices[A20](https://arxiv.org/html/2606.05557#A20)–[A22](https://arxiv.org/html/2606.05557#A22)\)\.

#### Benchmark and calibration\.

The 100\-query four\-scene benchmark is author\-written; inter\-annotator agreement on the 5\-subcategory partition is substantial \(κ=0\.61\\kappa\{=\}0\.61, two independent annotators; details in Appendix[A14](https://arxiv.org/html/2606.05557#A14)\)\. The few\-shot calibration examples are load\-bearing for gap estimation: removing them reduces the gain to non\-significance, though replacing them with benchmark\-disjoint examples preserves it \(Appendix[A12](https://arxiv.org/html/2606.05557#A12)\)\. Human evaluation usesN=8N\{=\}8raters \(Krippendorff’sα=0\.43\\alpha\{=\}0\.43\); directional agreement is strong \(74%74\\%consensus\) but magnitude estimates carry substantial uncertainty\.

#### Cross\-backbone\.

Three of four tested backbones reproduce the gain;gemini\-2\.5\-flashfails the IntentFrame JSON parser and silently falls back to a heuristic \(Appendix[A13](https://arxiv.org/html/2606.05557#A13)\)\.

## References

- W\. Chen, Y\. Su, J\. Zuo, C\. Yang, C\. Yuan, C\. Chan, H\. Yu, Y\. Lu, Y\. Hung, C\. Qian, Y\. Qin, X\. Cong, R\. Xie, Z\. Liu, M\. Sun, and J\. Zhou \(2024a\)AgentVerse: facilitating multi\-agent collaboration and exploring emergent behaviors\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=EHg5GDnyq1)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Chen, W\. Du, W\. Zhang, K\. Liu, J\. Liu, M\. Zheng, J\. Zhuo, S\. Zhang, D\. Lin, K\. Chen, and F\. Zhao \(2024b\)T\-eval: evaluating the tool utilization capability of large language models step by step\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 9510–9529\.External Links:[Link](https://aclanthology.org/2024.acl-long.515/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.515)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Gao, Y\. Xiong, X\. Gao, K\. Jia, J\. Pan, Y\. Bi, Y\. Dai, J\. Sun, H\. Wang, H\. Wang,et al\.\(2023\)Retrieval\-augmented generation for large language models: a survey\.arXiv preprint arXiv:2312\.109972\(1\),pp\. 32\.Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- S\. Hong, M\. Zhuge, J\. Chen, X\. Zheng, Y\. Cheng, J\. Wang, C\. Zhang, Z\. Wang, S\. K\. S\. Yau, Z\. Lin, L\. Zhou, C\. Ran, L\. Xiao, C\. Wu, and J\. Schmidhuber \(2024\)MetaGPT: meta programming for a multi\-agent collaborative framework\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=VtmBAGCN7o)Cited by:[§1](https://arxiv.org/html/2606.05557#S1.p1.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- H\. Kim, M\. Sclar, X\. Zhou, R\. Bras, G\. Kim, Y\. Choi, and M\. Sap \(2023\)FANToM: a benchmark for stress\-testing machine theory of mind in interactions\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 14397–14413\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.890/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.890)Cited by:[Appendix A20](https://arxiv.org/html/2606.05557#A20.p1.5)\.
- M\. Kosinski \(2023\)Theory of mind may have spontaneously emerged in large language models\.arXiv preprint arXiv:2302\.020834\(169\),pp\. 2\.Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px4.p1.1)\.
- P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela \(2020\)Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.InProceedings of the 34th International Conference on Neural Information Processing Systems,NIPS ’20,Red Hook, NY, USA\.External Links:ISBN 9781713829546Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- G\. Li, H\. A\. A\. K\. Hammoud, H\. Itani, D\. Khizbullin, and B\. Ghanem \(2023\)CAMEL: communicative agents for ”mind” exploration of large language model society\.InThirty\-seventh Conference on Neural Information Processing Systems,External Links:[Link](https://openreview.net/forum?id=3IyL2XWDkG)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Li, J\. Liu, Y\. Wang, Y\. Wu, and M\. Xu \(2026\)BMAM: brain\-inspired multi\-agent memory framework\.arXiv preprint arXiv:2601\.20465\.Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- Z\. Li, Y\. Li, H\. Xie, and S\. J\. Qin \(2025\)CondAmbigQA: a benchmark and dataset for conditional ambiguous question answering\.InThe 2025 Conference on Empirical Methods in Natural Language Processing,External Links:[Link](https://openreview.net/forum?id=rmTvB2Ximp)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px3.p1.1)\.
- X\. Liu, H\. Yu, H\. Zhang, Y\. Xu, X\. Lei, H\. Lai, Y\. Gu, H\. Ding, K\. Men, K\. Yang, S\. Zhang, X\. Deng, A\. Zeng, Z\. Du, C\. Zhang, S\. Shen, T\. Zhang, Y\. Su, H\. Sun, M\. Huang, Y\. Dong, and J\. Tang \(2024\)AgentBench: evaluating LLMs as agents\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=zAdUB0aCTQ)Cited by:[Appendix A8](https://arxiv.org/html/2606.05557#A8.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px1.p1.1)\.
- A\. Maharana, D\. Lee, S\. Tulyakov, M\. Bansal, F\. Barbieri, and Y\. Fang \(2024\)Evaluating very long\-term conversational memory of LLM agents\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 13851–13870\.External Links:[Link](https://aclanthology.org/2024.acl-long.747/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.747)Cited by:[Appendix A21](https://arxiv.org/html/2606.05557#A21.p1.5)\.
- C\. Packer, V\. Fang, S\. G\. Patil, K\. Lin, S\. Wooders, and J\. E\. Gonzalez \(2023\)MemGPT: towards llms as operating systems\.CoRRabs/2310\.08560\.External Links:[Link](https://doi.org/10.48550/arXiv.2310.08560)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- J\. S\. Park, J\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein \(2023\)Generative agents: interactive simulacra of human behavior\.InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology,UIST ’23,New York, NY, USA\.External Links:ISBN 9798400701320,[Link](https://doi.org/10.1145/3586183.3606763),[Document](https://dx.doi.org/10.1145/3586183.3606763)Cited by:[Appendix A8](https://arxiv.org/html/2606.05557#A8.SS0.SSS0.Px1.p1.2),[§1](https://arxiv.org/html/2606.05557#S1.p1.1),[§1](https://arxiv.org/html/2606.05557#S1.p2.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1),[§3](https://arxiv.org/html/2606.05557#S3.p1.6),[§4](https://arxiv.org/html/2606.05557#S4.p1.2)\.
- G\. Pasternak, D\. Rajagopal, J\. White, D\. Atreja, M\. Thomas, G\. Hurn\-Maloney, and A\. Lewis \(2025\)Beyond reactivity: measuring proactive problem solving in llm agents\.arXiv preprint arXiv:2510\.19771\.Cited by:[Table 9](https://arxiv.org/html/2606.05557#A7.T9.1.5.4.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Qin, S\. Liang, Y\. Ye, K\. Zhu, L\. Yan, Y\. Lu, Y\. Lin, X\. Cong, X\. Tang, B\. Qian, S\. Zhao, L\. Hong, R\. Tian, R\. Xie, J\. Zhou, M\. Gerstein, dahai li, Z\. Liu, and M\. Sun \(2024\)ToolLLM: facilitating large language models to master 16000\+ real\-world APIs\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=dHng2O0Jjr)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px1.p1.1)\.
- M\. Sap, R\. Le Bras, D\. Fried, and Y\. Choi \(2022\)Neural theory\-of\-mind? on the limits of social intelligence in large LMs\.InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing,Y\. Goldberg, Z\. Kozareva, and Y\. Zhang \(Eds\.\),Abu Dhabi, United Arab Emirates,pp\. 3762–3780\.External Links:[Link](https://aclanthology.org/2022.emnlp-main.248/),[Document](https://dx.doi.org/10.18653/v1/2022.emnlp-main.248)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px4.p1.1)\.
- T\. Schick, J\. Dwivedi\-Yu, R\. Dessí, R\. Raileanu, M\. Lomeli, E\. Hambro, L\. Zettlemoyer, N\. Cancedda, and T\. Scialom \(2023\)Toolformer: language models can teach themselves to use tools\.InProceedings of the 37th International Conference on Neural Information Processing Systems,NIPS ’23,Red Hook, NY, USA\.Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px1.p1.1)\.
- M\. Sclar, S\. Kumar, P\. West, A\. Suhr, Y\. Choi, and Y\. Tsvetkov \(2023\)Minding language models’ \(lack of\) theory of mind: a plug\-and\-play multi\-character belief tracker\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 13960–13980\.External Links:[Link](https://aclanthology.org/2023.acl-long.780/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.780)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px4.p1.1)\.
- N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\)Reflexion: language agents with verbal reinforcement learning\.InProceedings of the 37th International Conference on Neural Information Processing Systems,NIPS ’23,Red Hook, NY, USA\.Cited by:[Table 9](https://arxiv.org/html/2606.05557#A7.T9.1.8.7.1),[item 4](https://arxiv.org/html/2606.05557#A8.I1.i4.p1.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px1.p1.1)\.
- Y\. Tang, H\. Tang, T\. Cao, L\. Nguyen, A\. Zhang, X\. Cao, C\. Liu, W\. Ding, and Y\. Li \(2026\)ProAgentBench: evaluating llm agents for proactive assistance with real\-world data\.arXiv e\-prints,pp\. arXiv–2602\.Cited by:[Table 9](https://arxiv.org/html/2606.05557#A7.T9.1.6.5.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- Thinking Machines Lab \(2026\)Interaction models: a scalable approach to Human\-AI collaboration\.Note:Blog post, Thinking Machines Lab\.[https://thinkingmachines\.ai/blog/interaction\-models/](https://thinkingmachines.ai/blog/interaction-models/)Accessed: 2026\-05\-12Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- E\. Tulvinget al\.\(1972\)Episodic and semantic memory\.Organization of memory1\(381\-403\),pp\. 1\.Cited by:[Appendix A23](https://arxiv.org/html/2606.05557#A23.SS0.SSS0.Px2.p1.1),[§3\.3](https://arxiv.org/html/2606.05557#S3.SS3.SSS0.Px3.p1.6)\.
- T\. Ullman \(2023\)Large language models fail on trivial alterations to theory\-of\-mind tasks\.arXiv preprint arXiv:2302\.08399\.Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px4.p1.1)\.
- A\. S\. Vezhnevets, J\. P\. Agapiou, A\. Aharon, R\. Ziv, J\. Matyas, E\. A\. Duéñez\-Guzmán, W\. A\. Cunningham, S\. Osindero, D\. Karmon, and J\. Z\. Leibo \(2023\)Generative agent\-based modeling with actions grounded in physical, social, or digital space using concordia\.arXiv preprint arXiv:2312\.03664\.Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- L\. Wang, C\. Ma, X\. Feng, Z\. Zhang, H\. Yang, J\. Zhang, Z\. Chen, J\. Tang, X\. Chen, Y\. Lin, W\. X\. Zhao, Z\. Wei, and J\. Wen \(2023a\)A survey on large language model based autonomous agents\.Frontiers of Computer Science18\.External Links:[Link](https://api.semanticscholar.org/CorpusID:261064713)Cited by:[§1](https://arxiv.org/html/2606.05557#S1.p1.1)\.
- L\. Wang, W\. Xu, Y\. Lan, Z\. Hu, Y\. Lan, R\. K\. Lee, and E\. Lim \(2023b\)Plan\-and\-solve prompting: improving zero\-shot chain\-of\-thought reasoning by large language models\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 2609–2634\.External Links:[Link](https://aclanthology.org/2023.acl-long.147/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.147)Cited by:[Table 9](https://arxiv.org/html/2606.05557#A7.T9.1.9.8.1),[item 5](https://arxiv.org/html/2606.05557#A8.I1.i5.p1.1),[§1](https://arxiv.org/html/2606.05557#S1.p2.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px1.p1.1),[§5\.2](https://arxiv.org/html/2606.05557#S5.SS2.SSS0.Px2.p1.5)\.
- Z\. Xi, W\. Chen, X\. Guo, W\. He, Y\. Ding, B\. Hong, M\. Zhang, J\. Wang, S\. Jin, E\. Zhou, R\. Zheng, X\. Fan, X\. Wang, L\. Xiong, Y\. Zhou, W\. Wang, C\. Jiang, Y\. Zou, X\. Liu, Z\. Yin, S\. Dou, R\. Weng, W\. Qin, Y\. Zheng, X\. Qiu, X\. Huang, Q\. Zhang, and T\. Gui \(2025\)The rise and potential of large language model based agents: a survey\.Science China Information Sciences68\(2\),pp\. 121101\.External Links:[Document](https://dx.doi.org/10.1007/s11432-024-4222-0),[Link](https://doi.org/10.1007/s11432-024-4222-0)Cited by:[§1](https://arxiv.org/html/2606.05557#S1.p1.1)\.
- L\. Xu, Z\. Hu, D\. Zhou, H\. Ren, Z\. Dong, K\. Keutzer, S\. Ng, and J\. Feng \(2024\)MAgIC: investigation of large language model powered multi\-agent in cognition, adaptability, rationality and collaboration\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 7315–7332\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.416/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.416)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- B\. Yang, L\. Xu, L\. Zeng, Y\. Guo, S\. Jiang, W\. Lu, K\. Liu, H\. Xiang, X\. Jiang, G\. Xing,et al\.\(2025\)ProAgent: harnessing on\-demand sensory contexts for proactive llm agent systems\.arXiv preprint arXiv:2512\.06721\.Cited by:[Table 9](https://arxiv.org/html/2606.05557#A7.T9.1.4.3.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- B\. Yang, L\. Xu, L\. Zeng, K\. Liu, S\. Jiang, W\. Lu, H\. Chen, X\. Jiang, G\. Xing, and Z\. Yan \(2026\)ContextAgent: context\-aware proactive LLM agents with open\-world sensory perceptions\.InThe Thirty\-ninth Annual Conference on Neural Information Processing Systems,External Links:[Link](https://openreview.net/forum?id=tRXt10xKc5)Cited by:[Table 9](https://arxiv.org/html/2606.05557#A7.T9.1.3.2.1),[Appendix A8](https://arxiv.org/html/2606.05557#A8.SS0.SSS0.Px1.p1.2),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. R\. Narasimhan, and Y\. Cao \(2023\)ReAct: synergizing reasoning and acting in language models\.InThe Eleventh International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=WE_vluYUL-X)Cited by:[Table 9](https://arxiv.org/html/2606.05557#A7.T9.1.7.6.1),[item 3](https://arxiv.org/html/2606.05557#A8.I1.i3.p1.1),[§1](https://arxiv.org/html/2606.05557#S1.p2.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.05557#S3.SS2.p1.3)\.
- L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. Xing, H\. Zhang, J\. E\. Gonzalez, and I\. Stoica \(2023\)Judging LLM\-as\-a\-judge with MT\-bench and chatbot arena\.InThirty\-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track,External Links:[Link](https://openreview.net/forum?id=uccHPGDlao)Cited by:[Appendix A8](https://arxiv.org/html/2606.05557#A8.SS0.SSS0.Px2.p1.1)\.
- W\. Zhong, L\. Guo, Q\. Gao, H\. Ye, and Y\. Wang \(2024\)MemoryBank: enhancing large language models with long\-term memory\.InProceedings of the Thirty\-Eighth AAAI Conference on Artificial Intelligence and Thirty\-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence,AAAI’24/IAAI’24/EAAI’24\.External Links:ISBN 978\-1\-57735\-887\-9,[Link](https://doi.org/10.1609/aaai.v38i17.29946),[Document](https://dx.doi.org/10.1609/aaai.v38i17.29946)Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- X\. Zhou, H\. Zhu, L\. Mathur, R\. Zhang, H\. Yu, Z\. Qi, L\. Morency, Y\. Bisk, D\. Fried, G\. Neubig, and M\. Sap \(2024\)SOTOPIA: interactive evaluation for social intelligence in language agents\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=mM7VurbA4r)Cited by:[§A17\.2](https://arxiv.org/html/2606.05557#A17.SS2.p1.1),[Appendix A8](https://arxiv.org/html/2606.05557#A8.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.
- K\. Zhu, H\. Du, Z\. Hong, X\. Yang, S\. Guo, Z\. Wang, Z\. Wang, C\. Qian, X\. Tang, H\. Ji, and J\. You \(2025\)MultiAgentBench : evaluating the collaboration and competition of LLM agents\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 8580–8622\.External Links:[Link](https://aclanthology.org/2025.acl-long.421/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.421),ISBN 979\-8\-89176\-251\-0Cited by:[§2](https://arxiv.org/html/2606.05557#S2.SS0.SSS0.Px2.p1.1)\.

## Appendix A1Bounded\-Probing Algorithm

Algorithm 1Bounded Proactive Probing \(Section[3\.2](https://arxiv.org/html/2606.05557#S3.SS2)\)0:Agent state

ss, tool registry

𝒯=\{t1,…,tK\}\\mathcal\{T\}=\\\{t\_\{1\},\\dots,t\_\{K\}\\\}, max steps

BB, LLM planner

ϕ\\phi
0:Probe result

P=\(summary,trace\)P=\(\\text\{summary\},\\text\{trace\}\)
1:

context←BuildContext​\(s\)\\text\{context\}\\leftarrow\\textsc\{BuildContext\}\(s\)\{agent, location, time, nearby\}

2:

trace←⟨⟩\\text\{trace\}\\leftarrow\\langle\\,\\rangle\{empty probe trace\}

3:for

i=1i=1to

BBdo

4:

d←ϕ​\(𝒯,context,trace\)d\\leftarrow\\phi\(\\mathcal\{T\},\\text\{context\},\\text\{trace\}\)\{LLM decides: call a tool or stop\}

5:if

d\.action=stopd\.\\text\{action\}=\\texttt\{stop\}then

6:break

7:endif

8:

r←𝒯\.Execute\(d\.tool,d\.args\)r\\leftarrow\\mathcal\{T\}\.\\textsc\{Execute\}\(d\.\\text\{tool\},d\.\\text\{args\}\)
9:

trace\.Append\(\(d\.tool,d\.args,r\)\)\\text\{trace\}\.\\textsc\{Append\}\(\(d\.\\text\{tool\},d\.\\text\{args\},r\)\)
10:

context←Update​\(context,r\)\\text\{context\}\\leftarrow\\textsc\{Update\}\(\\text\{context\},r\)
11:endfor

12:

summary←Summarize​\(trace\)\\text\{summary\}\\leftarrow\\textsc\{Summarize\}\(\\text\{trace\}\)
13:return

\(summary,trace\)\(\\text\{summary\},\\text\{trace\}\)

## Appendix A2Adaptive\-Budget Evidence \(Supplementary\)

This appendix expands the adaptive\-budget claim summarised in Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2): under a fixed globalexplore\_max\_steps=3 ceiling, the per\-query probe count actually issued byAURAvaries across subcategories by2\.8×2\.8\\times, and is driven by theIntentFrame’s gap field rather than the configured ceiling\. Figure[4](https://arxiv.org/html/2606.05557#A2.F4)shows the per\-subcategory implicit\-need scores, Figure[3](https://arxiv.org/html/2606.05557#A2.F3)the gap\-vs\-probes scatter, and Table[4](https://arxiv.org/html/2606.05557#A2.T4)the per\-subcategory means with paired latencies\.

0\.40\.40\.50\.50\.60\.60\.70\.70\.80\.80\.90\.90112233IntentFrame gapggProbes issued \(per query\)availabilitymoodappropriatenesslatent\_goalsecond\_orderceilingB​\(g\)B\(g\)Figure 3:Per\-subcategory mean ofIntentFrame\.gapvs\. probes actually issued by the LLM \(5 queries×\\times3 seeds per marker;explore\_max\_steps=3\)\. The dashed line is the deterministic ceiling mapB​\(g\)B\(g\)from Section[3\.3](https://arxiv.org/html/2606.05557#S3.SS3), capped at the configured ceiling 3\. Probe count varies2\.8×2\.8\\times\(0\.80–2\.20\) under identical configuration; all four budget values\{0,1,2,3\}\\\{0,1,2,3\\\}appear; Pearsonr​\(g,probes\)=0\.19r\(g,\\text\{probes\}\)\{=\}0\.19\. The*second\_order*subcategory has the highest gap but issues only 1 probe because one targetedget\_agent\_belief\_aboutcall already returns the belief\.availabilitymoodappropriatenesslatent\_goalsecond\_order00\.50\.511Implicit ScoreLiteralNoIntent \(ReAct\-style\)AURAIntent

Figure 4:Per\-subcategory implicit\-need surfacing on the implicit\-intent benchmark \(25 queries, 3 seeds,N=15N\{=\}15per cell\)\.AURAIntent improves most on lexically decoupled private\-state queries \(e\.g\.*availability*\); on*second\_order*queries \(“does X think Y?”\) the surface already cues belief retrieval and NoIntent ties\.### A2\.1Adaptive Budget: LLM\-Directed Probe Allocation

A practical concern for any bounded\-probing system is whether the configured probe ceiling simply becomes a fixed per\-query cost\. We therefore measure the*effective*number of probes issued after theIntentFramesets a ceiling and the Explore loop decides whether to stop early\. This analysis does not try to define autonomous\-agent behavior; it checks the narrower mechanism claim that intent inference changes probe allocation across queries\.

Using the same 25\-query×\\times3\-condition×\\times3\-seed run as Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\(forAURA\-full only;N=75N=75\), we measure the probe\-count distribution conditional on query subcategory, while holdingexplore\_max\_steps=3constant across all runs\. The configured budget is therefore a ceiling, not a target\.

Table 4:Per\-subcategory distributions underAURAFull \(fixedexplore\_max\_steps=3,N=15N=15per row: 5 queries×\\times3 seeds\)\. Gap is the IntentFrame’s literal/implicit divergence; Probes is the number of tool calls actually issued after early stopping\.#### Observed allocation patterns\.

*\(i\) Effective probe count is not constant\.*Mean probe count per query ranges from0\.800\.80\(availability\) to2\.202\.20\(appropriateness\), despite the same globalexplore\_max\_steps=3 setting\.

*\(ii\) All four budget values are used\.*Across the 75 runs, the agent issued\{0,1,2,3\}\\\{0,1,2,3\\\}tool calls—the full effective range allowed by the run\. This is consistent with the intended design: the configured budget is a ceiling, not a target\.

*\(iii\) Gap is an input, not the decision itself\.*The*second\_order*subcategory has the highest observed gap \(0\.800\.80\) yet uses only1\.001\.00probe on average\. The gap\-to\-budget rule maps0\.800\.80to a ceiling of three probes, but the Explore loop usually stops after one targeted belief\-state call because that call already returns the needed value\. The Pearson correlation between gap and actual probe count isr=0\.19r=0\.19, so the gap is not a proxy for the final number of tool calls\.

Taken together, these traces support the mechanism claim thatIntentFrame\-conditioned probing changes the cost and evidence gathered for different queries under the same run configuration\.

#### Scope of the claim\.

The architecture remains a hybrid system: Perceive, Scene, and Memory are code\-determined, while intent inference, probe selection, and response generation use LLM calls\. The adaptive\-budget result should therefore be read as evidence about one control point in the system, not as a broad claim about every stage of the pipeline\.

Table 5:Per\-subcategory implicit score \(n=15n\{=\}15per cell, 5 queries×\\times3 seeds\)\. Intent inference’s marginal gain concentrates on availability queries \(surface hides implicit need\) and ties NoIntent on second\-order queries \(surface cues “X thinks Y”\)\.

## Appendix A3Private\-State Evolution Rules

Private fields on each agent \(availability,emotional\_state,unspoken\_goal,beliefs\_about\_others\) update each simulation tick under a deterministic, transparent rule table \(source:demo/town/private\_state\_evolution\.py\)\. The intent is to make AURATown’s private state a live function of agent action and environment context rather than a static lookup\. Rules are evaluated in order; the first match wins\.

Table 6:Private\-state evolution rules \(first match wins\)\. Action keywords match againstcurrent\_actionvia simple regex; workplace map: Lin Wei→\\toSunrise Cafe, Chen Mei→\\toGeneral Store, Zhang Hao→\\tohome/library, Liu Yang/Wang Jun→\\toLibrary\.#### Beliefs\-about\-others refresh\.

Each agent’sbeliefs\_about\_othersdictionary entries refresh*only*when this agent is co\-located with the target agent in the same simulation tick\. Beliefs about non\-co\-located agents remain at the most recently observed value, so they go stale when peers move\. This is the substrate the*second\_order*subcategory of the implicit\-intent benchmark \(§[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\) probes: a query like “does Lin Wei think Zhang Hao is at home?” asks for the believer’s stale memory, not the target’s current ground\-truth\.

The full rule table is unit\-tested attests/test\_private\_state\_evolution\.py\(14 cases, all rules \+ co\-location/staleness invariants\)\. The evolution itself is pure\-function and zero\-LLM, so it is reproducible across seeds and is not a confound when ablating other mechanisms\.

## Appendix A4AURATown Agent Profiles and Map

AURATownis a 60×\\times60\-grid simulation of five named agents living in twenty named locations \(homes, commerce, civic, parks\)\. The five agents \(Table[7](https://arxiv.org/html/2606.05557#A4.T7)\) and their starting locations \(Figure[5](https://arxiv.org/html/2606.05557#A4.F5)\) are fixed across all experiments\. Each agent has both public state \(location, current action, nearby agents\) visible to all queries and private state \(availability,emotional\_state,unspoken\_goal,beliefs\_about\_others\) accessible only via probe tools\.

010203040506001020304050601234567891011121314151617181920LWZCY5 agents:L Lin WeiZ Zhang HaoC Chen MeiY Liu YangW Wang Junhome1Lin Wei2Wang Jun3Zhang Hao4Chen Mei5Liu Yangcommerce6Tea House7Sunrise Cafe8Golden Wheat9Wellness Pharm\.10Chen’s Store11Art Gallerycivic12Temple13Town Hall14Town Library15AURA Academyopen / park16Town Park17Town Square18Comm\. Garden19Flower Garden20Riverside WalkFigure 5:AURATown— 60×\\times60 grid with 20 named locations at their real coordinates \(fromdemo/town/assets/town\_map\.json\), coloured by type \(home5,commerce6,civic4,open/park5\)\. Rings labelled L/Z/C/Y/W mark the five agents at their starting homes perdemo/town/agents\.py\. Public state \(location, action, nearby agents\) is visible in the scene snapshot; private state \(availability,emotional\_state,unspoken\_goal,beliefs\_about\_others\) is hidden and only retrievable via the probe tools listed in Table[8](https://arxiv.org/html/2606.05557#A5.T8)\.Table 7:Agent profiles inAURATown\.
## Appendix A5Probe Tool Descriptions

Factual grounding \(§[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)\) uses an eight\-tool environment registry \(Table[8](https://arxiv.org/html/2606.05557#A5.T8)\); each tool returns a structured fragment of the simulation state\. The implicit\-intent setup \(§[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\) uses a separate five\-tool scripted registry \(get\_all\_agents,get\_nearby\_agents,get\_agent\_plan,get\_agent\_private\_state,get\_agent\_belief\_about\) over a fixed scene snapshot rather than event\-history\. The two registries are disjoint by design — the factual\-grounding registry probes the world state through scene\-aware tools, the implicit\-intent registry probes individual agents’ public and private states\. The chat/demo deployment additionally exposes both registries, so a deployed agent has access to both base environment tools and private\-state/belief probes\.

Table 8:Environment tools available to the probe planner\.
## Appendix A6Implementation Details

This section records the LLM, memory, simulation, and infrastructure parameters that hold throughout the experiments\. All numbers report mean±\\pmstd over the three seeds\{42,123,456\}\\\{42,123,456\\\}unless explicitly noted\.

#### LLM Configuration\.

All LLM calls usegpt\-4o\-minivia the OpenAI API with:

- •Action decision: temperature 0\.7, default max tokens
- •Probe planning: temperature 0\.2, max tokens 200
- •Conversation generation: temperature 0\.8, max tokens 600
- •Reflection: temperature 0\.5, default max tokens
- •Importance scoring: temperature 0\.1, max tokens 50

#### Memory Configuration\.

Maximum 200 items per agent, retrieval weightswr=0\.3,wp=0\.4,wv=0\.3w\_\{r\}=0\.3,w\_\{p\}=0\.4,w\_\{v\}=0\.3, recency decayλ=0\.01\\lambda=0\.01, reflection thresholdθ=10\\theta=10\.

#### Simulation Parameters\.

30\-minute ticks, 6:00–23:00 day cycle, conversation cooldown 2 ticks, probe cooldown 2 ticks, movement speed 3 grid units/tick\.

#### Infrastructure\.

Python backend serving HTTP API on port 7861 with threading for concurrent state/chat/step requests\. React frontend with canvas\-based pixel\-art rendering, viewport camera system, and real\-time state updates\.

## Appendix A7Concrete Examples, Comparisons, and Diagrams

This appendix collects the figures and the proactive\-frameworks comparison referenced from the body but moved out for space\.

#### Motivating contrast \(Figure[6](https://arxiv.org/html/2606.05557#A7.F6)\)\.

A vanilla LLM andAURAon the same query in the same scene state\. Vanilla returns the literal location only\.AURAinfers an implicit availability need, probes the cafe scene for nearby agents and the target’s private state, and returns the location with a heads\-up alert that she is currently busy\.

![Refer to caption](https://arxiv.org/html/2606.05557v1/figures/vanilla-vs-aura.png)Figure 6:Motivating contrast on “Where is Lin Wei?”\. Vanilla \(left\) vs\.AURA\(right\) on the same scene state\.
#### IntentFrame example \(Figure[7](https://arxiv.org/html/2606.05557#A7.F7)\)\.

An exampleIntentFrameproduced by the LLM\-backedIntentInferreron a single AURATown query\. Six fields are emitted \(literal\_need,implicit\_need,gap,recommended\_probes,should\_alert,confidence\) and consumed by the Explore and Interact stages \(§[3\.3](https://arxiv.org/html/2606.05557#S3.SS3)\)\.

![Refer to caption](https://arxiv.org/html/2606.05557v1/figures/intentframe-thinking.png)Figure 7:TheIntentFrameproduced for a single user query\.
#### Long\-context prompt\-content sketch \(Figure[8](https://arxiv.org/html/2606.05557#A7.F8)\)\.

Figure[8](https://arxiv.org/html/2606.05557#A7.F8)sketches the token\-cost difference between packing the full environment state into the prompt andAURA’s selective probing on the sameAURATownscene\. Token counts from theAURATownprompt log; latency and FA from Tables[1](https://arxiv.org/html/2606.05557#S5.T1),[2](https://arxiv.org/html/2606.05557#S5.T2)\. This is an illustrative sketch, not a head\-to\-head experiment: the Static\-Context baseline \(FA0\.4500\.450\) is the closest in\-suite analogue but does not include private state or memory, so a true full\-state long\-context baseline is left for future work\.

\(A\) Long\-context model
Pack everything into the prompt; let the LLM attend\.
Prompt content \(per query\):
<scene snapshot\>∼\\sim180 tok<5 agents×\\timespublic state\>∼\\sim300 tok<5 agents×\\timesprivate state\>∼\\sim420 tok<recent events / memories\>∼\\sim650 tok<system \+ few\-shot\>∼\\sim500 tok↪\\hookrightarrow∼\\sim2050 prompt tokens / query
↪\\hookrightarrowO\(\#agents\)growth as the world scalesTrade\-off:pays for*all*state every query, even when the user’s literal need is one location lookup\. Long\-context attention dilution can lower precision on packed prompts\.Static\-Context baseline \(factual grounding\):FA0\.4500\.450,1\.81\.8s\.\(B\)AURA\(gap\-routed selective probing\)
Infer the gap; fetch only what closes it\.
Prompt content \(per query\):
<scene snapshot\>∼\\sim180 tok<IntentFrame: literal/implicit/gap\>∼\\sim210 tok<probe results×\\times1\.0∼\\\!\\sim\\\!2\.2\>∼\\sim120–280 tok<system \+ few\-shot\>∼\\sim500 tok↪\\hookrightarrow∼\\sim1010–1170 prompt tokens / query
↪\\hookrightarrowO\(\#probes\)=0–33, gap\-conditionedTrade\-off:pays one extra LLM call \(the IntentInferrer\) but skips fetching irrelevant private state\. Targeted probe returns are more easily decoded than a packed prompt\.AURAFull \(factual grounding\):FA0\.6400\.640,3\.93\.9s;AURAIntent \(implicit\-intent\):implicit\-score0\.8030\.803,13\.813\.8s\.Figure 8:AURAversus a long\-context approach onAURATown\. A long\-context model pays the full per\-agent public\+private\+memory cost per query \(panel A\);AURA’sIntentFrameroutes selective probing so the prompt grows with the inferred gap, not the number of agents \(panel B\)\. Token estimates are from theAURATownprompt log; FA / latency from Tables[1](https://arxiv.org/html/2606.05557#S5.T1),[2](https://arxiv.org/html/2606.05557#S5.T2)\. Static\-Context \(the long\-context lower\-bound in our suite\) reaches FA0\.4500\.450;AURAFull reaches0\.6400\.640\.
#### Proactive frameworks comparison \(Table[9](https://arxiv.org/html/2606.05557#A7.T9)\)\.

Table[9](https://arxiv.org/html/2606.05557#A7.T9)positionsAURArelative to recent proactive and tool\-using agent frameworks along three orthogonal axes \(decision target, trigger, selection\) and reports per\-query latency for frameworks we re\-implemented under our backbone\. “–” marks frameworks that address an orthogonal decision \(*when to assist*from sensor/behaviour patterns\) and were not re\-run on our factual benchmark\.

Table 9:WhereAURAsits among proactive / tool\-using agent frameworks\. “Trigger” = when the framework decides to act; “Selection” = how it decides what to fetch; “Latency” = mean per\-query wall\-clock when run on our 50\-query factual benchmark with backbonegpt\-4o\-mini, or “–” if not re\-implemented here\.FrameworkDecision targetTriggerSelectionLatencyContextAgent\(Yanget al\.,[2026](https://arxiv.org/html/2606.05557#bib.bib43)\)when to surface a hintsensor stream changemulti\-dim\. context extract\.–ProAgent\(Yanget al\.,[2025](https://arxiv.org/html/2606.05557#bib.bib44)\)predict user needAR\-glass sensor \+ profilehierarchical context fusion–PROBE\(Pasternaket al\.,[2025](https://arxiv.org/html/2606.05557#bib.bib49)\)search–identify–resolveidle / unspecified issuethree\-stage pipeline–ProAgentBench\(Tanget al\.,[2026](https://arxiv.org/html/2606.05557#bib.bib46)\)benchmark \(timing\+content\)N/A \(real\-user sessions\)N/A–ReAct\(Yaoet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib4)\)during\-reasoning actper LLM stepLLM picks per step6\.0 sReflexion\(Shinnet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib7)\)during \+ retryper step \+ reflectLLM picks per step20\.6 sPlan\-and\-Solve\(Wanget al\.,[2023b](https://arxiv.org/html/2606.05557#bib.bib58)\)plan\-then\-executeonce, from surface querysurface\-query plan5\.2 sAURA\(ours\)what to fetchpost\-query, gap\-routedgap→\\toprobe set3\.9 s

## Appendix A8Experimental Setup \(Full\)

#### Baselines\.

We compareAURAagainst six baselines on the factual\-grounding benchmark \(Table[1](https://arxiv.org/html/2606.05557#S5.T1)\) and four on the implicit\-intent benchmark \(Table[2](https://arxiv.org/html/2606.05557#S5.T2)\):

1. 1\.Vanilla LLM: directgpt\-4o\-minigeneration with no environmental context \(factual grounding\)\.
2. 2\.Static Context: LLM receives a packed scene snapshot \(location, time, nearby agents\) but no proactive information gathering \(factual grounding\)\.
3. 3\.ReAct Agent\(Yaoet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib4)\): Thought–Action–Observation interleaved*during*reasoning\. Same eight base environment tools asAURAon factual grounding; same five\-tool scripted registry on the implicit\-intent benchmark\.
4. 4\.Reflexion\(Shinnet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib7)\): ReAct \+ self\-reflection, up to 2 retry rounds \(factual grounding\)\.
5. 5\.Plan\-and\-Solve\(Wanget al\.,[2023b](https://arxiv.org/html/2606.05557#bib.bib58)\): Phase\-1 plan\-from\-surface\-query, Phase\-2 execute, Phase\-3 synthesise \(factual grounding \+ implicit\-intent benchmark\)\.
6. 6\.AURA\(No Probe\):AURAarchitecture with the probe budget forced to0\(factual\-grounding ablation\)\.
7. 7\.AURA\(Full\): complete pipeline with IntentFrame\-derived budget and tool selection\.

We do not re\-run Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib1)\)or ContextAgent\(Yanget al\.,[2026](https://arxiv.org/html/2606.05557#bib.bib43)\)as paired baselines: Generative Agents’ passive observation string is subsumed by our Static Context; ContextAgent targets a different decision \(*when to assist*from sensor streams\) and is discussed as related work in Appendix[A7](https://arxiv.org/html/2606.05557#A7.SS0.SSS0.Px4)rather than an in\-suite comparator\.

#### Models and metrics\.

Backbonegpt\-4o\-mini, temperature 0\.7;gpt\-4o\-miniLLM\-as\-judge at temperature 0\.1\(Zhenget al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib50)\), combined with a rule\-based pre\-filter on location and time consistency\. Judge and agent share the same model family \(gpt\-4o\-mini\), which reduces judge–policy independence; we mitigate by reporting a strict precision rescore \(Appendix[A9](https://arxiv.org/html/2606.05557#A9)\) that uses only the judge’s CORRECT/CONTRADICTED claim\-level classifications and drops the soft completeness component\. Metrics:*Grounding Accuracy*\(GA, 5 sub\-dimensions, inspired by AgentBench\(Liuet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib11)\)\),*Factual Accuracy*\(FA, 50 environment\-grounded queries in 5 categories\),*SOTOPIA Social Score*\(7 dimensions fromZhouet al\.\([2024](https://arxiv.org/html/2606.05557#bib.bib10)\)\),*Context Utilization*\(CU\),*Latency*\.

## Appendix A9Strict Precision Rescore for Factual Grounding \(Robustness Check\)

The factual accuracy \(FA\) reported in Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)uses an LLM\-as\-judge whose final score combines a hard precision component \(correct/\(correct\+contradicted\) over the judge’s claim\-level classifications, weight0\.70\.7\) with a soft*completeness*component \(a 0/0\.5/1 rubric for whether the response addressed the question, weight 0\.3\)\. Soft components are sensitive to phrasing and judge mood\.

We rescore the same per\-query records with strict precision only, drop the completeness term, and additionally report the per\-query*hallucination rate*\(fraction of \(query, seed\) cells with≥1\\geq 1contradicted claim\) and*perfect\-response rate*\(0 contradicted AND≥1\\geq 1correct\)\. The rescore script and output JSON are listed in Appendix[A24](https://arxiv.org/html/2606.05557#A24)\.

Table 10:Strict rescore over the same 50 factual\-grounding queries×\\times3 seeds \(per\-query, per\-seed cells\)\.*StrictP*= correct/\(correct \+ contradicted\) excluding cells where the judge produced only unverifiable claims\.*Halluc\. %*= fraction of cells with≥1\\geq 1contradicted claim\.*Perfect %*= 0 contradicted AND≥1\\geq 1correct\.*Lenient FA*= the original 0\.7⋅\\cdotP \+ 0\.3⋅\\cdotC from Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)for comparison\.Three honest takeaways relative to the lenient table in Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1):

- •vs\. Vanilla:the architectural\-effect headline holds and gets larger \(\+0\.609\+0\.609strict precision,p<10−4p<10^\{\-4\}\)\.
- •vs\. ReAct \(the fair\-tools baseline\):the strict precision gap is\+0\.090\+0\.090\(p=0\.064p\{=\}0\.064, query\-level pairedtt\-test,n=50n\{=\}50\) — not significant\.AURAalso has a higher hallucination rate than ReAct \(66\.7%66\.7\\%vs\.51\.3%51\.3\\%,Δ=\+15\.3\\Delta\{=\}\+15\.3pp,p=0\.013p\{=\}0\.013\)\. Reading these two together, the architectural pipeline’s contribution*above*a tool\-using ReAct baseline is small under strict scoring\.
- •vs\.AURANoProbe:strict precision rises by\+0\.062\+0\.062\(p=0\.032p\{=\}0\.032\); the lenient version of this contrast wasp=0\.299p\{=\}0\.299\. The probing mechanism is more clearly supported under strict scoring than under the soft FA, while the overall architectural margin shrinks\.

The per\-category strict precision \(Table[11](https://arxiv.org/html/2606.05557#A9.T11)\) localises the gains:AURA\(Full\) wins on temporal \(0\.8920\.892\) and memory \(0\.7400\.740\) categories and loses to ReAct on social and spatial\. The original “social win for probing” from the lenient analysis does not survive strict scoring; we update Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)’s scope condition accordingly in the discussion\.

Table 11:Per\-category strict precision \(mean across query\-seed cells,n≤30n\{\\leq\}30per cell\)\.
## Appendix A10Privacy\-Sensitive Distractor Slice

The factual\-grounding primary FA metric rewards factual completeness but does not penalise unnecessary access\. We therefore authored a 30\-query factual slice whose gold answers require public facts only\. Each query carries aforbidden\_toolslist over high\-disclosure tools \(memory\.recent,world\.events\_recent,agent\.plan\) plus query\-specific extras; a violation is any fired forbidden tool in the query–seed cell\. We run the same three seeds as the factual\-grounding benchmark\.

Table 12:Privacy\-sensitive factual slice \(30 queries×\\times3 seeds\)\.*Viol\. %*is the fraction of query–seed cells where any forbidden tool fired\.Against GapRouted, Fixed\-Probe gains\+0\.080\+0\.080FA only marginally \(p=0\.056p\{=\}0\.056\) while adding\+100\+100pp forbidden\-tool violations\. Plan\-and\-Solve and ReAct are tied in FA with GapRouted \(p=0\.856p\{=\}0\.856andp=0\.655p\{=\}0\.655\) but incur\+78\.9\+78\.9pp and\+25\.6\+25\.6pp violations\. This slice makes the factual\-grounding tradeoff explicit: saturated access is often accurate, but it violates stated access constraints by construction; gap routing is not FA\-dominant, but it operates on the low\-violation, low\-probe side of the Pareto frontier\.

## Appendix A11Cost and Latency Across Conditions

The Pareto framing in Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)ranks conditions on access cost \(probes\) and disclosure, not wall\-clock\. Table[13](https://arxiv.org/html/2606.05557#A11.T13)reports median and mean per\-query latency from the same multi\-seed runs so the wall\-clock side of the tradeoff is auditable\. Medians guard against rare upstream\-API timeouts that inflate means \(most pronounced for the implicit\-intent v1*tom*row, where one604604s API hiccup on seed456456pulls mean to13\.8513\.85s while median stays at5\.315\.31s\)\.

Table 13:Wall\-clock per query \(seconds; median / mean\)\. Factual columns aggregate the 50q factual benchmark; Privacy is the 30q distractor slice; Implicit\-Intent v1 is the 25q implicit\-need set under clean prompt\. Probe count is the per\-condition mean tool\-call count \(factual / Privacy use the eight\-tool registry; implicit\-intent uses the five\-tool probe registry\)\.AURAGapRouted pays an extra IntentInferrer LLM round\-trip \(≈2\\approx 2s\) over a one\-shot baseline; the cost does not scale with the probe count\.FactualPrivacy sliceImplicit\-Intent v1ConditionLatProbesLatProbesLatProbesVanilla LLM2\.142\.14/2\.282\.2801\.831\.83/2\.012\.010——Static Context1\.641\.64/1\.771\.7701\.541\.54/1\.591\.590——Literal————1\.231\.23/1\.291\.290NoIntent \(ReAct\)5\.105\.10/6\.016\.011\.051\.053\.183\.18/4\.874\.871\.871\.871\.951\.95/2\.202\.201\.051\.05Plan\-and\-Solve4\.754\.75/5\.195\.194\.644\.644\.464\.46/4\.514\.514\.164\.16——Reflexion16\.4516\.45/20\.6420\.642\.152\.15————Fixed\-Probe2\.372\.37/3\.233\.238\.008\.001\.471\.47/1\.651\.658\.008\.00——AURAGapRouted4\.08\\mathbf\{4\.08\}/4\.34\\mathbf\{4\.34\}1\.40\\mathbf\{1\.40\}3\.97\\mathbf\{3\.97\}/3\.99\\mathbf\{3\.99\}0\.73\\mathbf\{0\.73\}5\.31\\mathbf\{5\.31\}/13\.85\\mathbf\{13\.85\}†\\dagger1\.48\\mathbf\{1\.48\}

†\\daggerMedian is more representative; mean is inflated by one604604s upstream\-API outlier across the7575query–seed cells\.

Two observations qualify the Pareto picture\. First,AURAGapRouted is*not*the fastest condition despite firing the fewest probes: the IntentInferrer is a single extra LLM call \(≈2\\approx 2s ongpt\-4o\-mini\), so GapRouted’s median latency sits above Fixed\-Probe \(4\.084\.08vs\.2\.372\.37s on factual grounding;3\.973\.97vs\.1\.471\.47s on the privacy slice\)\. The probe\-count win does not translate into a wall\-clock win at small budgets\. Second, Reflexion’s median latency \(16\.516\.5s\) is the regime outlier; its retry loop pays the cost without buying accuracy\. We treat latency as a reportable axis rather than a paper claim: the cost\-of\-selectivity story holds on probe count and disclosure, not on wall\-clock\.

## Appendix A12Prompt Ablation for IntentFrame Calibration

We audit whether theIntentFramegain comes from benchmark\-overlapping few\-shot examples or from calibrated gap estimation\. Table[14](https://arxiv.org/html/2606.05557#A12.T14)compares three prompt variants on the same 25 implicit\-intent queries and three seeds\. The leaked row uses the original benchmark\-overlapping examples and is reported only for diagnosis; the clean row is the final system; the no\-few\-shot row keeps the same rubric but removes all examples\.

Table 14:Prompt ablation for theIntentFramecontroller\.pqp\_\{q\}is the query\-level paired test for Intent vs\. NoIntent\.Clean few\-shot is only0\.0370\.037below the leaked prompt, so the main effect is not driven by memorising names or locations\. In contrast, removing examples reduces Intent to0\.6770\.677and makes the Intent\-vs\.\-NoIntent gain non\-significant\. Mechanistically, the no\-few\-shot prompt underestimates the gap: mean inferred gap drops from0\.5890\.589to0\.4760\.476, and high\-gap cells \(g≥0\.60g\{\\geq\}0\.60\) drop from43/7543/75to22/7522/75\. The few\-shot examples are therefore load\-bearing calibration for gap\-to\-budget routing\.

## Appendix A13Cross\-Backbone Intent\-vs\.\-NoIntent \(Robustness\)

To check that the Intent\-vs\.\-NoIntent gain reported ongpt\-4o\-miniis not specific to one backbone, we re\-ran the 25 implicit\-intent queries on three additional production\-grade LLMs \(each via its vendor’s official API; judge fixed atgpt\-4o\-minito keep the scoring rubric constant\)\. Table[15](https://arxiv.org/html/2606.05557#A13.T15)summarises the result\. Thegpt\-4o\-minirow is the clean calibrated prompt used in the main paper; the other backbones are single\-seed robustness probes\. Three of four backbones reproduce the gain;gemini\-2\.5\-flashregresses, but the regression is explained by JSON\-schema parse failure on23/2523/25IntentFramecalls and silent fallback to the deterministic heuristic — a format\-compliance boundary\.

Table 15:Cross\-backbone Intent\-vs\.\-NoIntent contrast on implicit score \(25 implicit\-intent queries, query\-level paired comparison with cluster bootstrap onquery\_id\)\. Thegpt\-4o\-minirow is the clean calibrated prompt and averages over 3 seeds; the other rows are single\-seed robustness probes\. Three of four backbones reproduce the gain;gemini\-2\.5\-flashfails the JSON parser on23/2523/25IntentFramecalls \(run log\_run\_gemini25flash\.log\) so its row reflects heuristic\-fallback under a Gemini tag\.
## Appendix A14Expanded Implicit\-Intent v2 Details

To test whether the 25\-query scene is over\-specialised, we authored an expanded implicit\-intent v2 set with 4 scene snapshots×\\times5 subcategories×\\times5 queries \(100100queries;300300scored cells per condition\)\. Scene A preserves the original 25 queries; scenes B–D vary location, time of day, agent rosters, private states, and stale belief\-vs\-truth mismatches\. Each v2 query recordsgold\_required\_toolsandforbidden\_tools; all second\-order queries forbid directget\_agent\_private\_stateaccess because the correct evidence is the believer’s recorded belief, not the target’s ground truth\.

#### Subcategory examples\.

Representative surface queries by subcategory:*availability*“where is Lin Wei?” \(implicit: “is she free?”\);*mood*“how is Chen Mei today?” \(implicit: “is she in a receptive emotional state?”\);*appropriateness*“is now a good time to invite Lin Wei for coffee?” \(requires integrating schedule and private state\);*latent\_goal*“what is Lin Wei up to?”;*second\_order*“does Lin Wei think Zhang Hao is free?” \(the correct answer must report the believer’s recorded belief, not the target’s ground truth

#### Inter\-annotator agreement\.

Two independent annotators \(computer\-science graduate students, distinct from the authors and naive to the AURA architecture\) re\-labelled the 25\-query pilot set under the 5\-subcategory definitions given in the task instructions\. They reached Cohen’sκ=0\.61\\kappa\{=\}0\.61\(substantial under Landis\-Koch; raw agreement68%68\\%,17/2517/25\)\. All 8 disagreements concentrate on two boundaries: \(i\)*mood*vs\.*appropriateness*/*availability*\(4/8; e\.g\., “Does Zhang Hao look busy?” admits both an availability reading and a mood reading\), and \(ii\)*appropriateness*vs\.*literal*/*availability*\(3/8; e\.g\., “Can I ask Wang Jun for a favor right now?” splits between a literal\-permission reading and a context\-aware appropriateness reading\)\. Per\-class agreement is highest on*second\_order*\(24/25 between annotators\) and*latent\_goal*\(24/25\); collapsing*mood*into a single*context\-aware availability*super\-category yieldsκ=0\.68\\kappa\{=\}0\.68\. Raw label dumps are released alongside the queries\.

Table 16:Expanded implicit\-intent v2 check \(100 queries×\\times3 seeds,gpt\-4o\-mini\)\. Scores are implicit\-need coverage\. The final two rows reportAURAIntent minus NoIntent;ppvalues are paired tests over query\-seed cells\.
#### Scene\-level read\.

Three of four scenes are clearly positive \(A/B/C\), and the post\-event night scene D ties: agents are spatially distributed in D and the public state already telegraphs availability, so the gap mechanism has nothing to add\.

#### Subcategory breakdown \(Intent vs\. NoIntent, cell\-leveln=60n\{=\}60per cat\)\.

*availability*\+0\.29\+0\.29\(p=2\.7×10−11p\{=\}2\.7\\times 10^\{\-11\}\),*appropriateness*\+0\.11\+0\.11\(p=8\.2×10−4p\{=\}8\.2\\times 10^\{\-4\}\),*mood*\+0\.07\+0\.07\(p=1\.7×10−3p\{=\}1\.7\\times 10^\{\-3\}\),*second\_order*−0\.02\-0\.02\(p=0\.32p\{=\}0\.32, NS\) after a synthesis\-prompt fix that drops the public\-state dump on belief\-state queries and adds a strict “report belief, not actual state” instruction, and*latent\_goal*−0\.09\-0\.09\(p=6\.2×10−4p\{=\}6\.2\\times 10^\{\-4\}, residual deficit acknowledged in Limitations\)\.

#### Cross\-backbone v2\.

Re\-running the 100\-query set onclaude\-haiku\-4\.5\(seed 42 only, via OpenRouter\) reproduces the gain at*larger*magnitude: Intent0\.8760\.876vs\. NoIntent0\.7900\.790, pairedΔ=\+0\.086\\Delta\{=\}\{\+\}0\.086,p=3\.7×10−3p\{=\}3\.7\\times 10^\{\-3\}\. The scene\-D tie pattern replicates \(Δ=\+0\.008\\Delta\{=\}\{\+\}0\.008,p=0\.88p\{=\}0\.88, NS\), confirming the night\-scene null is a property of the regime rather than a single\-backbone artifact\.

## Appendix A15Factual\-Grounding Per\-Category Paired Contrast

The aggregateAURAFull vs\.AURANo\-Probe contrast on factual grounding is a near\-null \(\+0\.038\+0\.038FA,p=0\.060p\{=\}0\.060paired query\-level\), but this aggregate hides a heterogeneous per\-category structure\. Splitting the 50 queries into 5 categories of 10 \(spatial / social / temporal / memory / planning\), Table[17](https://arxiv.org/html/2606.05557#A15.T17)shows that probing produces a significant FA improvement only on the*social*category\. The other four categories are saturated by the Perceive/Scene channel, so probing has nothing to add\. This is the empirical scope condition for proactive probing referenced in the main text\.

Table 17:Per\-category pairedtt\-test of proactive probing \(AURAFull vs\.AURANo\-Probe\),N=10N\{=\}10queries per category per seed, 3 seeds\. Probing produces a significant effect only on the*social*category\.
## Appendix A16Routine\-Action Grounding \(Null Result, Full Table\)

The routine\-grounding check measures whether richer environmental access changes Grounding Accuracy on routine daily simulation\. The protocol is 100 simulation steps×\\times5 conditions×\\times3 seeds=1,500=1\{,\}500judgments per condition; the metric averages four GA sub\-dimensions \(location consistency, time appropriateness, social awareness, plan adherence\)\. Memory utilisation saturates at≈1\.0\\approx 1\.0for every method and is omitted from Table[18](https://arxiv.org/html/2606.05557#A16.T18)\. The pairedtt\-test column reports two\-sidedppvs\. Vanilla\. All five methods fall within0\.0240\.024absolute GA spread; every contrast is non\-significant\. We report this null openly: when most actions are trivially grounded, GA cannot distinguish architectures\.

Table 18:Grounding Accuracy \(GA\) on 100\-step routine daily simulation, 3 seeds, mean±\\pmstd\. Dimension scores: location consistency, time appropriateness, social awareness, plan adherence\.All pairwise pairedtt\-tests vs\. Vanilla givep\>0\.5p\>0\.5; a non\-parametric Wilcoxon signed\-rank test gives the same conclusion\. The GA spread across five different architectures is0\.0240\.024, well inside per\-condition seed variance\. This is evidence the metric is saturated on this workload, not that probing fails; factual grounding \(Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)\) reaches a9\.1×9\.1\\timesgap on queries that stress grounding\.

## Appendix A17Additional Diagnostic Checks

### A17\.1Component Ablation \(Saturated\-Workload Null\)

We ablate each pipeline component on a routine\-action workload of100100simulation steps and5050chat queries per configuration, three seeds, source dataevaluation/results/rq3\_ablation\_study\_multiseed\.json\. Table[19](https://arxiv.org/html/2606.05557#A17.T19)gives the multi\-seed mean GA / FA / latency deltas relative toAURAFull\. \(An earlier draft of this table reported single\-seed numbers measured at 20 steps and 20 queries; we replace it here with the full multi\-seed configuration to match the actual run\.\)

Table 19:Component ablation, multi\-seed mean \(Δ\\Deltavs\.AURAFull\); GA/FA absolute, latency in seconds\. Re\-computed fromrq3\_ablation\_study\_multiseed\.json\(3 seeds, 100 sim steps, 50 chat queries per config\)\.#### Interpretation\.

Component\-level deltas on the routine\-action workload are uniformly small \(\|Δ​GA\|≤0\.020\|\\Delta\\text\{GA\}\|\\leq 0\.020,\|Δ​FA\|≤0\.028\|\\Delta\\text\{FA\}\|\\leq 0\.028\)\. Removing the probing component does*not*hurt GA or FA on this workload; it saves1\.31\.3s of latency\. The same null pattern that drives the routine\-grounding saturation result \(Section[A16](https://arxiv.org/html/2606.05557#A16)\) drives the component\-ablation null: routine actions like “sleeping at home at 6 AM” pass any reasonable grounding check, and the LLM\-as\-judge factual scoring with completeness has a±0\.03\\pm 0\.03noise floor that swamps the per\-component contribution\. Memory has the largest negative GA effect \(−0\.017\-0\.017\) but with FA actually slightly higher when memory is removed — consistent with the same noise floor, not a clean component effect\. The picture changes on the factual\-grounding chat workload \(Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1)\), where the same architecture moves FA from0\.070\.07\(Vanilla\) to0\.640\.64\(Full\); the bottleneck the ablation in this table fails to expose is the same one the routine\-grounding check fails to expose: action\-grounding on routine days does not stress the structured\-environment\-access channel\.

### A17\.2Emergent Social Behaviours \(SOTOPIA\)

200\-step multi\-agent simulation, 36 conversations evaluated on SOTOPIA’s 7 dimensions\(Zhouet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib10)\)\. Overall quality 7\.87/10, strongest dimensions*goal*\(9\.5\) and*believability*\(9\.0\); 44 emergent behaviours across 4 categories: collaboration \(32\), routine adaptation \(7\), conflict resolution \(4\), group formation \(1\)\.

Table 20:SOTOPIA dimension averages \(AURAfull run, 200 sim steps\)\.![Refer to caption](https://arxiv.org/html/2606.05557v1/figures/rq4-sotopia-radar.png)Figure 9:SOTOPIA 7\-dimension social evaluation \(AURAfull, 200\-step run\)\. Each axis is normalised to\[0,1\]\[0,1\]of its native range \(shown next to each label\); raw values annotated\. Strongest dimensions:*believability*\(9\.0/109\.0/10\),*goal*\(9\.5/109\.5/10\); weakest:*social rules*\(−2\.13\-2\.13\) and*secret*\(−0\.67\-0\.67\) \(both negative\-only scales\)\. The shape is*lopsided\-positive*:AURAagents excel on individual\-level dimensions and underperform on multi\-agent normative dimensions, consistent with the per\-agent\-mechanism scope of the design\.The collective\-behaviour analysis of these numbers is in Appendix[A26](https://arxiv.org/html/2606.05557#A26)\(Section[A26\.1](https://arxiv.org/html/2606.05557#A26.SS1)\)\.

### A17\.3Human Evaluation

We collected pairwise A/B annotations from 8 independent raters on the 50 chat scenarios, four dimensions each \(*response helpfulness*,*environmental awareness*,*agent believability*,*factual accuracy*\) on a 5\-point Likert scale, side\-randomised and blinded to system identity\. This givesN=400N\{=\}400paired query\-rater observations per dimension\. Both responses on each item were generated with the fixed system prompt at the sameAURATownsimulation tick \(warmup 10 steps, seed 42\); raters saw the query, the asking agent’s name, the category, and the two anonymous responses but not the underlying simulation scene state\.

Table 21:Human eval primary analysis: rater\-aggregated paired Wilcoxon\. Each rater’sΔ\\Deltais averaged across the 50 scenarios first; the test is then overN=8N\{=\}8paired rater\-means per dimension\. This is conservative relative to a 400\-cell paired test, which would treat repeated\-measures within a rater as independent\. The 95% CI column is a 5000\-resample cluster bootstrap onrater\_idover the rawΔ\\Deltavalues\.![Refer to caption](https://arxiv.org/html/2606.05557v1/figures/rq5-violin.png)Figure 10:Per\-raterΔ\\Delta\(AURA−\-Vanilla\) across the 4 dimensions \(N=8N\{=\}8raters, 50 scenarios each\)\. Each coloured dot is one rater’s meanΔ\\Deltaacross the 50 scenarios; the grey violin is the pooledΔ\\Deltadistribution; the black diamond is the pooled mean\. Letters identify raters \(MMis the one whoseΔ\\Deltafalls below zero on factual accuracy and is the basis for theN=7N\{=\}7sensitivity analysis\)\. 7/8 raters’ means lie above zero on every dimension\.#### Headline\.

AURAreceives higher ratings than Vanilla LLM on all four dimensions under the conservative rater\-aggregated test \(N=8N\{=\}8paired rater\-means per dimension; Wilcoxonp=0\.017p\{=\}0\.017each, sign 7\-1\-0 for every dimension\)\. All four cluster\-bootstrap CIs on the per\-raterΔ\\Deltaexclude zero\. The largest gap is on*environmental awareness*\(Δ=\+1\.86\\Delta\{=\}\+1\.86, CI\[\+1\.08,\+2\.62\]\[\+1\.08,\+2\.62\]\), consistent with the environment\-mediated design\. The cell\-level paired test on the 400 \(rater×\\timesscenario\) cells givesp<10−4p<10^\{\-4\}but treats within\-rater repeated measures as independent, so we report the rater\-aggregatedp=0\.017p\{=\}0\.017as the primary statistic\. Fig\.[10](https://arxiv.org/html/2606.05557#A17.F10)visualises the per\-rater means against the pooled distribution\.

#### Per\-rater profile \(transparency\)\.

Per\-rater meanΔ\\Deltaranged from−0\.30\-0\.30to\+3\.18\+3\.18across the 8 raters; per\-rater AURA\-better rates were\{24\.0,69\.0,81\.5,82\.5,85\.0,85\.0,98\.5,100\.0\}%\\\{24\.0,69\.0,81\.5,82\.5,85\.0,85\.0,98\.5,100\.0\\\}\\%\. Seven of eight raters preferredAURAon aggregate; one rater \(24% AURA\-wins,Δ=−0\.30\\Delta\{=\}\-0\.30\) reported in post\-task debrief that they had penalised plausible\-but\-unverifiable specifics inAURA’s responses as factual errors\. We retain that rater in the primary analysis to avoid post\-hoc exclusion bias and discuss the implication under “Methodological limitations” below\.

#### Direction\-agreement at the cell level\.

For each of the 200 \(scenario, dimension\) cells, we count raters preferringAURA, Vanilla, or scoring tied\. Defining “consensus” as≥6\\geq 6of 8 raters agreeing on direction, we find:𝟎/𝟐𝟎𝟎\\mathbf\{0/200\}cells have a Vanilla consensus,𝟏𝟒𝟖/𝟐𝟎𝟎\\mathbf\{148/200\}\(74%74\\%\) have anAURAconsensus, and5252are split\. Per\-dimensionAURA\-consensus rates: helpfulness76%76\\%, env\-awareness80%80\\%, believability72%72\\%, factual\-accuracy68%68\\%\. Average per\-cellAURA\-preference rate:78\.2%78\.2\\%\.

#### Inter\-rater reliability\.

Pooled across all dimensions, Krippendorff’sαord\\alpha\_\{\\text\{ord\}\}on the raw 1–5 scores is0\.430\.43\(per\-dimension range0\.340\.34–0\.470\.47\), indicating moderate agreement on absolute quality\. Item\-levelα\\alphaon the preference differenceΔ\\Deltais much lower \(αord∈\[0\.03,0\.17\]\\alpha\_\{\\text\{ord\}\}\{\\in\}\[0\.03,0\.17\]per dimension\) because raters differ systematically in scale\-use \(one rater uses 5/1 polar ratings, others use 4/2\)\. The*direction*of preference is nonetheless highly consistent \(above:74%74\\%AURA\-consensus,0%0\\%Vanilla\-consensus\)\. We report all three numbers and treat the pattern as: agreement on direction is strong, agreement on magnitude is weak, and headline effect sizes pool through this magnitude variance\.

#### Sensitivity analysis\.

Excluding the one reverse\-preference rater post\-hoc yieldsN=7N\{=\}7and strengthens all four dimensions: helpfulnessΔ=\+1\.85\\Delta\{=\}\+1\.85,dz=1\.45d\_\{z\}\{=\}1\.45; environmental awarenessΔ=\+2\.15\\Delta\{=\}\+2\.15,dz=1\.62d\_\{z\}\{=\}1\.62; believabilityΔ=\+1\.86\\Delta\{=\}\+1\.86,dz=1\.41d\_\{z\}\{=\}1\.41; factual accuracyΔ=\+1\.65\\Delta\{=\}\+1\.65,dz=1\.37d\_\{z\}\{=\}1\.37\. We report this only as a robustness check; the headline numbers in Table[21](https://arxiv.org/html/2606.05557#A17.T21)are theN=8N\{=\}8primary analysis\.

#### Per\-category structure\.

The category ordering is consistent with the factual\-grounding per\-category analysis: temporal \(\+1\.83\+1\.83avg\) and memory \(\+1\.77\+1\.77\) at the top, spatial \(\+1\.29\+1\.29\) and planning \(\+1\.37\+1\.37\) lower\. Social \(\+1\.77\+1\.77\) matches the factual\-grounding finding that proactive probing helps most where residual environmental uncertainty is concentrated\.

#### Independent fabrication scan\.

Independently of the human ratings, we manually scanned all 50AURAresponses for fabricated proper\-name entities \(locations or characters absent from theAURATownsetup, which has 20 named locations and 5 named characters\)\. Two scenarios contain explicit fabrications: scenario id 6 lists “Bookstore, Art Supply Store, Clothing Boutiques, Craft Stores, Gift Shops” \(none exist among the 20 locations\); scenario id 7 references “Main Street” and “The Cozy Corner Cafe” \(AURATownhas neither that street nor a cafe by that name; the only cafe is*Sunrise Cafe*\)\. The other 48 responses cite only roster\-real locations and characters, giving a static\-entity fabrication rate of2/50=4%2/50=4\\%\. The scan is reproducible from the released response set against the canonical roster indemo/town/assets/town\_map\.json\.

#### Methodological limitations\.

\(1\) Sample size\.N=8N\{=\}8is small relative to managed crowdsourcing studies\. The consistency of direction across raters and the large effect sizes mitigate but do not substitute for a larger study; primary statistics should be read as evidence of a large effect with substantial uncertainty around the precise magnitude\.\(2\) Recruitment\.Raters were recruited individually rather than through a managed platform \(Prolific, MTurk\)\.\(3\) Dynamic\-state factual accuracy is incompletely measured\.The form did not display simulation scene state at query time, so raters’ factual\_accuracy judgments mix verifiable static\-entity checks \(catchable from the on\-form roster\) with trust\-prior on dynamic\-state claims \(agent positions, current activities\)\. The independent fabrication scan above bounds the static rate at4%4\\%; precise quantification of dynamic\-state error requires deterministically recapturing scene state at each query’s generation time, which is future work\.\(4\) IRR is moderate\.α≈0\.4\\alpha\\approx 0\.4on raw scores reflects scale\-use heterogeneity across raters; we publish all per\-rater data so this variance is auditable\.

### A17\.4Probe\-Budget Sweep \(Full Multi\-Seed\)

We vary the probe budgetBBfrom 0 to 5 and measure GA and latency per step \(500 GA judgments per budget per seed, averaged across 3 seeds\{42,123,456\}\\\{42,123,456\\\}\)\. Fig\.[11](https://arxiv.org/html/2606.05557#A17.F11)visualises the GA\-vs\-latency Pareto frontier\.

011223344550\.840\.840\.860\.860\.880\.88B⋆=1B^\{\\star\}\{=\}1Probe budgetBBGrounding Accuracy \(mean±\\pmstd, 3 seeds\)GAPareto frontier202030304040Latency per step \(s\)Latency

Figure 11:Probe\-budget sweep, 3 seeds, 500 GA judgments per\(B,seed\)\(B,\\text\{seed\}\)cell\. Pareto frontier:\{B=0,B=1\}\\\{B\{=\}0,B\{=\}1\\\}\(red rings\);B⋆=1B^\{\\star\}\{=\}1adds\+0\.012\+0\.012GA at\+4\.1\+4\.1s, everyB≥2B\\geq 2is dominated\. The same single probe buys\+13\.7\+13\.7pp implicit score in the social sub\-regime \(Table[5](https://arxiv.org/html/2606.05557#A2.T5)\); routine GA is saturated\.Table 22:Probe budget sweep, 3 seeds, 500 judgments perBBper seed\. Pareto frontier: onlyB∈\{0,1\}B\\in\\\{0,1\\\}\.The Pareto frontier contains exactly two points,\{B=0,B=1\}\\\{B\{=\}0,B\{=\}1\\\};B=1B\{=\}1is the peak\.B=0→B=1B\{=\}0\\to B\{=\}1adds\+0\.0122\+0\.0122GA \(\+1\.4%\+1\.4\\%relative\) at\+4\.1\+4\.1s latency; everyB≥2B\\geq 2is Pareto\-dominated\. An earlier single\-seed pilot reportedB∗=2B^\{\*\}\{=\}2; the 3\-seed replication withdraws that claim\. On the routine\-grounding metric the curve is consistent with a monotonically saturating information channel rather than a clean crossover: one probe fills the scene snapshot, subsequent probes contribute little\. The same single probe buys\+13\.7\+13\.7pp implicit score in the social sub\-regime \(Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2), Table[5](https://arxiv.org/html/2606.05557#A2.T5)\), confirming that the budget\-sweep’s small absolute gain is a feature of the metric, not the mechanism\.

### A17\.5Judge Disagreement

We quantify judge disagreement between the rule\-based pre\-filter and the LLM judge across 7,500 multi\-seed judgments in Table[23](https://arxiv.org/html/2606.05557#A17.T23)\. Disagreement varies by dimension: time appropriateness shows the highest rate \(61\.9%61\.9\\%, rule stricter in99\.5%99\.5\\%\), location consistency48\.4%48\.4\\%\(92\.3%92\.3\\%rule\-stricter\), plan adherence32\.0%32\.0\\%\(83\.0%83\.0\\%LLM\-stricter\), social awareness13\.5%13\.5\\%\(98\.8%98\.8\\%LLM\-stricter\), memory utilisation0\.0%0\.0\\%\. We resolve disagreements by taking the stricter score \(conservative grounding estimate\)\.

Table 23:Rule\-based vs\. LLM judge disagreement across 7,500 grounding judgments \(3 seeds×\\times5 conditions×\\times500 per condition\)\.

## Appendix A18Heuristic vs\. LLM IntentInferrer \(Backend Ablation\)

TheIntentFramepipeline of Section[3\.3](https://arxiv.org/html/2606.05557#S3.SS3)admits two backends: a deterministicHeuristicIntentInferrer\(rule\-based gap estimation from a small vocabulary of social/private\-state markers\) andLLMIntentInferrer\(gpt\-4o\-mini producing the frame as JSON\)\. The main implicit\-intent run uses the LLM backend\. We isolate the inferrer’s contribution by running the SAME 25 implicit\-intent queries×\\times3 seeds with the heuristic backend; all other plumbing \(gap\-to\-budget map, directed probe loop, judge\) is identical\.

Table 24:Per\-subcategory implicit score: heuristic\-backend vs\. LLM\-backendAURAIntent \(25 queries×\\times3 seeds,N=15N\{=\}15per cell\)\. The heuristic collapses to≈\\approxLiteral performance on*availability*and*latent\_goal*because its surface\-keyword gap estimator does not fire on lexically\-decoupled queries; the LLM backend recovers the gap and routes the appropriate probes\.#### Reading\.

The heuristic recovers\+0\.18\+0\.18over Literal on aggregate but lags the clean calibrated LLM backend by−0\.44\-0\.44\. Per\-subcategory: the heuristic stays near Literal on availability \(0\.270\.27vs\.0\.290\.29\) and ties Literal on latent\_goal \(0\.000\.00vs\.0\.000\.00\) because its trigger vocabulary \(“available”, “mood”, “appropriate”, “up to”\) does not match the surface of those query classes \(“where is X?”, “what is X up to?”\); the gap estimator returns0and the heuristic falls through to the literal answer\. On second\_order it does well \(0\.770\.77\) because “thinks”, “believes”, and “perspective” do appear in the trigger set\. The contrast quantifies what the LLM backend buys:\+0\.44\+0\.44aggregate at7\.3×7\.3\\timesthe heuristic latency, with the LLM doing the work precisely where surface cues fail\.

#### Implication\.

Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)’s headline gain is attributable to the LLM backend, not the gap\-to\-budget mapping or the probe\-loop machinery\. Distilling a faster intent classifier from the LLM backend is a practical optimisation; the rule\-based shortcut is not\.

## Appendix A19Cross\-Domain Sanity Checks

The main\-body results live in the social\-simulation regime where the Environment Agent pipeline was designed\. To probe how the architecture transfers, we ran four additional benchmarks outside that regime; we report them here as sanity checks rather than as headline contributions\.All four runs are single\-seed\(vs\. 3 seeds for the main implicit\-intent and factual\-grounding experiments\) and per\-run sample sizes vary: InteractiveBench Puzzle2020episodes, Trust66games/condition, Math2020questions, MemoryArenan=1n\{=\}1paper across55subtasks\. Trust \(n=6n\{=\}6\) and MemoryArena \(n=1n\{=\}1\) are underpowered for statistical inference and are reported as observational anchors only; we do not run paired tests on them\.

![Refer to caption](https://arxiv.org/html/2606.05557v1/figures/cross-domain.png)Figure 12:Cross\-domain results \(all single\-seed; sample sizes per panel vary, see text\)\. Bar height is the raw metric in each domain’s native units \(accuracy, payoff/round, partial\-success score\)\.AURAfull beats Vanilla on Puzzle \(\+5\+5pp absolute,n=20n\{=\}20episodes\), ties on Math \(n=20n\{=\}20, judge\-saturated\), and underperforms on Trust \(n=6n\{=\}6games/cond\. — underpowered, descriptive only\), MemoryArena \(n=1n\{=\}1paper×\\times5 subtasks — anecdotal\), and GAIA \(n=139n\{=\}139, see Appendix[A22](https://arxiv.org/html/2606.05557#A22)\)\. Together with the scope condition in Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1), these data points sketch the empirical envelope of the probing mechanism: it transfers when residual environmental uncertainty after passive perception is non\-trivial*and*tool returns are structurally extractable\.#### InteractiveBench — Puzzle \(text adventure\)\.

20 multi\-turn puzzle episodes per condition, single seed, gpt\-4o\-mini\.AURA\(Full\) reaches0\.700\.70accuracy, Vanilla LLM0\.650\.65,AURA\(No\-Probe\)0\.450\.45\. The architecture\-without\-probing under\-performs Vanilla here because the No\-Probe pipeline injects an empty environment context that consumes turns; restoring probing recovers and slightly exceeds Vanilla \(\+5\+5pp absolute,\+7\.7%\+7\.7\\%relative\)\. This is the cleanest cross\-domain signal we observe\.

#### InteractiveBench — Trust \(iterated prisoner’s dilemma\)\.

6 games per condition against three opponent strategies \(TFT, GrimTrigger, Random\)\. Aggregate payoff per round: Vanilla1\.311\.31,AURANo\-Probe1\.641\.64,AURAFull1\.151\.15\. The aggregate is dominated by opponent\-mix imbalance: against deterministic strategies all conditions reach the cooperative equilibrium of2\.002\.00payoff/round; against the noisy random opponent, the sample sizes differ \(nrounds=17,16,28n\_\{\\text\{rounds\}\}=17,16,28\) and theAURAFull mean is0\.960\.96\. We interpret this as the negative counterpart to Puzzle: when the “environment” is an adversarial opponent whose private state cannot be probed, additional context gathering does not help and may slow convergence\. We do not claim a positive result on Trust\.

#### InteractiveBench — Math \(LaTeX\-conditional QA\)\.

20 questions per condition, single seed; all three conditions reach≈1\.9%\\approx 1\.9\\%\. The LLM judge is conservative onLaTeX\-formatted answers and returns repeated parse errors \(e\.g\., “Invalid \\escape”\) that we manually verified mark correct agent responses as incorrect\. We report the result as null and attribute it to judge brittleness, not to a meaningful gap between conditions; a per\-domain custom judge is future work\.

#### MemoryArena \(research\-paper Q&A, formal reasoning\) —anecdotaln=1n\{=\}1\.

We ran a single\-paper smoke test \(1 paper×\\times5 subtasks, single seed\), which is too small to support a quantitative claim and we do not run statistical tests on it\. We report the observation for completeness:AURA\(memory only\) andAURA\(no\_probe\) reach5/55/5partial\-success \(P​S=1\.0PS\{=\}1\.0\);AURA\(full\) reaches2/52/5\(P​S=0\.4PS\{=\}0\.4\)\. The full\-pipeline under\-performs because probe results pollute the long mathematical context with environment\-style snippets—the same “hallucinated context” failure mode observed atB≥3B\\geq 3\(probes add noise faster than they reduce errors\), here amplified by the long\-context formal\-reasoning regime\. Sample is too small to draw a quantitative claim, but the direction is consistent with the scope condition: the probing mechanism degrades when the task is symbolic and the tool returns are off\-distribution\.

#### Cross\-domain takeaway\.

The architecture transfers positively to a multi\-turn puzzle domain, fails to transfer to adversarial games where opponent state cannot be probed, is masked by judge brittleness onLaTeXmath, and under\-performs on long\-context formal reasoning where probe outputs pollute the context\. Together with the GAIA negative\-transfer result in Appendix[A22](https://arxiv.org/html/2606.05557#A22), these data points describe the empirical envelope of the probing mechanism beyond the primary social\-simulation setting\.

## Appendix A20FANToM External ToM Benchmark \(Null Result\)

To probe whether AURA’s intent\-directed mechanism transfers to a standard, externally\-validated theory\-of\-mind benchmark, we ran a 400\-question stratified split of FANToM\(Kimet al\.,[2023](https://arxiv.org/html/2606.05557#bib.bib56)\):8080questions per type across the five FANToM question types —*beliefQA*\(multi\-choice belief\),*answ\_bin*,*answ\_list*\(answerability binary / list\),*info\_bin*,*info\_list*\(info\-accessibility binary / list\)\. The8080*beliefQA*questions are sub\-stratified20\+20\+20\+2020\{\+\}20\{\+\}20\{\+\}20across first\-/second\-order×\\timesaccessible/inaccessible\. Backbone isgpt\-4o\-mini, seed=4242\(the OpenAIseedkwarg is accepted in this run, so backbone draws are deterministic when supported\)\. Source:run\_fantom\_full\.py; data:evaluation/results/fantom\_full\_seed42\.json\.

Table 25:FANToM accuracy by question type \(n=80n\{=\}80each,400400total\)\. Three conditions on the same backbone:*Literal*= single\-call, no tools;*NoIntent*= ReAct\-style tool use up to budget 3;*Intent*= full AURA pipeline with IntentInferrer\. Last column: per\-question paired contrasts\.#### Statistical tests \(per\-question paired,n=400n\{=\}400\)\.

Intent vs\. Literal:Δ​acc=\+0\.003\\Delta\\text\{acc\}\{=\}\{\+\}0\.003, pairedttp=0\.92p\{=\}0\.92; McNemarp=1\.0p\{=\}1\.0\. Intent vs\. NoIntent:Δ​acc=\+0\.025\\Delta\\text\{acc\}\{=\}\{\+\}0\.025, pairedttp=0\.26p\{=\}0\.26; McNemarp=0\.31p\{=\}0\.31\. Both contrasts null\.

#### Interpretation\.

FANToM’s narrative ToM questions ship the full conversation in\-context, so a competent backbone can answer literally without retrieval; the IntentFrame’s gap calibration finds little to direct, and the additional probe budget adds latency \(6\.56\.5s vs\.0\.90\.9s for Literal\) without lifting accuracy\. FANToM is therefore a clean negative case for AURA’s mechanism — the residual uncertaintyH​\(ℰ∣ℬ\)H\(\\mathcal\{E\}\\mid\\mathcal\{B\}\)after passive context assembly is already low, so probing has nothing to reduce\. This bounds the probing claim’s scope: AURA helps when private state lies behind a tool\-mediated information frontier \(multi\-agent simulation, scene\-grounded queries with hidden private state\), and does not help on narrative ToM transcripts already in\-context\.

#### Backbone\-capability corroboration\.

The Intent condition’s IntentInferrer fell back to the deterministic heuristic0/4000/400times ongpt\-4o\-mini\(vs\.23/2523/25ongemini\-2\.5\-flashin Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\)\. The contrast confirms that the cross\-backbone Gemini regression in Table[15](https://arxiv.org/html/2606.05557#A13.T15)is a backbone\-capability failure \(Gemini’s structured\-JSON adherence\) rather than a problem with the IntentInferrer’s prompt or the AURA pipeline\.

## Appendix A21LoCoMo Long\-Term Conversational Memory \(Partial Transfer\)

To probe whether the AURA pipeline transfers to the long\-term\-conversational\-memory recall regime, we ran a 200\-question stratified split of LoCoMo\(Maharanaet al\.,[2024](https://arxiv.org/html/2606.05557#bib.bib57)\)\(10 conversations×\\times19–32 sessions×\\times∼20\{\\sim\}20turns/session,2,2062\{,\}206QA total across 5 categories\)\. Backbonegpt\-4o\-mini, seed=4242, 8 parallel workers, OpenAIseedkwarg accepted\. Source:run\_locomo\_smoke\.py\(re\-runnable atLOCOMO\_N=200\); data:evaluation/results/locomo\_smoke\.json\. Adapter atlocomo\_eval\.pymaps a conversation’s session list to AURA’s\(scene, memories, query, available\_tools\)quadruple, with three simulated probe tools \(get\_session\(n\),search\_by\_speaker,list\_sessions\_on\_date\)\. Scoring follows the LoCoMo paper exactly: token\-F1 with stem normalisation, multi\-answer F1 for cat\-1, semicolon\-alternative for cat\-3\.

Table 26:LoCoMo 200\-question result \(1 seed, gpt\-4o\-mini\)\.*Literal*= single LLM call with session catalog only;*NoIntent*= ReAct\-style tool use;*Intent*= full AURA pipeline with IntentInferrer\. F1 reported because token\-EM is near\-zero across all conditions \(LoCoMo gold answers are dense phrases that natural prose rarely matches set\-equal — same as the upstream paper\)\.*Fallback*= LLMIntentInferrer parse failures\.#### Statistical tests \(per\-question paired,n=200n\{=\}200\)\.

Intent vs\. Literal:Δ​F1=\+0\.151\\Delta\\text\{F1\}\{=\}\{\+\}0\.151, pairedt=8\.11t\{=\}8\.11,𝐩<𝟏𝟎−𝟏𝟓\\mathbf\{p<10^\{\-15\}\}— highly significant\. Intent vs\. NoIntent:Δ​F1=\+0\.020\\Delta\\text\{F1\}\{=\}\{\+\}0\.020, pairedt=1\.09t\{=\}1\.09,p=0\.28p\{=\}0\.28—not significant\.

#### Interpretation\.

The architecture\-and\-tools layer \(NoIntent: probe\-tool harness with no intent reframing\) carries most of the gain from0\.0420\.042to0\.1730\.173F1; AURA’s IntentInferrer adds only an additional\+0\.020\+0\.020F1, which is not statistically significant on this sample\. Read together with the FANToM null \(Appendix[A20](https://arxiv.org/html/2606.05557#A20)\), LoCoMo gives the same picture: the architectural pipeline transfers to long\-term\-recall and to narrative ToM, but the intent\-direction stage’s marginal contribution*above*a tool\-using baseline is regime\-specific\. It is significant on AURATown’s hand\-designed implicit\-intent set \(Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\) where private\-state needs are deliberately hidden behind the surface form, and small\-or\-null on benchmarks where the underlying QA does not require this kind of lexical\-vs\-implicit gap inference\.

The0/2000/200IntentInferrerfallback rate \(alongside0/4000/400on FANToM\) is the cleanest external corroboration that the originalgemini\-2\.5\-flashcross\-backbone regression \(Section[5\.2](https://arxiv.org/html/2606.05557#S5.SS2)\) is a backbone format\-compliance failure on Gemini’s part, not a problem with AURA’s prompt or schema\.

## Appendix A22GAIA Cross\-Domain Run \(Negative Transfer\)

We ran the GAIA Level\-1/2 question set \(139 questions per condition, seed 42, gpt\-4o\-mini\) under two conditions:*probe*\(theAURAExplore stage with the OpenAI Responses APIweb\_searchtool\) and*direct*\(single\-pass LLM, no tools\)\.

Table 27:GAIA Level\-1/2 \(seed 42,n=139n\{=\}139per condition\)\.#### Reading\.

On Level\-1 \(single\-step factual lookup\) probing*degrades*accuracy by−1\.9\-1\.9pp absolute; on Level\-2 \(multi\-step reasoning\) the two conditions tie\. Total cost: probing spends22×22\\timesmore wall time and3\.53\.5tool calls per question for no aggregate benefit\.

#### Why the transfer fails\.

GAIA’s environment is a search engine wrapped in an LLM\-simulated browser\. Each “probe” is itself an LLM call summarising a web page, not a structured\-state read against a ground\-truth simulator as inAURATown\. Two consequences: \(i\) probe outputs inherit the backbone’s failure modes \(hallucinated facts compound across the probe loop\), and \(ii\) the residualH​\(ℰ∣ℬ\)H\(\\mathcal\{E\}\\mid\\mathcal\{B\}\)that probing was designed to reduce on social\-availability queries is not the bottleneck on web\-grounded factoid questions, where the backbone’s parametric knowledge dominates\. This is the boundary case of the scope condition stated in Section[5\.1](https://arxiv.org/html/2606.05557#S5.SS1): probing helps when \(a\) residual uncertainty after passive perception is non\-trivial*and*\(b\) tool returns are structurally extractable\. GAIA fails \(b\)\.

#### What we report\.

We report GAIA as evidence about a regime in which the bounded\-probing mechanism does not transfer\. Coupling this with the InteractiveBench Trust negative and the MemoryArena formal\-reasoning degradation, the empirical envelope ofAURA’s probing contribution is structured social environments with extractable, non\-hallucinated state probes\.

## Appendix A23Memory and Enrichment Protocol Details

#### Weighted retrieval\.

For each memorymm, queryqq, current timett:

score​\(m,q,t\)\\displaystyle\\text\{score\}\(m,q,t\)=wr⋅exp\(−λ\(t−m\.t\)\)\\displaystyle=w\_\{r\}\\cdot\\exp\(\-\\lambda\(t\-m\.t\)\)\+wp⋅m\.importance10\\displaystyle\\quad\+w\_\{p\}\\cdot\\tfrac\{m\.\\text\{importance\}\}\{10\}\+wv⋅sim\(q,m\.c\)\\displaystyle\\quad\+w\_\{v\}\\cdot\\text\{sim\}\(q,m\.c\)\(1\)Default weights\(wr,wp,wv\)=\(0\.3,0\.4,0\.3\)\(w\_\{r\},w\_\{p\},w\_\{v\}\)=\(0\.3,0\.4,0\.3\), decayλ=0\.01\\lambda=0\.01\.sim​\(⋅\)\\text\{sim\}\(\\cdot\)is a keyword\-based relevance function in the current prototype\.

#### Memory types\.

FollowingTulving and others \([1972](https://arxiv.org/html/2606.05557#bib.bib24)\):*Observation*\(direct percepts\),*Conversation*\(dialogues\),*Reflection*\(synthesised insights everyθ=10\\theta=10observations\),*Plan*\(daily schedules and goals\)\.

#### Importance scoring\.

When an LLM endpoint is available, importance is scored via a dedicated prompt \(“Rate the importance of this event 1–10”\); otherwise a keyword\-based heuristic assigns scores based on emotional salience\.

#### Three\-stage enrichment protocol\.

User queryqqto agentaia\_\{i\}triggers: \(1\) context gathering \(perception \+ recent memories\), \(2\) proactive probing \(1\-step for latency\), \(3\) enriched generation with the structured context prepended\. Enrichment metadata \(tools called, context gathered\) is exposed to the user for interpretability\.

## Appendix A24Reproducibility Checklist

#### Code and data\.

All experiments are driven by the runners inscripts/andrun\_experiments\.py, with results saved toevaluation/results/\. Each table or figure traces to a specific result JSON viaevaluation/results/MANIFEST\.md; the manifest covers factual\-grounding multi\-seed, factual\-grounding Fixed\-Probe and GapRouted Pareto controls, the privacy\-sensitive distractor slice, implicit\-intent clean\-prompt multi\-seed, fixed\-private/oracle\-intent controls, no\-few\-shot prompt ablation, Plan\-and\-Solve and Reflexion adapter outputs, FANToM 400q, LoCoMo 200q, the four cross\-backbone runs, implicit\-intent IAA returns, and the strict\-precision rescore\. The code release will include theAURAagent library, theAURATownsimulation, all 50 environment\-grounded queries with their templates, all 100 implicit\-intent queries \(4 scenes×\\times25\) with subcategory and target labels, and all per\-condition / per\-seed details that back the numbers reported here\.

#### Models and APIs\.

Agent backbone and judge are bothgpt\-4o\-minivia the OpenAI Chat Completions API\. Cross\-backbone tests additionally useclaude\-haiku\-4\-5\(Anthropic\),qwen\-plus\(Alibaba\), andgemini\-2\.5\-flash\(Google\), each via its vendor’s official API\. Backbone temperature is 0\.7 for action decisions, 0\.1 for the judge and the IntentInferrer\. We pass theseedparameter to the OpenAI Chat Completions API where the SDK accepts it; on rejection, the LLM engine falls back to a no\-seed call and flips an internal flag for the rest of the run\. Cross\-backbone runs do not all support theseedkwarg, so the multi\-seed claim is bounded as: “Python\-level random sampling and OpenAIseedwhen supported\.”

#### Determinism caveats\.

Multi\-seed paired tests for factual grounding / component ablation / implicit\-intent vary the seed both at Python’srandommodule \(controlling query ordering, agent shuffles, and template instantiation\) and at the OpenAI APIseedkwarg\. Stochastic backbone decoding above the seed kwarg may still introduce per\-seed variance; this is acknowledged in the Limitations and is consistent with the OpenAI API documentation thatseedprovides a best\-effort, not strict, determinism guarantee\. The simulation server itself derives its world generation fromTownConfig\.world\_seedand the agent backbone derives its outputs fromTownConfig\.llm\_seed; both are propagated through/api/reset\.

#### Strict\-paired factual grounding\.

The factual\-grounding runner \(run\_experiments\.py:run\_rq2\) operates in two phases\. Phase A: reset with the experiment seed, warm 10 steps, then advance one tick per query position to capture 50 ground\-truth snapshots\. Phase B: each AURA condition resets with the same seed, warms 10, and replays the snapshots; chats run withread\_only=Trueso they do not write event log or memory, leaving the per\-tick trajectory deterministic across conditions\. External baselines \(Vanilla, Static, ReAct, Reflexion, Plan\-and\-Solve\) receive the same Phase\-A snapshots\. Output JSON stamps a\_paired\_snapshot\_metablock recordingseed,n\_snapshots, andshared\_snapshots\_across\_conditions=true\.

#### Aggregate scripts\.

rescore\_rq2\_strict\.pyrecomputes the strict precision rescore \(Appendix[A9](https://arxiv.org/html/2606.05557#A9)\) from the multi\-seed factual\-grounding details\.aggregate\_rq2\_multiseed\.pycollapses per\-seed details into the canonical multi\-seed summary\.run\_rq2\_fixed\_probe\.pyandrun\_rq2\_aura\_gap\_routed\.pyregenerate the saturated and gap\-routed factual\-grounding controls;rq2\_pareto\_analysis\.pycomputes FA/probe/disclosure contrasts andplot\_rq2\_pareto\.pyrenders Figure[2](https://arxiv.org/html/2606.05557#S5.F2)\.run\_privacy\_distractor\.pyregenerates the 30\-query forbidden\-tool slice\.run\_implicit\_intent\_full\.pyandrun\_implicit\_intent\_v2\.pyregenerate the primary 25\-query and expanded 100\-query \(4\-scene\) implicit\-intent runs respectively; the v2 runner supports\-\-resumefor interrupted multi\-seed jobs and a\-\-validate\-onlypre\-flight pass on the benchmark JSON\.compute\_irr\.pyproduces the IAA Cohen’sκ\\kappaon the implicit\-intent subcategory labels\.rq5\_rater\_aggregated\.pyproduces the per\-rater aggregated Wilcoxon used in Table[21](https://arxiv.org/html/2606.05557#A17.T21)\.run\_fantom\_full\.pyregenerates the FANToM 400\-question external bench\.run\_locomo\_smoke\.pyregenerates the LoCoMo run\.audit\_paper\_numbers\.pyre\-derives every cited headline from the JSON files and is the recommended pre\-submission check\.

#### Hardware\.

All experiments run on a single CPU machine; the LLM is invoked via remote API\. No GPU is required\. Wall\-clock budgets per run: factual\-grounding multi\-seed∼30\{\\sim\}30min/seed; implicit\-intent∼10\{\\sim\}10min/seed; FANToM 400 questions∼9\{\\sim\}9min; LoCoMo 200 questions∼6\{\\sim\}6min; PnS\-on\-implicit\-intent multi\-seed∼16\{\\sim\}16min total; Reflexion \+ PnS on factual grounding multi\-seed combined∼2\.5\{\\sim\}2\.5h\. Total API spend for all reported experiments under $15 USD\.

## Appendix A25Ethics Statement

#### Research artefacts\.

This paper studies an LLM agent architecture and accompanying social simulation\. Human involvement consists of the 8 voluntary annotators of the helpfulness study and 2 additional annotators for the implicit\-intent label audit\. Annotators were friends and colleagues of the authors who consented to evaluate anonymous response pairs or query labels; no demographic data was collected; no personally identifying information appears in saved annotations\. Annotators received no compensation; each task was a one\-time, optional 15–30 minute evaluation\. We retain a dissenting rater in the primary helpfulness analysis to avoid post\-hoc exclusion bias and we report rater\-level transparency \(per\-rater preference rates, Krippendorff’sα\\alpha\) so that readers can inspect inter\-annotator dynamics\.

#### Simulation content\.

AURATown’s 5 agents have hand\-authored profiles \(occupation, personality, daily routine\)\. The names are fictional\. Agent profiles deliberately span ages 20–68 and include contrasting personality types but are not intended to represent any real demographic distribution\. The simulation does not emulate real people, real businesses, or real locations\.

#### Misuse considerations\.

The IntentFrame mechanism produces a structured estimate of a user’s implicit information need from a surface query\. In a benign deployment this is used to surface relevant context the user did not lexically request; in an adversarial deployment, the same machinery could be used to surface information the user has not consented to share\. The Explore stage’s tool whitelist is the principal mitigation: tools are scoped to the simulation’s structured environment state, not to external data sources\. Practitioners deploying a similar IntentInferrer over real personal\-data tools should constrain the whitelist accordingly and surface the inferred implicit need to the user \(the heads\-up prefix in ourInteractstage is a minimal version of this\)\.

#### Energy and compute\.

All experiments run on remote API endpoints; the only local compute is a single CPU process running the AURATown simulation server\. We did not benchmark API energy use; total token consumption across all reported runs is approximately 8 million input tokens and 200 thousand output tokens againstgpt\-4o\-mini\.

#### Limitations as ethics\.

The paper’s empirical claims are scoped to the regime where structured environment access matters \(single\-user situated queries with hidden private state\)\. Where the mechanism does not transfer \( narrative ToM in FANToM, web\-grounded factoid in GAIA, formal reasoning in MemoryArena \) we report negative results in Appendix[A20](https://arxiv.org/html/2606.05557#A20),[A22](https://arxiv.org/html/2606.05557#A22), and[A19](https://arxiv.org/html/2606.05557#A19)\. We avoid abstracting from these specific findings to general claims about LLM theory\-of\-mind or proactive assistance\.

## Appendix A26Collective Behaviour and Supplementary Discussion

### A26\.1Collective Behaviour in Multi\-AgentAURATown

In the 200\-step SOTOPIA run, 5 agents with independentAURApipelines \(IntentInferrer disabled\) produce 44 emergent behaviours: collaboration \(32, 73%\), routine adaptation \(7, 16%\), conflict resolution \(4, 9%\), and group formation \(1\)\. SOTOPIA scores \(Table[20](https://arxiv.org/html/2606.05557#A17.T20)\) show strength on*goal*\(9\.5\) and*believability*\(9\.0\) but weakness on*social\_rules*\(−2\.13\-2\.13\) and*secret*\(−0\.67\-0\.67\), indicating that structured environment access supports local coordination but not higher\-order social constraint handling\. These data supplement the SOTOPIA evaluation; they do not constitute a separate theory\-of\-mind claim\.