File size: 5,986 Bytes
d861bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cbdf2bb
 
d861bc9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cbdf2bb
 
 
 
 
 
 
d861bc9
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
---
license: cc-by-4.0
language:
- en
pretty_name: AURA Implicit-Intent Benchmark (AURATown)
annotations_creators:
- expert-generated
source_datasets:
- original
task_categories:
- question-answering
- text-classification
tags:
- theory-of-mind
- implicit-intent
- social-intelligence
- situated-agents
- llm-agents
- tool-use
- proactive-probing
size_categories:
- n<1K
configs:
- config_name: implicit_intent
  default: true
  data_files:
  - split: test
    path: implicit_intent.jsonl
- config_name: implicit_intent_v1
  data_files:
  - split: test
    path: implicit_intent_v1.jsonl
- config_name: scenes
  data_files:
  - split: test
    path: scenes.jsonl
- config_name: privacy_distractor
  data_files:
  - split: test
    path: privacy_distractor.jsonl
- config_name: factual_grounding
  data_files:
  - split: test
    path: factual_grounding.jsonl
- config_name: grounding_templates
  data_files:
  - split: test
    path: grounding_templates.jsonl
---

# AURA Implicit-Intent Benchmark (AURATown)

A small, **author-authored** evaluation suite for studying **implicit-need
surfacing** by situated LLM agents. 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 *available*, *in a good mood*, or *worth interrupting now*.
This benchmark separates the **literal** answer (readable from public scene
state) from the **implicit** need (which requires private/hidden state),
and labels which tools are *required* vs *forbidden* to answer it.

All queries are grounded in **AURATown**, a small grid-based social simulation
with 5 named agents (Lin Wei, Zhang Hao, Chen Mei, Liu Yang, Wang Jun) and a set
of named locations. Each scene fixes a time of day, locations, and per-agent
public/private state so the benchmark tests cross-scene robustness rather than
memorisation of one configuration.

> **Paper:** *AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated
> LLM Agents* — Li, Liu, Cai, Xu (2026). arXiv:[2606.05557](https://arxiv.org/abs/2606.05557).

## Configs

| Config | Rows | What it is |
|---|---:|---|
| `implicit_intent` *(default)* | 100 | Primary benchmark: 4 scenes × 25 queries, 5 subcategories. |
| `implicit_intent_v1` | 25 | Pilot subset (= Scene A of v2, single scene). Used for the IAA study. |
| `scenes` | 4 | Scene metadata: public/private state + beliefs (nested state JSON-encoded). |
| `privacy_distractor` | 30 | Factual questions that must **not** touch private/historical state; each carries `allowed_tools` / `forbidden_tools`. |
| `factual_grounding` | 50 | Plain environment-grounded factual queries (spatial/social/temporal/memory/planning). |
| `grounding_templates` | 22 | Parameterised templates that generate verifiable questions from simulation state. |

The original, un-flattened JSON files are also shipped verbatim under `raw/`.

## Fields (`implicit_intent` config)

| Field | Type | Description |
|---|---|---|
| `id` | int | Stable query id. |
| `scene` | str | Scene key, e.g. `A_cafe_morning`. |
| `scene_summary` | str | One-line natural-language scene context. |
| `subcategory` | str | One of `availability`, `mood`, `appropriateness`, `latent_goal`, `second_order`. |
| `agent_subject` | str | Agent the query is about. |
| `target` | str / null | For `second_order` queries, the third party being reasoned about. |
| `query` | str | The surface user query. |
| `literal_requires` | list[str] | Public state needed for the literal answer. |
| `implicit_requires` | list[str] | Private state needed for the implicit answer. |
| `implicit_need` | str | One-line statement of what the user is *really* asking. |
| `gold_required_tools` | list[str] | Tools an oracle must call to surface the implicit need. |
| `forbidden_tools` | list[str] | Tools that would over-reach / leak private state. |

## Construction & annotation

- Queries were **hand-authored** by the paper authors. Scene A reuses the 25
  pilot queries verbatim; scenes B/C/D are new and authored to keep surface
  forms disjoint from the pilot. Stale-belief templates are adapted from
  Ullman (2023).
- **Inter-annotator agreement**: two independent annotators relabelled the
  pilot 25 queries' subcategory under the same 5-class scheme; Cohen's
  κ = 0.61 (substantial, Landis–Koch), with disagreements concentrated on the
  mood/appropriateness/availability boundary. The author labels are retained as
  gold. (IAA raw response files are *not* included in this release.)

## Intended use & scope

Designed to evaluate whether an agent **surfaces the implicit need** behind a
situated query while respecting tool/privacy boundaries — **not** general QA.
On purely factual lookup the implicit-intent machinery is *not* expected to help
(see `factual_grounding` / `privacy_distractor`, which are control slices).

## Limitations

- **Small** (100 primary queries) and **single-simulator** (AURATown only); the
  agent roster and scene set are fixed.
- Author-written gold labels; the IAA (κ=0.61) is supportive but the sample is
  small. Treat subcategory labels as a useful partition, not ground truth.
- English only; agent names are romanised Mandarin but all query text is English.

## Related external benchmarks

This suite is *original*; cross-domain checks in the paper use the external
benchmarks **FANToM** (Kim et al., 2023), **LoCoMo** (Maharana et al., 2024),
**GAIA**, and **SOTOPIA** (Zhou et al., 2024). Those datasets are **not**
redistributed here — obtain them from their original sources.

## Citation

```bibtex
@misc{li2026aura,
  title         = {AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents},
  author        = {Li, Yang and Liu, Jiaxiang and Cai, Jiang and Xu, Mingkun},
  year          = {2026},
  eprint        = {2606.05557},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2606.05557}
}
```

## License

CC-BY-4.0. You may share and adapt with attribution.