emodiary commited on
Commit
1276054
·
verified ·
1 Parent(s): df8c310

Enable Dataset Viewer for EgoBench tasks

Browse files
Files changed (2) hide show
  1. README.md +420 -410
  2. data/test.parquet +3 -0
README.md CHANGED
@@ -1,410 +1,420 @@
1
- ---
2
- pretty_name: "EgoBench"
3
- language:
4
- - en
5
- task_categories:
6
- - visual-question-answering
7
- tags:
8
- - video
9
- - multimodal
10
- - benchmark
11
- - tool-use
12
- - agent
13
- - egocentric-video
14
- size_categories:
15
- - 1K<n<10K
16
- license: other
17
- license_name: egobench-mixed-license
18
- license_link: "https://huggingface.co/datasets/emodiary/EgoBench/blob/main/LICENSE"
19
- ---
20
-
21
- # EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents
22
-
23
- Our paper, **“EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents,”** has been accepted to **NeurIPS 2026**.
24
-
25
- <div align="center">
26
- <a href="figure/fig1.pdf">
27
- <img src="figure/fig1.png" alt="Overview of EgoBench tasks and evaluation workflow" width="100%">
28
- </a>
29
- <p><em>EgoBench task overview.</em></p>
30
- </div>
31
-
32
- Welcome to **EgoBench**, the first interactive multimodal agent benchmark grounded in **egocentric (first-person) videos**. EgoBench is designed to bridge the evaluation gap for AI agents operating in open, real-world environments. It jointly assesses three critical capabilities: **multimodal perception**, **tool-augmented multi-hop reasoning**, and **dynamic user interaction**.
33
-
34
- EgoBench constructs a dynamic environment containing 1,590 tasks across five scenario categories: Retail, Kitchen, Restaurant, Warehouse, and Household. Restaurant 6 is the multi-restaurant selection variant within the Restaurant category.
35
-
36
- ---
37
-
38
- <div align="center">
39
- <h3>📌 Key Features</h3>
40
- </div>
41
-
42
- **1. Egocentric Visual Collection**
43
- - Built upon the **Ego4D** dataset and self-collected first-person videos.
44
- - Tasks contain rich **spatiotemporal cues** (e.g., "the bottle on the left," "the one I just picked up") to test the agent's ability to resolve references in dynamic scenes.
45
-
46
- **2. Strict Capability Coupling**
47
- - Tasks are designed to enforce the **joint application** of visual perception and tool invocation based multi-hop reasoning.
48
- - We implement a three-stage synergistic pipeline to generate such tasks: Specifically, we first collect ego-centric videos with explicit spatiotemporal cues as visual anchors, then build a comprehensive tool library and database with a visual–information gap, and finally design tasks that require agents to perform multimodal perception, retrieve hidden contextual information via tools, reason logically, and modify database states through tool use.
49
-
50
- **3. Multi-Agent User Simulation**
51
- - Features an **Actor-Evaluator-Summarizer** architecture to generate high-fidelity, goal-aligned user responses.
52
- - Supports three interaction modes: `Dynamic Easy Mode`, `Dynamic Hard Mode` (with distractions), and `Static Mode`, to comprehensively test agents' interactive ability.
53
-
54
- **4. Deterministic Evaluation Framework**
55
- - Ensures objectivity through **process-based** (tool-call coverage) and **result-based** (database state equivalence) validation.
56
- - Eliminates reliance on subjective LLM judges.
57
-
58
-
59
- ## Architecture
60
-
61
- The system operates as a **two-agent dialogue loop**:
62
-
63
- 1. **Simulated User** — Simulates a customer with a specific task. In *easy* mode, the customer is cooperative and provides information step-by-step. In *hard* mode, the customer has low patience, gives vague or fragmented information, and may provide negative feedback. In *static* mode, the customer sends a single comprehensive request with no follow-up.
64
-
65
- 2. **Service Agent** — The LLM being evaluated. It receives tool definitions for the scenario, processes the user's request, invokes tools via structured JSON output, and responds in natural language. Tool calls are intercepted, executed against an scenario database, and the results are fed back into the conversation.
66
-
67
- 3. **Evaluation Pipeline** — After simulation, tool calls and database state are compared against scenario ground truth using fuzzy matching to compute success metrics. Results feed into visualization and report generation scripts.
68
-
69
- ```
70
- ┌─────────────┐ message ┌──────────────┐
71
- │ Simulated │ ───────────────▶ │ Service Agent │
72
- │ User │ ◀─────────────── │ (LLM under │
73
- │ │ response │ test) │
74
- └─────────────┘ └──────┬───────┘
75
- │ tool call (JSON)
76
- ▼
77
- ┌──────────────┐
78
- │ Scenario DB │
79
- │ │
80
- └──────────────┘
81
- ```
82
-
83
- ## Supported Scenarios
84
-
85
- | Scenario | #Variants | #Tasks |
86
- |------------|-----------|--------|
87
- | Retail | 10 | 454 |
88
- | Kitchen | 4 | 175 |
89
- | Restaurant | 6 | 316 |
90
- | Warehouse | 25 | 375 |
91
- | Household | 18 | 270 |
92
- | **Total** | **63** | **1,590** |
93
-
94
- Each scenario variant provides a different database configuration (product catalogs, menus, etc.) loaded from `tools/{scenario}/{scenario}_init.py`.
95
-
96
- ## Supported Models
97
-
98
- | Model | Provider | API |
99
- |-------|----------|-----|
100
- | `glm-5v-turbo` | Zhipu AI | Zhipu AI SDK |
101
- | `qwen3-vl-235b` | Alibaba Qwen | Custom endpoint |
102
- | `Qwen3.5-397B-A17B` | Alibaba Qwen | OpenAI-compatible |
103
- | `qwen3.6-plus` | Alibaba Qwen | OpenAI-compatible |
104
- | `gemini-3.1-pro-preview` | Google | OpenAI-compatible proxy |
105
- | `kimi-k2.6` | Moonshot | OpenAI-compatible |
106
- | `mimo-v2.5-omni` | Xiaomi MiMo | OpenAI-compatible |
107
- | `doubao-seed-2-0-pro-260215` | ByteDance Doubao | OpenAI-compatible |
108
- | `manual` | — | Terminal input for manual testing |
109
-
110
- ## Project Structure
111
-
112
- ```
113
- ├── run/ # Core simulation framework
114
- │ ├── multi_agent.py # Main entry point
115
- │ ├── utils.py # LLM API calls, tool execution, message building
116
- │ ├── prompts.py # Prompt templates for all agents
117
- │ └── apis/ # LLM API wrappers
118
- │ ├── unified.py # Unified call_llm dispatcher
119
- │ ├── zhipu.py # Zhipu AI
120
- │ ├── qwen.py # Qwen
121
- │ ├── qwen3_5.py # Qwen3.5
122
- │ ├── kimi.py # Kimi
123
- │ ├── mimo.py # MiMo
124
- │ └── doubao.py # Doubao
125
- ├── tools/ # Scenario databases & tool definitions
126
- │ ├── retail/ # RetailDB, retail_init.py, retail_tools.json
127
- │ ├── kitchen/ # KitchenDB, kitchen_init.py, kitchen_tools.json
128
- │ ├── restaurant/ # RestaurantDB plus Restaurant 6 multi-restaurant backend/schema
129
- │ ├── warehouse/ # WarehouseDB and variants 1-25
130
- │ ├── household/ # HouseholdDB and variants 1-18
131
- │ ├── database_init.py # DB initialization utilities
132
- │ └── user_words.py # Extra phrases for hard-mode user agent
133
- ├── scenarios/
134
- │ └── final/ # Task definitions and ground truth (JSON)
135
- │ ├── video/ # Version-controlled benchmark video assets
136
- │ ├── retail1.json .. retail10.json
137
- │ ├── kitchen1.json .. kitchen4.json
138
- │ ├── restaurant1.json .. restaurant6.json
139
- │ ├── warehouse1.json .. warehouse25.json
140
- │ └── household1.json .. household18.json
141
- ├── analysis_scripts/ # Evaluation & visualization
142
- │ ├── evaluate_interaction.py # Core evaluation script
143
- │ ├── print_eval.py # Pretty-print evaluation results
144
- │ ├── analyze_errors.py # Error categorization
145
- │ ├── analyze_tool_correctness.py
146
- │ ├── visualize_results.py # Chart generation
147
- │ ├── plot_results.py # Additional plotting
148
- │ ├── generate_comprehensive_report.py
149
- │ └── run_eval.sh # Run evaluation across all models
150
- ├── run_all.sh # Launch all model runs in parallel
151
- ├── run_scenario.sh # Configurable multi-model runner
152
- ├── run_glm.sh / run_qwen3.sh / … # Model-specific runners
153
- └── requirements.txt
154
- ```
155
-
156
- <div align="center">
157
- <h3>🚀 Quick Start</h3>
158
- </div>
159
-
160
- ## Getting Started
161
-
162
- ### Prerequisites
163
-
164
- - Python 3.10+
165
- - API keys for the models you want to evaluate
166
-
167
- ### API Key Configuration
168
-
169
- All API keys and base URLs are managed through environment variables. A `.env.example` file is provided as a template.
170
-
171
- 1. Copy the example file and fill in your keys:
172
- ```bash
173
- cp .env.example .env
174
- ```
175
-
176
- 2. Edit `.env` and replace the placeholder values with your actual API keys and endpoints:
177
- ```bash
178
- # Main API key for LLM platform models
179
- # Used by: Qwen3.5, Qwen3-VL, Gemini, DeepSeek, GLM-4.5V, Doubao, Qwen3.6, etc.
180
- export API_KEY="your-api-key-here"
181
-
182
- # Base URL for LLM API endpoints (Qwen, Doubao, Gemini, etc.)
183
- # Default: https://api.example.com/v1
184
- export LLM_API_BASE_URL="https://your-llm-api-endpoint.com/v1"
185
-
186
- # Kimi (Moonshot) API key
187
- # Used by: kimi-k2.6
188
- export KIMI_API_KEY="your-kimi-api-key-here"
189
-
190
- # Kimi API base URL (optional, defaults to Moonshot's official endpoint)
191
- export KIMI_API_BASE_URL="https://api.moonshot.cn/v1"
192
-
193
- # Xiaomi MiMo API key
194
- # Used by: mimo-v2.5-omni
195
- export MIMO_API_KEY="your-mimo-api-key-here"
196
-
197
- # MiMo API base URL (optional, defaults to Xiaomi's official endpoint)
198
- export MIMO_API_BASE_URL="https://api.xiaomimimo.com/v1"
199
-
200
- # Zhipu AI API key
201
- # Used by: glm-5v-turbo, glm-4.5v
202
- export ZHIPU_API_KEY="your-zhipu-api-key-here"
203
- ```
204
-
205
- 3. Source the `.env` file before running simulations:
206
- ```bash
207
- source .env
208
- ```
209
-
210
- > **Note:** The shell scripts (`run_glm.sh`, `run_kimi.sh`, etc.) automatically source `.env` when executed. The `.env` file is excluded from version control via `.gitignore`.
211
-
212
- | Environment Variable | Required | Default | Description |
213
- |---------------------|----------|---------|-------------|
214
- | `API_KEY` | Yes* | - | API key for Qwen, Gemini, DeepSeek, GLM, Doubao models |
215
- | `LLM_API_BASE_URL` | Yes* | `https://api.example.com/v1` | Base URL for LLM API endpoints |
216
- | `KIMI_API_KEY` | Yes* | - | Moonshot Kimi API key |
217
- | `KIMI_API_BASE_URL` | Yes* | `https://api.moonshot.cn/v1` | Kimi API base URL |
218
- | `MIMO_API_KEY` | Yes* | - | Xiaomi MiMo API key |
219
- | `MIMO_API_BASE_URL` | Yes* | `https://api.xiaomimimo.com/v1` | MiMo API base URL |
220
- | `ZHIPU_API_KEY` | Yes* | - | Zhipu AI API key |
221
-
222
- *Required only if using the corresponding models.
223
-
224
- ### Video Configuration
225
-
226
- The required video files are version-controlled in this Git repository under
227
- `scenarios/final/video/`. Clone the repository to download them together with
228
- the task definitions and ground truth.
229
-
230
- The framework uses `<repository>/scenarios/final/video` by default. To keep the
231
- videos elsewhere, pass
232
- `--video_dir` or set `EGOBENCH_VIDEO_DIR`:
233
-
234
- ```bash
235
- export EGOBENCH_VIDEO_DIR="/absolute/path/to/video"
236
- ```
237
-
238
- Legacy scenario entries containing a URL are resolved by filename against this
239
- local directory. Extension-only differences such as `.MOV` versus `.mp4` are
240
- also handled automatically.
241
-
242
- ### Installation
243
-
244
- ```bash
245
- pip install -r requirements.txt
246
- ```
247
-
248
- ### Running a Simulation
249
-
250
- ```bash
251
- # Set up API keys (required before first run)
252
- cp .env.example .env # then edit .env with your keys
253
- source .env
254
-
255
- # Run a single simulation
256
- python run/multi_agent.py \
257
- --scenario retail \
258
- --scenario_number 1 \
259
- --user_mode easy \
260
- --service_model_name Qwen3.5-397B-A17B \
261
- --multi_agent_user \
262
- --summary_user
263
-
264
- # Run with limited samples
265
- python run/multi_agent.py \
266
- --scenario kitchen \
267
- --scenario_number 2 \
268
- --user_mode hard \
269
- --service_model_name doubao-seed-2-0-pro-260215 \
270
- --num_samples 10
271
-
272
- # Run a Household scenario
273
- python run/multi_agent.py \
274
- --scenario household \
275
- --scenario_number 1 \
276
- --user_mode easy \
277
- --service_model_name kimi-k2.6 \
278
- --num_samples 1
279
-
280
- # Rerun only previously failed scenarios
281
- python run/multi_agent.py \
282
- --scenario restaurant \
283
- --scenario_number 1 \
284
- --user_mode easy \
285
- --service_model_name qwen3-vl-235b \
286
- --num_samples 0
287
- ```
288
-
289
- ### Key Arguments
290
-
291
- | Argument | Values | Default | Description |
292
- |----------|--------|---------|-------------|
293
- | `--scenario` | `retail`, `kitchen`, `restaurant`, `warehouse`, `household` | `retail` | Domain scenario |
294
- | `--scenario_number` | scenario-specific integer | `1` | Retail 1-10, Kitchen 1-4, Restaurant 1-6, Warehouse 1-25, Household 1-18 |
295
- | `--user_mode` | `easy`, `hard`, `static` | `easy` | User agent difficulty |
296
- | `--service_model_name` | See [Supported Models](#supported-models) | `qwen3-vl-235b` | LLM to evaluate |
297
- | `--multi_agent_user` | flag | off | Enable contradiction checker for user responses |
298
- | `--summary_user` | flag | off | Enable dialogue summarization to manage context length |
299
- | `--num_samples` | int | `0` | Limit scenarios per run (`0` = all) |
300
-
301
- ### Running Evaluation
302
-
303
- ```bash
304
- # Evaluate a specific model
305
- cd analysis_scripts
306
- python evaluate_interaction.py --model_name glm-4.5v --num_samples 10
307
-
308
- # Evaluate all models
309
- bash run_eval.sh
310
- ```
311
-
312
- ### Batch Execution
313
-
314
- ```bash
315
- # Run all models in parallel
316
- bash run_all.sh
317
-
318
- # Run a specific model across all scenarios and modes
319
- bash run_glm.sh
320
- bash run_qwen3.5.sh
321
- ```
322
-
323
- ## User Modes
324
-
325
- ### Easy Mode
326
- The user agent acts as a cooperative customer who provides information step-by-step, responds clearly to questions, and signals completion with "STOP" when all requirements are met.
327
-
328
- ### Hard Mode
329
- The user agent simulates a challenging customer with low patience, provides vague or incomplete information initially, may give negative feedback ("Bad Service Agent"), and passively introduces contradictions. Additional distractor phrases from `tools/user_words.py` are injected to increase difficulty.
330
-
331
- ### Static Mode
332
- A single-turn interaction where the user agent sends one comprehensive request containing all requirements at once. The service agent must process everything without follow-up clarification.
333
-
334
- ## Evaluation
335
-
336
- The evaluation pipeline in `analysis_scripts/evaluate_interaction.py` compares each service agent's tool calls against the scenario ground truth:
337
-
338
- - **Tool-call matching** — Verifies that the correct tools were invoked with the correct parameters
339
- - **Fuzzy matching** — Handles minor variations in parameter values (e.g., product names) using fuzzy string matching
340
- - **Database verification** — Re-executes tool calls against the scenario database to verify result correctness
341
- - **Success metrics** — Computes per-scenario and aggregate success rates across difficulty levels
342
-
343
- ### Visualization & Reports
344
-
345
- ```bash
346
- # Generate charts and reports
347
- cd analysis_scripts
348
- python visualize_results.py
349
- python plot_results.py
350
- python generate_comprehensive_report.py
351
- ```
352
-
353
- Output charts include success rates by difficulty/scenario, error distributions, token usage correlations, and dialogue round analysis. Charts are saved to `eval_result/figures/`.
354
-
355
- ## Output Format
356
-
357
- Simulation results are saved to `results/{model_name}/{scenario}{number}_{mode}.json`:
358
-
359
- ```json
360
- {
361
- "scenario_id": 0,
362
- "mode": "easy",
363
- "instruction": "...",
364
- "image_description": "...",
365
- "dialogue": [{"role": "user", "turn": 1, "content": "..."}, ...],
366
- "tool_calls": [{"turn": 1, "calls": [...], "results": [...]}],
367
- "rounds_count": 5,
368
- "input_tokens": 12345,
369
- "output_tokens": 6789,
370
- "tool_calls_count": 8,
371
- "user_response_time_seconds": 2.3,
372
- "agent_response_time_seconds": 5.1,
373
- "execution_time_seconds": 42.0,
374
- "user_performance": {...}
375
- }
376
- ```
377
-
378
- Evaluation results are saved to `eval_result/{model_name}/{scenario}{number}_{mode}_eval.json`.
379
-
380
- ![Code License](https://img.shields.io/badge/code-MIT-blue.svg)
381
- ![Data License](https://img.shields.io/badge/data-CC%20BY%204.0-lightgrey.svg)
382
-
383
- ## License
384
-
385
- EgoBench uses separate licenses for code and data:
386
-
387
- - Source code, including Python files, shell scripts, and evaluation utilities,
388
- is licensed under the [MIT License](./LICENSE-CODE).
389
- - Benchmark task definitions, annotations, JSON files, scenario databases, and
390
- video files under `scenarios/final/video/` are licensed under the
391
- [Creative Commons Attribution 4.0 International License](./DATA_LICENSE.md).
392
-
393
- The included video assets are distributed with authorization from their
394
- respective rightsholders. Third-party components, if any, remain subject to
395
- their respective licenses. See the repository [licensing notice](./LICENSE)
396
- before using or redistributing individual components.
397
-
398
- ## Citation
399
-
400
- If you find EgoBench useful in your research, please cite our paper:
401
-
402
- ```bibtex
403
- @inproceedings{liu2026egobench,
404
- title = {EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents},
405
- author = {Yunqi Liu and Tong Niu and Zitong Wang and Zhenlong Dai and Yuqi Qing and Weiqiang Wang and Jian Liu},
406
- booktitle = {Advances in Neural Information Processing Systems},
407
- year = {2026},
408
- url = {https://arxiv.org/abs/2605.27820}
409
- }
410
- ```
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: "EgoBench"
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - visual-question-answering
7
+ tags:
8
+ - video
9
+ - multimodal
10
+ - benchmark
11
+ - tool-use
12
+ - agent
13
+ - egocentric-video
14
+ size_categories:
15
+ - 1K<n<10K
16
+ license: other
17
+ license_name: egobench-mixed-license
18
+ license_link: "https://huggingface.co/datasets/emodiary/EgoBench/blob/main/LICENSE"
19
+ configs:
20
+ - config_name: default
21
+ data_files:
22
+ - split: test
23
+ path: data/test.parquet
24
+
25
+ ---
26
+
27
+ # EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents
28
+
29
+ Our paper, **“EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents,”** has been accepted to **NeurIPS 2026**.
30
+
31
+ <div align="center">
32
+ <a href="figure/fig1.pdf">
33
+ <img src="figure/fig1.png" alt="Overview of EgoBench tasks and evaluation workflow" width="100%">
34
+ </a>
35
+ <p><em>EgoBench task overview.</em></p>
36
+ </div>
37
+
38
+ Welcome to **EgoBench**, the first interactive multimodal agent benchmark grounded in **egocentric (first-person) videos**. EgoBench is designed to bridge the evaluation gap for AI agents operating in open, real-world environments. It jointly assesses three critical capabilities: **multimodal perception**, **tool-augmented multi-hop reasoning**, and **dynamic user interaction**.
39
+
40
+ EgoBench constructs a dynamic environment containing 1,590 tasks across five scenario categories: Retail, Kitchen, Restaurant, Warehouse, and Household. Restaurant 6 is the multi-restaurant selection variant within the Restaurant category.
41
+
42
+ ---
43
+
44
+ <div align="center">
45
+ <h3>📌 Key Features</h3>
46
+ </div>
47
+
48
+ **1. Egocentric Visual Collection**
49
+ - Built upon the **Ego4D** dataset and self-collected first-person videos.
50
+ - Tasks contain rich **spatiotemporal cues** (e.g., "the bottle on the left," "the one I just picked up") to test the agent's ability to resolve references in dynamic scenes.
51
+
52
+ **2. Strict Capability Coupling**
53
+ - Tasks are designed to enforce the **joint application** of visual perception and tool invocation based multi-hop reasoning.
54
+ - We implement a three-stage synergistic pipeline to generate such tasks: Specifically, we first collect ego-centric videos with explicit spatiotemporal cues as visual anchors, then build a comprehensive tool library and database with a visual–information gap, and finally design tasks that require agents to perform multimodal perception, retrieve hidden contextual information via tools, reason logically, and modify database states through tool use.
55
+
56
+ **3. Multi-Agent User Simulation**
57
+ - Features an **Actor-Evaluator-Summarizer** architecture to generate high-fidelity, goal-aligned user responses.
58
+ - Supports three interaction modes: `Dynamic Easy Mode`, `Dynamic Hard Mode` (with distractions), and `Static Mode`, to comprehensively test agents' interactive ability.
59
+
60
+ **4. Deterministic Evaluation Framework**
61
+ - Ensures objectivity through **process-based** (tool-call coverage) and **result-based** (database state equivalence) validation.
62
+ - Eliminates reliance on subjective LLM judges.
63
+
64
+
65
+ ## Architecture
66
+
67
+ The system operates as a **two-agent dialogue loop**:
68
+
69
+ 1. **Simulated User** — Simulates a customer with a specific task. In *easy* mode, the customer is cooperative and provides information step-by-step. In *hard* mode, the customer has low patience, gives vague or fragmented information, and may provide negative feedback. In *static* mode, the customer sends a single comprehensive request with no follow-up.
70
+
71
+ 2. **Service Agent** — The LLM being evaluated. It receives tool definitions for the scenario, processes the user's request, invokes tools via structured JSON output, and responds in natural language. Tool calls are intercepted, executed against an scenario database, and the results are fed back into the conversation.
72
+
73
+ 3. **Evaluation Pipeline** — After simulation, tool calls and database state are compared against scenario ground truth using fuzzy matching to compute success metrics. Results feed into visualization and report generation scripts.
74
+
75
+ ```
76
+ ┌─────────────┐ message ┌──────────────┐
77
+ │ Simulated │ ───────────────▶ │ Service Agent │
78
+ │ User │ ◀─────────────── │ (LLM under │
79
+ │ │ response │ test) │
80
+ └─────────────┘ └──────┬───────┘
81
+ │ tool call (JSON)
82
+ ▼
83
+ ┌──────────────┐
84
+ │ Scenario DB │
85
+ │ │
86
+ └──────────────┘
87
+ ```
88
+
89
+ ## Supported Scenarios
90
+
91
+ | Scenario | #Variants | #Tasks |
92
+ |------------|-----------|--------|
93
+ | Retail | 10 | 454 |
94
+ | Kitchen | 4 | 175 |
95
+ | Restaurant | 6 | 316 |
96
+ | Warehouse | 25 | 375 |
97
+ | Household | 18 | 270 |
98
+ | **Total** | **63** | **1,590** |
99
+
100
+ Each scenario variant provides a different database configuration (product catalogs, menus, etc.) loaded from `tools/{scenario}/{scenario}_init.py`.
101
+
102
+ ## Supported Models
103
+
104
+ | Model | Provider | API |
105
+ |-------|----------|-----|
106
+ | `glm-5v-turbo` | Zhipu AI | Zhipu AI SDK |
107
+ | `qwen3-vl-235b` | Alibaba Qwen | Custom endpoint |
108
+ | `Qwen3.5-397B-A17B` | Alibaba Qwen | OpenAI-compatible |
109
+ | `qwen3.6-plus` | Alibaba Qwen | OpenAI-compatible |
110
+ | `gemini-3.1-pro-preview` | Google | OpenAI-compatible proxy |
111
+ | `kimi-k2.6` | Moonshot | OpenAI-compatible |
112
+ | `mimo-v2.5-omni` | Xiaomi MiMo | OpenAI-compatible |
113
+ | `doubao-seed-2-0-pro-260215` | ByteDance Doubao | OpenAI-compatible |
114
+ | `manual` | — | Terminal input for manual testing |
115
+
116
+ ## Project Structure
117
+
118
+ ```
119
+ ├── run/ # Core simulation framework
120
+ │ ├── multi_agent.py # Main entry point
121
+ │ ├── utils.py # LLM API calls, tool execution, message building
122
+ │ ├── prompts.py # Prompt templates for all agents
123
+ │ └── apis/ # LLM API wrappers
124
+ │ ├── unified.py # Unified call_llm dispatcher
125
+ │ ├── zhipu.py # Zhipu AI
126
+ │ ├── qwen.py # Qwen
127
+ │ ├── qwen3_5.py # Qwen3.5
128
+ │ ├── kimi.py # Kimi
129
+ │ ├── mimo.py # MiMo
130
+ │ └── doubao.py # Doubao
131
+ ├── tools/ # Scenario databases & tool definitions
132
+ │ ├── retail/ # RetailDB, retail_init.py, retail_tools.json
133
+ │ ├── kitchen/ # KitchenDB, kitchen_init.py, kitchen_tools.json
134
+ │ ├── restaurant/ # RestaurantDB plus Restaurant 6 multi-restaurant backend/schema
135
+ │ ├── warehouse/ # WarehouseDB and variants 1-25
136
+ │ ├── household/ # HouseholdDB and variants 1-18
137
+ │ ├── database_init.py # DB initialization utilities
138
+ │ └── user_words.py # Extra phrases for hard-mode user agent
139
+ ├── scenarios/
140
+ │ └── final/ # Task definitions and ground truth (JSON)
141
+ │ ├── video/ # Version-controlled benchmark video assets
142
+ │ ├── retail1.json .. retail10.json
143
+ │ ├── kitchen1.json .. kitchen4.json
144
+ │ ├── restaurant1.json .. restaurant6.json
145
+ │ ├── warehouse1.json .. warehouse25.json
146
+ │ └── household1.json .. household18.json
147
+ ├── analysis_scripts/ # Evaluation & visualization
148
+ │ ├── evaluate_interaction.py # Core evaluation script
149
+ │ ├── print_eval.py # Pretty-print evaluation results
150
+ │ ├── analyze_errors.py # Error categorization
151
+ │ ├── analyze_tool_correctness.py
152
+ │ ├── visualize_results.py # Chart generation
153
+ │ ├── plot_results.py # Additional plotting
154
+ │ ├── generate_comprehensive_report.py
155
+ │ └── run_eval.sh # Run evaluation across all models
156
+ ├── run_all.sh # Launch all model runs in parallel
157
+ ├── run_scenario.sh # Configurable multi-model runner
158
+ ├── run_glm.sh / run_qwen3.sh / … # Model-specific runners
159
+ └── requirements.txt
160
+ ```
161
+
162
+ <div align="center">
163
+ <h3>🚀 Quick Start</h3>
164
+ </div>
165
+
166
+ ## Getting Started
167
+
168
+ ### Prerequisites
169
+
170
+ - Python 3.10+
171
+ - API keys for the models you want to evaluate
172
+
173
+ ### API Key Configuration
174
+
175
+ All API keys and base URLs are managed through environment variables. A `.env.example` file is provided as a template.
176
+
177
+ 1. Copy the example file and fill in your keys:
178
+ ```bash
179
+ cp .env.example .env
180
+ ```
181
+
182
+ 2. Edit `.env` and replace the placeholder values with your actual API keys and endpoints:
183
+ ```bash
184
+ # Main API key for LLM platform models
185
+ # Used by: Qwen3.5, Qwen3-VL, Gemini, DeepSeek, GLM-4.5V, Doubao, Qwen3.6, etc.
186
+ export API_KEY="your-api-key-here"
187
+
188
+ # Base URL for LLM API endpoints (Qwen, Doubao, Gemini, etc.)
189
+ # Default: https://api.example.com/v1
190
+ export LLM_API_BASE_URL="https://your-llm-api-endpoint.com/v1"
191
+
192
+ # Kimi (Moonshot) API key
193
+ # Used by: kimi-k2.6
194
+ export KIMI_API_KEY="your-kimi-api-key-here"
195
+
196
+ # Kimi API base URL (optional, defaults to Moonshot's official endpoint)
197
+ export KIMI_API_BASE_URL="https://api.moonshot.cn/v1"
198
+
199
+ # Xiaomi MiMo API key
200
+ # Used by: mimo-v2.5-omni
201
+ export MIMO_API_KEY="your-mimo-api-key-here"
202
+
203
+ # MiMo API base URL (optional, defaults to Xiaomi's official endpoint)
204
+ export MIMO_API_BASE_URL="https://api.xiaomimimo.com/v1"
205
+
206
+ # Zhipu AI API key
207
+ # Used by: glm-5v-turbo, glm-4.5v
208
+ export ZHIPU_API_KEY="your-zhipu-api-key-here"
209
+ ```
210
+
211
+ 3. Source the `.env` file before running simulations:
212
+ ```bash
213
+ source .env
214
+ ```
215
+
216
+ > **Note:** The shell scripts (`run_glm.sh`, `run_kimi.sh`, etc.) automatically source `.env` when executed. The `.env` file is excluded from version control via `.gitignore`.
217
+
218
+ | Environment Variable | Required | Default | Description |
219
+ |---------------------|----------|---------|-------------|
220
+ | `API_KEY` | Yes* | - | API key for Qwen, Gemini, DeepSeek, GLM, Doubao models |
221
+ | `LLM_API_BASE_URL` | Yes* | `https://api.example.com/v1` | Base URL for LLM API endpoints |
222
+ | `KIMI_API_KEY` | Yes* | - | Moonshot Kimi API key |
223
+ | `KIMI_API_BASE_URL` | Yes* | `https://api.moonshot.cn/v1` | Kimi API base URL |
224
+ | `MIMO_API_KEY` | Yes* | - | Xiaomi MiMo API key |
225
+ | `MIMO_API_BASE_URL` | Yes* | `https://api.xiaomimimo.com/v1` | MiMo API base URL |
226
+ | `ZHIPU_API_KEY` | Yes* | - | Zhipu AI API key |
227
+
228
+ *Required only if using the corresponding models.
229
+
230
+ ### Video Configuration
231
+
232
+ The required video files are version-controlled in this Git repository under
233
+ `scenarios/final/video/`. Clone the repository to download them together with
234
+ the task definitions and ground truth.
235
+
236
+ The framework uses `<repository>/scenarios/final/video` by default. To keep the
237
+ videos elsewhere, pass
238
+ `--video_dir` or set `EGOBENCH_VIDEO_DIR`:
239
+
240
+ ```bash
241
+ export EGOBENCH_VIDEO_DIR="/absolute/path/to/video"
242
+ ```
243
+
244
+ Legacy scenario entries containing a URL are resolved by filename against this
245
+ local directory. Extension-only differences such as `.MOV` versus `.mp4` are
246
+ also handled automatically.
247
+
248
+ ### Installation
249
+
250
+ ```bash
251
+ pip install -r requirements.txt
252
+ ```
253
+
254
+ ### Running a Simulation
255
+
256
+ ```bash
257
+ # Set up API keys (required before first run)
258
+ cp .env.example .env # then edit .env with your keys
259
+ source .env
260
+
261
+ # Run a single simulation
262
+ python run/multi_agent.py \
263
+ --scenario retail \
264
+ --scenario_number 1 \
265
+ --user_mode easy \
266
+ --service_model_name Qwen3.5-397B-A17B \
267
+ --multi_agent_user \
268
+ --summary_user
269
+
270
+ # Run with limited samples
271
+ python run/multi_agent.py \
272
+ --scenario kitchen \
273
+ --scenario_number 2 \
274
+ --user_mode hard \
275
+ --service_model_name doubao-seed-2-0-pro-260215 \
276
+ --num_samples 10
277
+
278
+ # Run a Household scenario
279
+ python run/multi_agent.py \
280
+ --scenario household \
281
+ --scenario_number 1 \
282
+ --user_mode easy \
283
+ --service_model_name kimi-k2.6 \
284
+ --num_samples 1
285
+
286
+ # Rerun only previously failed scenarios
287
+ python run/multi_agent.py \
288
+ --scenario restaurant \
289
+ --scenario_number 1 \
290
+ --user_mode easy \
291
+ --service_model_name qwen3-vl-235b \
292
+ --num_samples 0
293
+ ```
294
+
295
+ ### Key Arguments
296
+
297
+ | Argument | Values | Default | Description |
298
+ |----------|--------|---------|-------------|
299
+ | `--scenario` | `retail`, `kitchen`, `restaurant`, `warehouse`, `household` | `retail` | Domain scenario |
300
+ | `--scenario_number` | scenario-specific integer | `1` | Retail 1-10, Kitchen 1-4, Restaurant 1-6, Warehouse 1-25, Household 1-18 |
301
+ | `--user_mode` | `easy`, `hard`, `static` | `easy` | User agent difficulty |
302
+ | `--service_model_name` | See [Supported Models](#supported-models) | `qwen3-vl-235b` | LLM to evaluate |
303
+ | `--multi_agent_user` | flag | off | Enable contradiction checker for user responses |
304
+ | `--summary_user` | flag | off | Enable dialogue summarization to manage context length |
305
+ | `--num_samples` | int | `0` | Limit scenarios per run (`0` = all) |
306
+
307
+ ### Running Evaluation
308
+
309
+ ```bash
310
+ # Evaluate a specific model
311
+ cd analysis_scripts
312
+ python evaluate_interaction.py --model_name glm-4.5v --num_samples 10
313
+
314
+ # Evaluate all models
315
+ bash run_eval.sh
316
+ ```
317
+
318
+ ### Batch Execution
319
+
320
+ ```bash
321
+ # Run all models in parallel
322
+ bash run_all.sh
323
+
324
+ # Run a specific model across all scenarios and modes
325
+ bash run_glm.sh
326
+ bash run_qwen3.5.sh
327
+ ```
328
+
329
+ ## User Modes
330
+
331
+ ### Easy Mode
332
+ The user agent acts as a cooperative customer who provides information step-by-step, responds clearly to questions, and signals completion with "STOP" when all requirements are met.
333
+
334
+ ### Hard Mode
335
+ The user agent simulates a challenging customer with low patience, provides vague or incomplete information initially, may give negative feedback ("Bad Service Agent"), and passively introduces contradictions. Additional distractor phrases from `tools/user_words.py` are injected to increase difficulty.
336
+
337
+ ### Static Mode
338
+ A single-turn interaction where the user agent sends one comprehensive request containing all requirements at once. The service agent must process everything without follow-up clarification.
339
+
340
+ ## Evaluation
341
+
342
+ The evaluation pipeline in `analysis_scripts/evaluate_interaction.py` compares each service agent's tool calls against the scenario ground truth:
343
+
344
+ - **Tool-call matching** — Verifies that the correct tools were invoked with the correct parameters
345
+ - **Fuzzy matching** — Handles minor variations in parameter values (e.g., product names) using fuzzy string matching
346
+ - **Database verification** — Re-executes tool calls against the scenario database to verify result correctness
347
+ - **Success metrics** — Computes per-scenario and aggregate success rates across difficulty levels
348
+
349
+ ### Visualization & Reports
350
+
351
+ ```bash
352
+ # Generate charts and reports
353
+ cd analysis_scripts
354
+ python visualize_results.py
355
+ python plot_results.py
356
+ python generate_comprehensive_report.py
357
+ ```
358
+
359
+ Output charts include success rates by difficulty/scenario, error distributions, token usage correlations, and dialogue round analysis. Charts are saved to `eval_result/figures/`.
360
+
361
+ ## Output Format
362
+
363
+ Simulation results are saved to `results/{model_name}/{scenario}{number}_{mode}.json`:
364
+
365
+ ```json
366
+ {
367
+ "scenario_id": 0,
368
+ "mode": "easy",
369
+ "instruction": "...",
370
+ "image_description": "...",
371
+ "dialogue": [{"role": "user", "turn": 1, "content": "..."}, ...],
372
+ "tool_calls": [{"turn": 1, "calls": [...], "results": [...]}],
373
+ "rounds_count": 5,
374
+ "input_tokens": 12345,
375
+ "output_tokens": 6789,
376
+ "tool_calls_count": 8,
377
+ "user_response_time_seconds": 2.3,
378
+ "agent_response_time_seconds": 5.1,
379
+ "execution_time_seconds": 42.0,
380
+ "user_performance": {...}
381
+ }
382
+ ```
383
+
384
+ Evaluation results are saved to `eval_result/{model_name}/{scenario}{number}_{mode}_eval.json`.
385
+
386
+ ![Code License](https://img.shields.io/badge/code-MIT-blue.svg)
387
+ ![Data License](https://img.shields.io/badge/data-CC%20BY%204.0-lightgrey.svg)
388
+
389
+ ## License
390
+
391
+ EgoBench uses separate licenses for code and data:
392
+
393
+ - Source code, including Python files, shell scripts, and evaluation utilities,
394
+ is licensed under the [MIT License](./LICENSE-CODE).
395
+ - Benchmark task definitions, annotations, JSON files, scenario databases, and
396
+ video files under `scenarios/final/video/` are licensed under the
397
+ [Creative Commons Attribution 4.0 International License](./DATA_LICENSE.md).
398
+
399
+ The included video assets are distributed with authorization from their
400
+ respective rightsholders. Third-party components, if any, remain subject to
401
+ their respective licenses. See the repository [licensing notice](./LICENSE)
402
+ before using or redistributing individual components.
403
+
404
+ ## Citation
405
+
406
+ If you find EgoBench useful in your research, please cite our paper:
407
+
408
+ ```bibtex
409
+ @inproceedings{liu2026egobench,
410
+ title = {EgoBench: An Interactive Egocentric Multimodal Benchmark for Tool-Using Agents},
411
+ author = {Yunqi Liu and Tong Niu and Zitong Wang and Zhenlong Dai and Yuqi Qing and Weiqiang Wang and Jian Liu},
412
+ booktitle = {Advances in Neural Information Processing Systems},
413
+ year = {2026},
414
+ url = {https://arxiv.org/abs/2605.27820}
415
+ }
416
+ ```
417
+
418
+ ## Dataset Viewer
419
+
420
+ The `test` split contains one row per benchmark task (1,590 rows). The table provides searchable task text, scenario identifiers, the original ground-truth tool calls as JSON text, and a direct link to the associated MP4 video. The original task definitions remain in `scenarios/final/*.json`; the viewer table is a browsing copy and does not replace those files.
data/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a0bbf1d6c827d740baedbd04a6db8e9d5676aa0b01c19312b0c416a3281556ac
3
+ size 682002