SPIEval

SPIEval: Evaluating Large Language Models as Mobile Assistants over Scattered Personal Information

SPIEval is a human-curated benchmark for evaluating whether large language models can act as mobile assistants by proactively retrieving and reasoning over personal information scattered across multiple applications. Given an underspecified user instruction, a model must search structured records, recover the information required for execution, and invoke the appropriate tools with correct arguments.

All personal information in SPIEval is entirely fictional and does not correspond to any real individual.

Highlights

• Realistic mobile-assistant setting: task-relevant information is distributed across structured records from 10 commonly used mobile applications.
• Five cognitive capabilities: Reasoning, Disambiguation, Integration, Preference Inference, and Multi-Intent Decomposition.
• Active information retrieval: models must formulate queries, inspect retrieved records, and decide when further retrieval is necessary.
• Interactive tool use: 11 retrieval tools and 10 execution tools support multi-turn interaction.
• Judge-free evaluation: final execution calls are automatically compared with human-annotated gold calls at the parameter level.
• Human-curated and privacy-safe: all tasks and records were manually constructed, and all personal data is fictional.

Benchmark Overview

Statistic	Value
Tasks	250
Cognitive capabilities	5
Tasks per capability	50
Personal records	4,335
Simulated applications	10
Retrieval tools	11
Execution tools	10
Execution operations	357

Leaderboard

Accuracy (%) under the standard paginated-retrieval protocol is reported below. Each value is the mean over three independent runs. Models are evaluated at their highest and lowest available reasoning-effort settings and ordered by overall accuracy.

Model	Reasoning	Disambiguation	Integration	Preference Inference	Multi-Intent Decomposition	Overall
GPT-5.5 (xhigh)	70.0	72.0	73.3	38.7	32.7	57.3
Gemini 3.1 Pro (high)	74.0	73.3	74.7	20.0	23.3	53.1
Claude Opus 4.8 (max)	68.7	58.7	67.3	34.7	32.0	52.3
DeepSeek-V4-Pro (max)	53.3	62.7	60.0	34.0	22.7	46.5
Gemini 3.1 Pro (low)	65.3	62.7	60.7	16.7	23.3	45.7
Claude Opus 4.8 (none)	48.7	40.7	46.0	26.7	19.3	36.3
Kimi K2.6 (high)	48.7	46.7	43.3	25.3	16.0	36.0
GLM-5.2 (max)	44.7	44.7	49.3	22.7	14.7	35.2
Hy3 (high)	44.0	41.3	46.0	26.0	14.0	34.3
DeepSeek-V4-Pro (none)	36.0	39.3	48.7	28.7	18.7	34.3
Qwen3.7-Plus (high)	44.0	38.7	46.7	23.3	14.7	33.5
Seed-2.1-Pro (high)	50.0	44.0	39.3	17.3	16.0	33.3
GLM-5.2 (none)	37.3	32.0	46.7	19.3	10.7	29.2
GPT-5.5 (none)	41.3	35.3	30.7	22.0	13.3	28.5
Seed-2.1-Pro (minimal)	34.7	32.7	27.3	18.7	9.3	24.5
Qwen3.7-Plus (none)	34.7	27.3	36.0	12.7	8.7	23.9
Hy3 (no_think)	25.3	24.7	23.3	13.3	6.0	18.5
Kimi K2.6 (none)	17.3	26.0	19.3	11.3	8.0	16.4

The simulated applications are:

1. Accommodation
2. Alarm
3. Contacts
4. Meeting
5. Notes
6. Schedule
7. SMS
8. Transactions
9. Transport
10. Voicemail

Cognitive Capabilities

Reasoning

The model follows dependencies across multiple records. Information retrieved at one step may be required to formulate the next query.

Disambiguation

The model distinguishes the target record from multiple plausible candidates using contextual and temporal evidence.

Integration

The model aggregates complementary information from multiple records or applications to complete an action.

Preference Inference

The model infers implicit user preferences from historical records rather than relying on preferences explicitly stated in the instruction.

Multi-Intent Decomposition

The model decomposes a compound request into multiple coordinated actions and recovers the parameters required by each action.

Repository Structure

SPIEval/
├── data/
│   ├── input_data.jsonl       # Instructions, annotations, and gold tool calls
│   ├── records_db.json        # Task-specific records from the simulated apps
│   └── tools.json             # Schemas for retrieval and execution tools
├── response_generator.py      # Interactive model evaluation environment
├── evaluation_main.py         # Parameter-level automatic evaluator
├── requirements.txt
├── LICENSE                     # Apache License 2.0 for the source code
└── README.md

Data Format

data/input_data.jsonl

Each line represents one task:

{
  "id": "推理1",
  "type": "推理",
  "messages": [
    {
      "role": "user",
      "content": "..."
    }
  ],
  "prompt_metadata": {
    "answer": [
      {
        "name": "Contacts_Create",
        "arguments": {
          "name": ["..."],
          "telephone": ["..."]
        }
      }
    ],
    "free_params": [["company", "relationship", "note"]]
  },
  "reasoning_path": "...",
  "distractors": "..."
}

The five values of type are 推理 (Reasoning), 辨析 (Disambiguation), 整合 (Integration), 偏好 (Preference Inference), and 多意图 (Multi-Intent Decomposition).

For each gold argument, the value is represented as a list of acceptable values. A null entry indicates that the corresponding parameter may be omitted. Parameters listed in free_params may take any value valid under the tool schema.

data/records_db.json

Records are indexed by task ID and then grouped by application:

{
  "推理1": {
    "contacts": [],
    "voicemail": [],
    "sms": [],
    "notes": [],
    "alarm": [],
    "meeting": [],
    "schedule": [],
    "transport": [],
    "accommodation": [],
    "transaction": []
  }
}

Records are task-specific. This prevents information from one task from leaking into another task and enables each task to have a controlled information environment.

data/tools.json

This file contains OpenAI-style function schemas for:

• 10 application-specific retrieval tools;
• 1 global retrieval tool;
• 10 execution tools.

Application-specific retrieval supports substring, regular-expression, and fuzzy matching, as well as field-specific and case-sensitive search. Under the standard paginated setting, a retrieval call returns at most five records per page.

Installation

Python 3.9 or later is recommended.

pip install -r requirements.txt

Model Integration

Before running inference, implement the client() factory in response_generator.py for the model provider being evaluated. The returned client must expose:

client.chat(messages=messages, tools=tools)

The response should follow the OpenAI-compatible chat-completion format and include tool calls in choices[0].message.tool_calls.

API credentials should be supplied through command-line arguments or environment variables. Do not commit credentials to the repository.

Running Evaluation

1. Generate model responses

The standard SPIEval protocol uses paginated retrieval:

python response_generator.py \
  --model MODEL_NAME \
  --input_file data/input_data.jsonl \
  --output_dir outputs/MODEL_NAME \
  --search_mode paginated \
  --thinking_level high \
  --repeat 3 \
  --max_workers 4

Additional controlled settings are provided for analysis:

• paginated: retrieval results are paginated, with at most five records per page;
• full: each retrieval call returns all matching records;
• no_search: all task-relevant records are placed directly in the system prompt.

Useful arguments include:

Argument	Description
--model	Model identifier passed to the client implementation
--base_url	Optional API endpoint
--api_key	Optional API key
--thinking_level	Provider-specific reasoning-effort setting
--repeat	Number of independent runs per task
--max_workers	Number of concurrent workers
--start_id, --end_id	Evaluate a selected range of repeated samples

The generated responses are saved to:

outputs/MODEL_NAME/response.jsonl

2. Score model responses

python evaluation_main.py \
  --input_data data/input_data.jsonl \
  --input_response_data outputs/MODEL_NAME/response.jsonl \
  --output_file outputs/MODEL_NAME/result.json \
  --overload

This command produces:

• result.json: mean accuracy and standard deviation overall and for each capability;
• details.jsonl: task-level correctness labels and diagnostic information.

Evaluation Metric

SPIEval uses strict binary task accuracy. A task is correct only when the predicted execution tools and all constrained parameter values match a valid human-annotated gold answer.

The evaluator additionally supports:

• multiple acceptable values for a parameter;
• unconstrained parameters whose values are validated only against the tool schema;
• order-invariant matching when a task requires multiple execution calls.

Intermediate retrieval trajectories are not used to determine correctness because multiple valid retrieval strategies may lead to the same final execution.

License

SPIEval uses separate licenses for its dataset and source code:

• Dataset: All files under data/ are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may share and adapt the dataset for any purpose, including commercial use, provided that appropriate credit is given, a link to the license is included, and any modifications are indicated.
• Code: The Python source code is released under the Apache License 2.0. See LICENSE for the full license text.

When using SPIEval, please cite the benchmark and include a link to this repository.

Citation

If you use SPIEval in your research, please cite:

@misc{ye2026spievalevaluatinglargelanguage,
      title={SPIEval: Evaluating Large Language Models as Mobile Assistants over Scattered Personal Information}, 
      author={Junjie Ye and Zhuohui Sheng and Shaofan Liu and Yulun Zhu and Wenjie Fu and Dingwei Zhu and Ming Zhang and Yujiong Shen and Weichao Wang and Xin Zhao and Shihan Dou and Tao Gui and Qi Zhang and Xuanjing Huang and Pluto Zhou},
      year={2026},
      eprint={2608.10692},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.10692}, 
}