MTAC-IFBench: Benchmarking Instruction-Following in Multi-Turn Agentic Coding

🌟 Overview

MTAC-IFBench benchmarks instruction following in multi-turn agentic coding.

Existing agentic coding benchmarks (e.g., SWE-bench, Terminal-Bench) focus on final functional correctness, while current instruction-following benchmarks confine themselves to single-turn chat or code generation. Neither answers the question that matters in a real development session: does the agent keep following the rules, turn after turn, as the requirements change?

In real multi-turn software development, an agent must comply with:

• Repository policy files (CLAUDE.md, AGENTS.md) that govern the whole session
• Per-turn constraints on the response, the code, the environment, and its own workflow
• Constraints that persist across turns without decaying as context grows
• Constraints that are added, revised, or overridden by later instructions

Each instance in MTAC-IFBench is a complete development session: a repository policy file, 5–10 progressive user instructions, a constraint checklist for every turn, and a function checklist for the finished project. Each checklist item is verified by a verification script or a judge agent otherwise. MTAC-IFBench identifies significant deficiencies in existing code agents in multi-turn instruction-following, with performance degrading rapidly as the interaction session grows longer. See our paper for full results.

📊 Data statistics

Metric	Value
Instances	100 (full) / 20 (lite)
Constraints	9,133
Avg. turns per instance	7.04 (sequences span 5 to 10 turns)
Avg. constraint checklist items	91.33 per instance, 12.97 per turn
Avg. function checklist items	16.07 per instance

📦 Constraint taxonomy

Constraints in MTAC-IFBench span 6 primary and 18 secondary categories, covering the generated response and code as well as the agent's environment interactions and workflow.

Primary Category	Secondary Categories	Description	Example
Content	Keyword, Persona, Format	Lexical elements, adopted persona, and structural templates in the generated code or responses	Every response must end with the modal particle "meow~".
Language	Response Language, Comment Language, File Encoding	The language used for responses and comments, and the character encoding of created files	All comments in every code file you generate must be in Chinese.
Quantity	Range, Exact Value, Complexity	Quantitative attributes of generated artifacts, as ranges, exact values, or complexity caps	Each file you generate must contain between 200 and 500 lines.
Style	Layout, Naming, Paradigm	Spatial organization, naming conventions, and the programming style used in implementation	Every function name in the code files you generate must follow snake case.
Environment	File Path, File Operation, Logging	How files are referenced, placed, and modified, and what record the agent leaves of its work	Before modifying an existing file each time, you must create a backup of the original file in the same directory.
Workflow	Tool Usage, Orchestration, Testing	Tool selection and sequencing of multi-step actions, and code verification practices	You must execute multiple independent tool calls in parallel as much as possible to improve efficiency.

⚙️ Data format

Each line is one instance, a JSON object with the following fields:

{
  "task_id": 1,
  "task_category": "Data Analysis",
  "repository_policy": "# 项目规范\n所有的回复内容都应以语气词\"喵～\"起头 ...",
  "repository_policy_checklist": [
    {
      "constraint": "所有的回复内容都应以语气词\"喵～\"起头",
      "validation_code": "def check_requirement(response: str, workspace_path: str) -> bool:\n    ...",
      "tags": ["Content", "Persona"]
    },
    ...
  ],
  "rounds": [
    {
      "round_id": 0,
      "instruction": "帮我写一个监控表格页面的基础框架。页面需要一个大标题 ...",
      "instruction_following_checklist": [
        {
          "constraint": "...", 
          "validation_code": "...", 
          "tags": ["Style", "Layout"]
        },
        ...
      ]
    },
    ...
  ],
  "function_checklist": [
    "查看页面整体结构，确认页面顶部包含大标题 ...",
    ...
  ]
}

Field	Description
task_id	Instance id (1–100)
task_category	Development domain: Frontend Development, Data Analysis, Application Development, UI/UX Optimization, Build & Deployment, or Machine Learning
repository_policy	Repository policy file content (e.g. CLAUDE.md / AGENTS.md), imposing global constraints over the whole session
repository_policy_checklist	Constraint checklist for the repository policy file; applies to every turn
rounds	The multi-turn instruction sequence. Each entry carries a round_id, the user instruction for that turn, and an instruction_following_checklist holding all constraints in force at that turn, including constraints from the repository_policy file
function_checklist	Functional requirements for the final project

🚀 Usage

Evaluation uses AgentProbe, a sandbox framework for coding-agent assessment.

1. Set up

git clone https://github.com/abelperry/AgentProbe.git && cd AgentProbe
uv sync
./scripts/init.sh && source .agentprobe-env

uv pip install huggingface_hub
python scripts/pull_benchmarks.py --repo mtacifbench=thu-coai/MTAC-IFBench

2. Place the data

The adapter reads benchmarks/mtacifbench/data/questions.jsonl:

cd benchmarks/mtacifbench/data
cp data/full/questions.jsonl questions.jsonl    # or data/lite/questions.jsonl
cp eval_config/judge.yaml .                     # or judge_if_function.yaml

judge.yaml scores instruction-following only, while judge_if_function.yaml also builds the final project and checks the function checklist.

3. Configure the agent

In examples/exp-mtacifbench.yaml, models: is the LLM to be evaluated and agents: is the harness driving it.

models:
  your-model:
    base_url: "${GATEWAY_BASE_URL}"
    api_key: "${GATEWAY_API_KEY}"
    model_name: "your-model"
    format: "anthropic"

agents:
  claude_code:
    type: "agent_probe.agents.claude_code.ClaudeCodeAgent"
    version: "2.1.14"
    offline: true
    offline_package_dir: ${OFFLINE_PACKAGE_DIR}
  # opencode:
  #   type: "agent_probe.agents.opencode.OpenCodeAgent"
  #   version: "1.1.21"
  #   params: {output_format: "json"}

Every agent listed runs against every model listed, so you can uncomment opencode to compare one model across both harnesses.

4. Start evaluation

export GATEWAY_BASE_URL=... GATEWAY_API_KEY=...
uv run agentprobe -c examples/exp-mtacifbench.yaml -l info

Results land under output/{experiment}/{dataset}/{agent}/{model}/, with aggregated metrics in metrics.jsonl.

👏 Citation

@article{wen2026mtacifbench,
  title   = {MTAC-IFBench: Benchmarking Instruction-Following in Multi-Turn Agentic Coding},
  author  = {Wen, Bosi and Wang, Cunxiang and Gui, Jiayi and Zhang, Haoke and
             Niu, Yilin and Ke, Pei and Yang, Dayong and Wang, Hongning and Huang, Minlie},
  journal = {arXiv preprint arXiv:2609.14992},
  year    = {2026}
}

Please kindly cite our paper if this paper and the codes are helpful.