Authorization Closure Graph: Minimal Repair for LLM Agents with Evolving User Instructions

arxiv: https://arxiv.org/pdf/2609.32428

Implementation of ACG:

• main.py: experiment entry point.
• src/acg/: authorization graph, repair, confirmation, transport, and Airline integration. Airline specifications are embedded in airline_spec.py.

Run

Use Python 3.12. Install tau2-bench separately at revision b351ed5f9281d4bdfa5629262f54c8781da0d5be with pip install -e . from its checkout. Its code, data, and license are not bundled here.

Set these environment variables before running:

• TAU2_DATA_DIR: the frozen benchmark data directory used for the experiments. The entry point verifies the Airline database, tasks, policy, and user-simulator instructions before making requests.
• MODEL_BASE_URL, MODEL_API_KEY: your OpenAI-compatible endpoint and credential.
• AGENT_MODEL: acting model name.
• USER_MODEL: simulator model; defaults to the acting model, using a separate conversation.
• AGENT_EXTRA_BODY, USER_EXTRA_BODY: provider-specific decoding JSON. For DeepSeek, use {"thinking":{"type":"disabled"}}; for GPT, use {"reasoning_effort":"none"}.

Optional role-specific settings are AGENT_BASE_URL, USER_BASE_URL, AGENT_API_KEY, and USER_API_KEY. Use AGENT_RESPONSE_MODEL or USER_RESPONSE_MODEL when a provider returns a different model identifier, and AGENT_TOOL_SCHEMA_MODE=inline_local_refs if required by the provider.

python main.py --output results

Defaults: all 50 Airline tasks, one trial, seed 300, temperature 0, 200 steps, and concurrency 3. Use --tasks 0 1 for a small run or --concurrency 8 to adjust concurrency. The output directory must not already exist. results.json contains complete conversations and official rewards; summary.json contains task success.