ToolAlignBench

Code and dataset for "Investigating Alignment Conflicts in Tool-Calling Enabled LLMs" (Published at the ICML 2026 Workshop on Pluralistic Alignment).

• Paper: https://openreview.net/pdf?id=KJTiUm8b7d
• Dataset on HuggingFace: huggingface.co/datasets/aryankeluskar/ToolAlignBench

Repository layout

Directory	Description
benchmark/	Prompts (128 markdown files) + per-domain tool taxonomy.
runner/	Drives an LLM through each scenario, creates one markdown transcript per (model, domain, scenario, run, condition).
analysis/	Parses transcripts → behavior labels → tables, statistical tests, and figures.
inference/	FastAPI shim that serves local HuggingFace models behind an OpenAI-compatible API with tool-calling. Used for the abliterated open-source models.
dataset/	Script that packages benchmark/ into the JSONL files uploaded to HuggingFace.

Quick start

git clone https://github.com/aryankeluskar/ToolAlignBench
cd ToolAlignBench
cp .env.example .env
# then fill in OPENROUTER_API_KEY

# Reproduce one model on one scenario type
cd runner
bun install
bun run start --safe

# Analyze the transcripts and reproduce figures
cd ../analysis
bun install
bun run analyze            
bun run plot               

The default model in runner/src/models.ts is meta-llama/llama-3.1-8b-instruct via OpenRouter. Uncomment additional entries to reproduce the full Table 1.

The four pharmaceutical-distribution scenarios are derived from SnitchBench (MIT License, Theo Browne). All other scenarios are original work.

Citation

@inproceedings{
    keluskar2026toolalignbench,
    title={ToolAlignBench: Investigating Alignment Conflicts in Tool-Calling Enabled {LLM}s},
    author={Keluskar, Aryan and Bhattacharjee, Amrita and Liu, Huan},
    booktitle={Pluralistic Alignment Workshop at ICML 2026},
    year={2026},
    url={https://openreview.net/forum?id=KJTiUm8b7d}
}

License

See LICENSE. The benchmark documents are released under CC BY 4.0 via the HuggingFace dataset card.

Responsible use

This benchmark contains synthetic documents describing fabricated organizational wrongdoing across regulated domains. The scenarios are designed to probe how LLM agents resolve conflicts between safety training and deployment instructions; they are not templates for real misconduct. The "abliterated" models we evaluate are publicly available and were used solely to isolate the effect of safety training on tool-calling behavior. See the paper's Impact Statement (§ Impact Statement) for the full discussion.