NOTICE
Cost-Aware Protocol Routing: Matched Protocol Outcomes

This dataset accompanies:

  "LLMs Can Predict Failure Risk, But Struggle to Predict Which Collaboration
   Protocol Pays Off: Cost-Aware Protocol Routing Across Reasoning Tasks"
  Chih-Hsuan Yang, Jingyan Jiang, Cheng-Hau Yang, Vikram Vasudevan,
  Huihuo Zheng, Venkatram Vishwanath, Rajeev Thakur.
  EMNLP 2026. arXiv:2608.14927

ATTRIBUTION FOR UPSTREAM BENCHMARKS
The derived outcomes in this release were measured on problems from the
following benchmarks. Their problem text is NOT redistributed here; only stable
identifiers are. Anyone who reconstructs the problems from upstream must comply
with the upstream license and cite the upstream work.

  Omni-MATH-2 (Apache-2.0)
    https://huggingface.co/datasets/martheballon/Omni-MATH-2
    Derived from Omni-MATH, https://arxiv.org/abs/2410.07985

  JEEBench (MIT), Copyright (c) 2023 Data Analytics and Intelligence Research
  (DAIR) Group, IIT Delhi
    https://github.com/dair-iitd/jeebench
    Arora, Singh, Mausam. "Have LLMs Advanced Enough? A Challenging Problem
    Solving Benchmark For Large Language Models." EMNLP 2023.

  SciBench (MIT), Copyright (c) 2023 Xiaoxuan Wang
    https://github.com/mandyyyyii/scibench
    Wang et al. "SciBench." ICML 2024.

  LAB-Bench (CC-BY-SA-4.0), FutureHouse
    https://huggingface.co/datasets/futurehouse/lab-bench
    Upstream revision 5c77cec648430f30611808808861eb86f81d5eaa.
    LAB-Bench ships a canary string and an accompanying do-not-train request.
    This release contains no LAB-Bench problem text, options, or answers, so it
    carries no canary; do not use the identifiers here to assemble a training
    corpus that would violate that request.

  MaScQA (CC-BY-NC-SA-4.0) is NOT part of this release. It has no matched Gemma
  run and is excluded from the paper. Its NonCommercial terms are also
  incompatible with this release's licensing.

MODEL TERMS
  openai/gpt-oss-120b   Apache-2.0 weights
  google/gemma-4-31B-it Gemma Terms of Use, https://ai.google.dev/gemma/terms

ACKNOWLEDGMENT
This research used resources of the Argonne Leadership Computing Facility, a
U.S. Department of Energy (DOE) Office of Science user facility at Argonne
National Laboratory (ANL) operated under Contract No. DE-AC02-06CH11357.

CONTACT
bellayang@anl.gov