facebook/omnilingual-gaia2

OmnilingualGAIA2

OmnilingualGAIA2 is the multilingual extension of the GAIA2 agentic benchmark: 6400 scenarios covering 10 target languages, each machine-translated from its English reference by a pipeline that enforces a cross-surface preservation contract.

For the methodology, per-language quality analysis, and the evaluation of frontier agents on this benchmark, see the paper.

Layout

One config per (language, subset) pair, a single test split each, with the same columns as meta-agents-research-environments/gaia2-cli:

column	description
scenario_id	Stable GAIA2 scenario identifier, shared across languages
scenario	Full scenario definition as a JSON string (see below)
language	ISO 639-3 code
script	ISO 15924 code
subset	GAIA2 capability subset

import json
from datasets import load_dataset

ds = load_dataset("facebook/omnilingual-gaia2", "spa_Latn_execution", split="test")
scenario = json.loads(ds[0]["scenario"])

scenario is the standard GAIA2 scenario object, with top-level keys metadata (scenario id, duration, start time, hints, tags), apps (the serialized app universe the agent acts on), events (the oracle event DAG), version, and augmentation.

Scenario IDs are aligned across languages, so any two configs sharing a subset can be compared row for row.

Coverage

language	execution	search	ambiguity	adaptability	total
cmn_Hans	160	160	160	160	640
deu_Latn	160	160	160	160	640
fra_Latn	160	160	160	160	640
hin_Deva	160	160	160	160	640
ind_Latn	160	160	160	160	640
ita_Latn	160	160	160	160	640
jpn_Jpan	160	160	160	160	640
por_Latn	160	160	160	160	640
spa_Latn	160	160	160	160	640
tur_Latn	160	160	160	160	640

Running the benchmark

The gaia2-cli runner now supports this dataset natively.

Set [target].language and it fetches and caches the per-language config for you:

[target]
dataset = "facebook/omnilingual-gaia2"
language = "spa_Latn"   # or cmn_Hans, deu_Latn, fra_Latn, ...
splits = "all"          # execution, search, ambiguity, adaptability

[judge]
prompt_version = "omnilingual-gaia2"

A complete annotated template — OpenClaw with a locally served agent model, a separate judge endpoint and pass@3 — ships with the runner as runner/examples/openclaw_qwen_omnilingual_gaia2_pass3.toml. Copy it, edit the language, models and endpoints, then:

# Validate the config and resolve the scenario selection without launching
gaia2-runner run-config --config ./my_omnilingual_run.toml --dry-run

# Run the evaluation
gaia2-runner run-config --config ./my_omnilingual_run.toml

Note on the verifier: [judge].prompt_version = "omnilingual-gaia2" selects the multilingual judge prompts used for the paper's numbers. Scoring non-English trajectories with the stock English verifier is not a fair comparison. Set it for parity with published results.

Provenance and limitations

The source dataset is meta-agents-research-environments/gaia2-cli. Translations are machine-generated with a translation pipeline involving google/gemma-4-31B-it and have not been fully human-verified. The pipeline is open source at gaia2-cli/mt, so GAIA2 can be translated into further languages.

Citation

@misc{caciolai2026omnilingualgaia2evaluatingmultilingualgap,
      title={OmnilingualGAIA2: Evaluating the Multilingual Gap in Frontier AI Agents},
      author={Andrea Caciolai and Pere-Lluís Huguet Cabot and Chierh Cheng and Albert Ventayol-Boada and Gabriel Mejia Gonzalez and Christophe Ropers and Lucas Bandarkar and Sebastian Ruder and Darlene Sakakihara and Elliot Yun and Pierre Andrews and Grégoire Mialon and Romain Froger and Marta R. Costa-jussà},
      year={2026},
      eprint={2608.08775},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.08775},
}