RIDGE: Revealing Item-level Differences of Generalist agents across digital and physical Environments

Item-level results and agent traces from Astra, Opus 5.5, and other Frontier Models Demonstrate Jagged Performance Across SoTA Agentic Tasks from Web Browsing to Robotics. We ran seven frontier models on web browsing and three on driving, tabletop manipulation, industrial procedures and toy assembly. We recorded every item's outcome and the model's output.

Dataset summary

Across web and physical tasks, a model with a higher average still does worse on some tasks and task categories.

• Web browsing (VisualWebArena): seven models (Fable 5, GPT-5.6, GPT-6 Astra, Kimi K3, Opus 4.7, Opus 5, Opus 5.5), 1,239 rollouts and 20,296 model turns, each with the screenshot the model saw and its reasoning and action.
• Physical domains (Bench2Drive, VLABench, IndEgo, Assembly101): three models (Gemini 3.5-Flash, GPT-6 Astra, Opus 5.5), 5,210 items scored for each model on identical inputs.

Intended uses

• Compare and explore model behaviors task by task and category by category, within and across domains.
• Compare scoring rules: VisualWebArena rows carry the benchmark's verdict and a relaxed re-scoring we tried.

Loading the data

from datasets import load_dataset

results = load_dataset("figai/RIDGE", "visualwebarena", split="train")

# The rollout tables have one file per model (kimik3, opus5, opus47, opus55, gpt56, fable5, astra).
episodes = load_dataset("figai/RIDGE", data_files="traces/visualwebarena/episodes/astra.parquet",
                        split="train")
steps = load_dataset("figai/RIDGE", data_files="traces/visualwebarena/steps/astra.parquet",
                     split="train", streaming=True)  # about 600 MB per model

Item-level results

config	rows	unit	cluster	metric
visualwebarena	177	web task	site	task success (benchmark checker)
bench2drive	3,746	driving segment	86 routes	weighted F1
vlabench	644	manipulation query	6 tasks	exact-match score
indego	300	anticipation point	268 recordings	top-1 accuracy
assembly101	520	anticipation point	27 recordings	top-1 action accuracy

Agent rollouts and model outputs

config	one row is	holds
vwa_episodes	model × task (1,239)	prompt, goal images, evaluator reference, raw and corrected success, invalid_reason, steps, hit_step_limit, final URL and answer, harness log
vwa_steps	model × API turn	screenshot before the turn, reasoning, text, action, API request and response
indego_inputs	item	prompt (<image> marks each image) and images, identical for all models
indego_outputs	model × item	score, prediction, raw output, finish reason, result row (JSON)
assembly101_outputs, bench2drive_outputs, vlabench_outputs	model × item	score, prediction, raw output, finish reason

Encrypted reasoning returned by the APIs is removed from the saved requests and responses.

Limitations

• Single-pass sweep. We ran each model once per item, favoring domain coverage over repeat runs. Repeating runs at the reported budget would measure the variance.
• Harness differences across models. The Opus models use computer-use actions and the rest use function calling, which confounds comparisons of absolute success across models. Kimi K3 uses OSWorld's 4,096-token output budget. Unlike the other models, it sometimes runs out before choosing an action, triggering retries until the task fails.
• Shared setup choices. VWA runs at 100 steps rather than the benchmark's default of 30. All harnesses use the same BrowserGym action subset without tab switching, so the two tasks that open with two tabs (527, 897) fail for every model. The offline evaluations adapt each benchmark's example setup.
• Programmatic verifiers confound competence. String-matching checkers misjudge some answers in both directions, and the BrowserGym checker port scores correct orders on two tasks (695, 699) as failures.
• We measured only where models differ. We did not study why.

invalid_reason flags the four tasks named above (527, 897, 695, 699). The report has further study design and harness details.

Obtaining tasks for Assembly101, Bench2Drive and VLABench

We do not redistribute their images (licenses below), so these three domains have outputs only. To see what a model saw, download the source dataset and look up the item id.

domain	license	source	item id → input
Assembly101	CC BY-NC 4.0	download scripts	<recording>#<k>: the k-th coarse action segment (from 0) of that recording, camera C10095_rgb
Bench2Drive	CC BY-NC-ND 4.0¹	images: rethinklab/Bench2Drive, annotations: Telkwevr/Bench2Drive-VL-base	<route>/<frame>: that frame of the route
VLABench	none stated	VLABench/vlm_evaluation_v1.0	<task>#<k>: <folder>/<task>/example<k>/input/

VLABench folders: M&T/ (select_billiards, select_chemistry_tube), CommenSence/ (select_nth_largest_poker), Complex/ (texas_holdem), Spatial/ (insert_bloom_flower). The 400 insert_bloom_flower_manufactured items are scenes we generated, so their images are not public.

¹ Its GitHub license; its Hugging Face card says Apache-2.0, and we follow the stricter one.

License and terms of use

Our contribution (the results tables, per-item metadata, flags, and the model outputs and traces) is released under CC BY 4.0. The third-party media embedded in the trace tables stay under their upstream licenses:

source	media in this release	upstream license
VisualWebArena	screenshots of its self-hosted sites, goal images as distributed with the benchmark	MIT
IndEgo	egocentric video frames	CC BY 4.0

Citation

@misc{fig_ridge_2026,
  title     = {RIDGE: Revealing Item-level Differences of Generalist agents across digital and physical Environments},
  author    = {Yangyue Wang and Harsh Sikka and Pranav Guruprasad and Sudipta Chowdhury},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/figai/RIDGE}
}