NitroGen: An Open Foundation Model for Generalist Gaming Agents
Peer-reviewed conference; CVPR 2026 Best Paper Honorable Mention, 2026 · paper · AI and machine learning
Introduces NitroGen, a vision-action model trained on 40,000 hours of gameplay videos from over 1,000 games, with actions extracted from on-screen input overlays. Fine-tuning on unseen games gives up to 52% relative improvement in task success over training from scratch.
Checked: PDF from openaccess.thecvf.com, fetched 5 Oct 2026. Page numbers are those of that PDF.
0 findings
None that passed a second check.
13 minor (typos and the like; most checked by the judge only)
Reference or label 'Prior to deep learning, game-playing AI relied on search and planning, as in GVGAI, Mario AI, and Ms. Pac-Man [24, 26, 33, 44, 54, 55]' cites [55], 'Deep reinforcement learning for general video game ai' (2018).
- p. 8 Sec. 5 Gaming agents
Prior to deep learning, game-playing AI re- lied on search and planning, as in GVGAI, Mario AI, and Ms. Pac-Man [24, 26, 33, 44, 54, 55]
- p. 11 Reference [55]
Deep reinforcement learning for general video game ai
Only the authors can say what is right here.
Judge's note: Reference [55] on p.11 is titled as a deep-RL paper, so it does not support a claim about pre-deep-learning search and planning. It may be meant only as a pointer to GVGAI. The citation needs an author check.
Second check: another model agreed
Report a problem with this finding
Wording Cradle [51] is the example of the non-API group in the Introduction ('(2) complicated perception modules for textual information extraction and object detection [51]'), but in Sec. 5 it is grouped with the API approaches ('high-level reasoning with structured APIs (e.g., Voyager, Cradle) [51, 60]').
- p. 1 Introduction
plicated perception modules for textual information extrac-tion and object detection [51]
- p. 8 Section 5, Gaming agents
high-level reasoning with struc-tured APIs (e.g., Voyager, Cradle) [51, 60]
- p. 1 Sec. 1 Introduction
plicated perception modules for textual information extrac-
- p. 8 Sec. 5 Gaming agents
tured APIs (e.g., Voyager, Cradle) [51, 60]
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: what Cradle's method actually is. On the page images, the p.1 Introduction contrasts (1) programmatic APIs [59–61] with (2) perception modules [51]. On p.8 the paper cites [51] (Cradle, 'general computer control', p.11 reference list) as an example of structured-API methods. Only the authors can say which description they intend.
Second check: another model agreed
Report a problem with this finding
Wording Page 2 still reads 'We will open-source the NITROGEN dataset, simulator, and pre-trained weights upon acceptance of the paper.', while the abstract says 'We release the dataset, evaluation suite, and model weights'.
- p. 2 Introduction, last paragraph
We will open-source the NITROGEN dataset, simulator, and pre-trained weights upon acceptance of the paper.
- p. 1 Abstract
We release the dataset, evaluation suite, and model weights
Judge's note: Both passages are on the page images (p.2 left column, p.1 abstract). The conditional wording about acceptance is stale in the published camera-ready version and does not match the abstract's present tense.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 5 says 'NITROGEN scales behavior cloning to internet data across hundreds of games', whereas the abstract, introduction and Fig. 3 say 'more than 1,000 games'.
- p. 8 Section 5, Gaming agents
scales behavior cloning to inter-net data across hundreds of games
- p. 1 Abstract
of gameplay videos across more than 1,000 games
- p. 4 Figure 3 (a)
233 452 303 76 15
- p. 8 Sec. 5 Related works, Gaming agents paragraph
net data across hundreds of games
Judge's note: The Fig. 3a bins on the p.4 image sum to 1,079 games. The related-work wording understates the paper's own count.
Second check: another model agreed
Report a problem with this finding
Wording 'It contains 38,739 videos from 818 different content creators, with an average video duration of 1 hour and 50 minutes' gives about 71,000 h (the raw collection). The sentence sits next to the description of 'the dataset', which Fig. 3 gives as 40,000 h after filtering, and the paper does not say whether the statistics are raw or filtered.
- p. 3 Section 2.1, after Figure 3 reference
It contains 38,739 videos from 818 different content creators, with an average video duration of 1 hour and 50 minutes.
- p. 4 Section 2.1, Stage 3
resulting in 55% of the data being kept
Judge's note: 38,739 videos × 110 min ≈ 71,021 h, which matches the 71,000 h of raw video (p.2), not the 40,000 h after filtering (Fig. 3 caption; Stage 3 keeps 55%). The paragraph follows directly from the raw-collection sentence, so the raw reading is plausible. But 'The dataset covers more than 1,000 unique games' also refers to Fig. 3, which shows the filtered data. The ambiguity is minor.
Second check: another model agreed
Report a problem with this finding
Wording The Fig. 3 caption says 'with the remainder distributed across seven genres', but the pie chart's remainder is six named genres (Sports, Metroidvania, Roguelike, RPG, Battle Royale, Racing) plus an aggregate 'Other' (9.4%).
- p. 4 Figure 3 caption
with the remainder distributed across seven genres
- p. 4 Figure 3(b) pie chart
Other 9.4%
Judge's note: Checked the Fig. 3b pie on the p.4 image. There are seven slices besides the top three, one of which is 'Other'. The slices sum to 100%. Calling the catch-all a genre is loose wording.
Second check: another model agreed
Report a problem with this finding
Wording Limitations state 'It can thus only play for a few seconds, when the visual context is clear enough to predict a next action'. Elsewhere, task attempts are 'typically lasting a few minutes' (Sec. 2.2), and a boss fight that 'takes ∼2 minutes for a human' reaches 90.5% success.
- p. 7 Section 4, Scope and design limitations
It can thus only play for a few seconds, when the visual context is clear enough
- p. 5 Section 2.2, Diverse evaluation tasks
with attempts typically lasting a few minutes
- p. 7 Section 3, self-imitation
A successful attempt takes ∼2 minutes for a human.
- p. 7 Sec. 4 Scope and design limitations
It can thus only play for a few seconds
Only the authors can say what is right here.
Judge's note: Checked p.5 (Sec. 2.2; Fig. 4 caption: 'a few seconds to a few minutes') and p.7 (self-imitation paragraph; Sec. 4). As written, the limitation conflicts with the durations of the tasks the paper reports the model completing. The intended meaning is probably about long-horizon play, so this is a wording issue.
Second check: another model agreed
Report a problem with this finding
Wording 'Across different data quantities, fine-tuning achieves an average relative improvement of 10% on the isometric roguelike, whereas the 3D action-RPG shows a 25% average relative improvement.' The 3D action-RPG is shown only at 30h, split by task type (Fig. 7b), so its 25% is not an average across data quantities.
- p. 7 Section 3, transfer paragraph
Across different data quantities, fine-tuning achieves an average relative improvement of 10% on the isometric roguelike, whereas the 3D action-RPG shows a 25% average relative improvement.
- p. 8 Figure 7(b) subtitle
varying task types (3D action-RPG, 30h data)
- p. 7 Section 3, 'The effectiveness of pre-training varies' paragraph
the 3D action-RPG shows a 25% average relative improvement
- p. 8 Figure 7b
73.3% 48.3% 60.0% 48.0% 66.6% 63.3%
Only the authors can say what is right here.
Judge's note: Fig. 7a varies data quantity (60/120/240h) for the roguelike only. Fig. 7b is the 3D action-RPG at 30h across three task types. The roguelike 10% matches the mean of the per-quantity gains. The 3D 25% matches the ratio of mean completion rates over task types (66.63/53.2 ≈ 1.25); the mean of the per-task gains is about 27%. The sentence's 'across different data quantities' framing therefore does not fit the 3D figure. Only the authors can say how the 25% was computed.
Second check: another model agreed
Report a problem with this finding
Typo Grammatical number mismatch: 'Figure 2b shows examples of successful match.' (should be 'successful matches').
- p. 3 Stage 1
Figure 2b shows examples of successful match.
Judge's note: Seen on the p.3 crop. Cosmetic grammar slip.
Second check: none (judge only)
Report a problem with this finding
Typo Comma splice: 'Frequent pausing and resuming during gameplay could potentially affect the game's physics engine in unknown ways, we verify that this process does not alter games' physics and behaviors'.
- p. 4 Section 2.2
Frequent pausing and resuming during game-
Judge's note: Seen on the p.4 image (Sec. 2.2): two independent clauses are joined by a comma.
Second check: none (judge only)
Report a problem with this finding
Typo The run-in heading 'NITROGEN is a strong candidate for RL fine-tuning' has no closing period and sits on its own line, while the other Sec. 3 headings end with a period and run into the text.
- p. 7 Section 3
NITROGEN is a strong candidate for RL fine-tuning We conduct
Judge's note: On the p.7 image, the heading stands alone without a period. The other bold headings in Sec. 3 (e.g. 'NITROGEN pre-training improves downstream fine-tuning on unseen environments.') end with periods. This is a formatting inconsistency only.
Second check: none (judge only)
Report a problem with this finding
Typo Subject–verb agreement error: 'Other approaches uses LLMs for high-level reasoning' (should be 'use').
- p. 8 Section 5, Gaming agents paragraph
Other approaches uses LLMs for high-level reasoning
Judge's note: Seen on the p.8 image, Sec. 5.
Second check: none (judge only)
Report a problem with this finding
Typo The π0 paper is listed twice: [7] 'π0: A vision-language-action flow model for general robot control, 2024. 6, 1' and [8] 'π0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 8'.
- p. 9 References [7]
π0: A vision-language-action flow model for general robot control, 2024. 6, 1
- p. 9 References [8]
π0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 8
Judge's note: On the p.9 image, [7] and [8] have the same first ten authors, the same title and the same year. [8] truncates the author list with 'et al.' This is a duplicated bibliography entry.
Second check: none (judge only)