BALROG

Benchmarking Agentic LLM/VLM Reasoning On Games

Davide Paglieri*, Bartłomiej Cupiał*, Samuel Coward, Ulyana Piterbarg,
Maciej Wolczyk, Akbir Khan, Eduardo Pignatelli, Łukasz Kuciński, Lerrel Pinto
Rob Fergus, Jakob Nicolaus Foerster, Jack Parker-Holder, Tim Rocktäschel

Leaderboard

Agent
% Progress
BabyAI
Crafter
TextWorld
BabaIsAI
MiniHack
NetHack
Date

🕹️ ✔️ Qwen2.5-72B-it

16.2 ± 1.6

34.0 ± 6.7

27.3 ± 3.6

11.2 ± 3.8

19.3 ± 3.6

5.0 ± 3.4

0.3 ± 0.3

2024-11-25

🕹️ ✔️ Claude-3.5-Sonnet-2024-10-22

32.6 ± 1.9

68.0 ± 6.6

32.7 ± 3.2

42.1 ± 5.4

37.5 ± 4.4

15.0 ± 5.6

0.6 ± 0.5

2024-11-11

🥇 🕹️ ✔️ Grok-4

43.6 ± 2.2

76.0 ± 6.0

57.3 ± 3.9

62.9 ± 7.9

45.8 ± 4.5

17.5 ± 6.0

1.8 ± 0.8

2025-07-23

🕹️ ✔️ Gemini-1.5-Flash-002

14.6 ± 1.4

50.0 ± 7.1

20.0 ± 0.7

0.0 ± 0.0

12.8 ± 2.3

5.0 ± 3.5

0.0 ± 0.0

2024-11-11

🕹️ ✔️ Llama-3.1-8B-it

15.1 ± 1.6

36.0 ± 6.8

25.5 ± 3.2

6.1 ± 2.4

18.3 ± 3.5

5.0 ± 3.4

0.0 ± 0.0

2024-11-11

🕹️ ✔️ Qwen-2.5-7B-it

7.8 ± 1.1

14.0 ± 4.9

16.4 ± 3.0

3.9 ± 1.0

12.5 ± 3.0

0.0 ± 0.0

0.0 ± 0.0

2024-11-25

🕹️ ✔️ DeepSeek-R1-Distill-Qwen-32B

19.5 ± 1.6

48.0 ± 7.1

15.0 ± 2.1

10.6 ± 3.7

40.0 ± 4.5

2.5 ± 2.5

0.7 ± 0.4

2025-01-26

🕹️ ✔️ Microsoft-Phi-4

11.6 ± 1.4

32.0 ± 6.6

13.6 ± 2.7

2.5 ± 0.9

16.7 ± 3.4

5.0 ± 3.4

0.0 ± 0.0

2025-01-13

🕹️ ✔️ Reka-Flash-3

29.2 ± 1.8

76.0 ± 6.0

33.6 ± 3.5

11.6 ± 4.6

43.3 ± 4.5

10.0 ± 4.7

0.6 ± 0.3

2025-03-13

🕹️ ✔️ Gemini-2.5-Flash

33.5 ± 2.2

68.0 ± 6.6

40.0 ± 4.8

29.6 ± 7.2

50.0 ± 4.6

12.5 ± 5.2

0.8 ± 0.4

2025-07-22

🕹️ ✔️ Qwen2-VL-7B-it

3.7 ± 0.8

4.0 ± 2.8

6.4 ± 1.7

1.6 ± 0.6

7.6 ± 2.4

2.5 ± 2.5

0.0 ± 0.0

2024-11-25

🕹️ ✔️ Grok-3-beta

29.5 ± 2.2

62.0 ± 6.9

25.0 ± 4.1

32.5 ± 6.9

33.3 ± 4.3

22.5 ± 6.6

1.6 ± 0.4

2025-04-25

🕹️ ✔️ Llama-3.2-3B-it

10.1 ± 1.3

20.0 ± 5.7

17.3 ± 2.8

3.5 ± 1.1

17.5 ± 3.5

2.5 ± 2.5

0.0 ± 0.0

2024-11-11

🕹️ ✔️ Llama-3.3-70B-it

23.0 ± 1.7

66.0 ± 6.7

28.6 ± 4.1

9.0 ± 2.9

29.2 ± 4.1

5.0 ± 3.4

0.4 ± 0.3

2024-12-09

🕹️ ✔️ GPT-4o-2024-05-13

32.3 ± 1.5

77.6 ± 3.7

33.1 ± 2.3

39.3 ± 5.2

33.7 ± 3.3

10.0 ± 4.7

0.4 ± 0.4

2024-11-11

🕹️ ✔️ Gemini-1.5-Pro-002

21.0 ± 1.2

58.4 ± 4.4

30.2 ± 2.9

0.0 ± 0.0

32.0 ± 3.3

5.0 ± 3.5

0.4 ± 0.4

2024-11-11

🕹️ ✔️ Qwen2-VL-72B-it

12.8 ± 1.6

24.0 ± 6.0

22.7 ± 2.7

16.5 ± 5.4

10.8 ± 2.8

2.5 ± 2.5

0.0 ± 0.0

2024-11-25

🕹️ ✔️ Mistral-Nemo-it-2407

17.6 ± 1.5

50.0 ± 7.1

27.7 ± 2.7

4.5 ± 1.3

20.8 ± 3.7

2.5 ± 2.5

0.3 ± 0.3

2024-12-09

🕹️ ✔️ Llama-3.1-70B-it

27.9 ± 1.4

73.2 ± 4.0

31.2 ± 2.7

15.0 ± 4.6

40.0 ± 3.4

7.5 ± 4.2

0.3 ± 0.3

2024-11-11

🕹️ ✔️ GPT-4o-mini-2024-07-18

17.4 ± 1.4

50.4 ± 4.5

15.9 ± 2.0

12.2 ± 3.5

15.6 ± 2.5

10.0 ± 4.7

0.0 ± 0.0

2024-11-11

🕹️ ✔️ Llama-3.2-90B-it

27.3 ± 1.4

72.0 ± 6.3

31.7 ± 1.4

11.2 ± 3.0

43.9 ± 3.5

5.0 ± 3.4

0.0 ± 0.0

2024-11-11

🕹️ ✔️ Claude-3.5-Haiku-2024-10-22

19.3 ± 1.8

52.0 ± 7.1

26.4 ± 2.8

18.0 ± 5.8

8.3 ± 2.5

10.0 ± 4.7

1.2 ± 0.4

2024-12-11

🕹️ ✔️ Llama-3.2-11B-it

16.8 ± 1.5

50.0 ± 7.1

26.2 ± 3.3

6.7 ± 2.2

15.6 ± 2.5

2.5 ± 2.5

0.0 ± 0.0

2024-11-11

🥉 🕹️ ✔️ DeepSeek-R1

34.9 ± 2.1

74.0 ± 6.2

36.4 ± 3.8

21.8 ± 6.1

50.8 ± 4.6

25.0 ± 6.8

1.4 ± 0.5

2025-04-10

🥈 🕹️ ✔️ Gemini-2.5-Pro-Exp-03-25

43.3 ± 2.3

80.0 ± 5.7

55.0 ± 6.0

49.2 ± 8.2

56.7 ± 4.5

17.5 ± 6.0

1.7 ± 0.2

2025-04-25

🕹️ ✔️ Llama-3.2-1B-it

6.6 ± 1.0

8.0 ± 3.8

12.7 ± 1.9

3.3 ± 0.9

10.8 ± 2.8

5.0 ± 3.4

0.0 ± 0.0

2024-11-11

Agent
% Progress
BabyAI
Crafter
BabaIsAI
MiniHack
NetHack
Date

🥈 🕹️ ✔️ Claude-3.5-Sonnet-2024-10-22

35.5 ± 2.0

82.0 ± 5.4

37.3 ± 3.1

34.5 ± 4.4

22.5 ± 6.6

1.2 ± 0.4

2024-11-11

🕹️ ✔️ Gemini-1.5-Flash-002

14.9 ± 1.4

43.2 ± 4.4

20.7 ± 4.4

8.3 ± 1.9

2.5 ± 2.5

0.0 ± 0.0

2024-11-11

🕹️ ✔️ Qwen2-VL-7B-it

4.4 ± 0.8

2.0 ± 2.0

5.5 ± 0.9

11.9 ± 3.0

5.0 ± 3.4

0.0 ± 0.0

2024-11-25

🕹️ ✔️ GPT-4o-2024-05-13

22.6 ± 1.4

62.0 ± 4.3

26.8 ± 3.7

18.6 ± 2.7

5.0 ± 3.4

0.4 ± 0.4

2024-11-11

🥉 🕹️ ✔️ Gemini-1.5-Pro-002

25.8 ± 1.4

58.4 ± 4.4

33.5 ± 2.1

31.4 ± 3.2

5.0 ± 3.4

0.5 ± 0.5

2024-11-11

🕹️ ✔️ Qwen2-VL-72B-it

12.2 ± 1.6

34.0 ± 6.7

18.6 ± 2.8

5.9 ± 2.2

2.5 ± 2.5

0.0 ± 0.0

2024-11-25

🕹️ ✔️ Llama-3.2-90B-it

21.0 ± 1.6

66.0 ± 6.7

14.5 ± 1.8

21.9 ± 2.9

2.5 ± 2.5

0.0 ± 0.0

2024-11-11

🕹️ ✔️ Llama-3.2-11B-it

8.4 ± 1.3

18.0 ± 5.4

15.9 ± 1.2

5.8 ± 1.6

2.5 ± 2.5

0.0 ± 0.0

2024-11-11

🥇 🕹️ ✔️ Gemini-2.5-Pro-Exp-03-25

35.7 ± 2.1

74.0 ± 6.2

37.3 ± 4.6

53.3 ± 4.6

12.5 ± 5.2

1.6 ± 0.4

2025-04-25

🕹️ ✔️ GPT-4o-mini-2024-07-18

15.4 ± 1.3

38.0 ± 4.3

19.9 ± 3.1

16.4 ± 2.6

2.5 ± 2.5

0.0 ± 0.0

2024-11-11

- The % Progress metric refers to the average completion percentage of BALROG environments of the model.
- ✔️ Checked indicates that we, the BALROG team, received access to the system and were able to reproduce the patch generations.
- 🕹️ Open refers to submissions that have open-source code. This does not necessarily mean the underlying model is open-source.
- The leaderboard is updated once a week on Monday.
- If you would like to submit your model to the leaderboard, please check the submission page.

About

BALROG is a benchmark designed to evaluate the agentic capabilities of large language and vision-language models (LLMs and VLMs) on long-horizon tasks, testing their ability to plan, reason spatially, and explore in dynamic environments. Our benchmark reveals that while current models show some success on simpler tasks, they struggle with more complex, procedurally generated environments like NetHack, especially when vision-based decision-making is involved. We provide an open, fine-grained evaluation framework to drive progress in autonomous agent research. Read more about agent quest in our paper!

Citation


  @inproceedings{paglieri2025balrog,  
    title={{BALROG}: Benchmarking Agentic {LLM} and {VLM} Reasoning On Games},
    author={Davide Paglieri and Bart{\l}omiej Cupia{\l} and Samuel Coward and Ulyana Piterbarg and Maciej Wolczyk and Akbir Khan and Eduardo Pignatelli and {\L}ukasz Kuci{\'n}ski and Lerrel Pinto and Rob Fergus and Jakob Nicolaus Foerster and Jack Parker-Holder and Tim Rockt{\"a}schel},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=fp6t3F669F}
  }
    

Usage: BALROG's website and leaderboard use the template made available by SWE-bench. If you would like to use this template for your own leaderboard, please visit their website and request permission

Correspondence to: d.paglieri@cs.ucl.ac.uk