Mastering PokeGym: Graph-Guided Multimodal Evolution at Test Time

Ruizhi Zhang; Ye Huang; Yuangang Pan; Chuanfu Shen; Zhilin Liu; Ting Xie; Haijun Lei; Lixin Duan

cs.CVApr 9, 2026

Mastering PokeGym: Graph-Guided Multimodal Evolution at Test Time

Authors: Ruizhi Zhang, Ye Huang, Yuangang Pan, Chuanfu Shen, Zhilin Liu, Ting Xie, Haijun Lei, Lixin Duan

Organizations: Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China · A*STAR · College of Computer Science and Software Engineering, Shenzhen University

Abstract

While artificial intelligence has mastered structured games like chess and Go, vision-language agents still struggle in visually-driven 3D games without access to game states. Existing game environments typically evaluate a fixed agent configuration, rather than an agent's ability to improve its configuration across consecutive episodes of the same task---a paradigm known as test-time learning (TTL). Furthermore, current TTL methods typically optimize single modalities---such as text prompts or actions---in isolation, ignoring the synergy between perception, reasoning, and control. To bridge these gaps, we first introduce PokeGym, a long-horizon benchmark built upon the 3D open-world game Pokémon Legends: Z-A, where agents act from visual observations without access to game states, designed to evaluate an agent's ability to learn and adapt across consecutive episodes of the task. To tackle this challenging environment, we propose Graph-Guided Evolutionary Multimodal Agent Configuration (G-EvoMAC), a graph-guided framework that jointly optimizes visual perception, strategy, and action set synergistically. Extensive experiments show that G-EvoMAC achieves a 60.18% average success rate on PokeGym, outperforming the strongest baseline by over 11 percentage points, validating the power of cross-modal co-evolution.

Figures & tables

Figure 1: Overview of PokeGym and G-EvoMAC. Top: PokeGym is a 3D vision-only benchmark for evaluating test-time learning through iterative multimodal evolution. Bottom: G-EvoMAC jointly evolves visual, language, and action modalities and models their cross-modal synergy via a GNN.

Environment	Method	Observation	Open	Narrative-	Urban	TTL
			World	driven	Env
NetHack	IMPALA	Text	$\times$	$\times$	$\times$	$\times$
TextWorld	BYU	Text	$\times$	$\times$	$\times$	$\times$
Pokémon Red	DRL	2D + Text	$\times$	✓	$\times$	$\times$
StarDojo	Zero-shot	2D + Text	✓	$\times$	$\times$	$\times$
MineDojo	MineAgent	3D + Text	✓	$\times$	$\times$	$\times$

Table 1: Comparison of game-based environments. Open World indicates unconstrained exploration; Narrative-driven indicates multi-stage narrative quests; Urban Environment indicates densely populated urban games; TTL marks whether the benchmark is designed to evaluate agents’ ability to refine their internal configuration across consecutive episodes of the same task.

Figure 2: Overview of PokeGym. Top: Task design, showing the proportion of the three task categories along with example visual trajectories. Bottom: The environment interface, where the agent perceives the game through visual observations and interacts with the emulator via high-level actions.

Figure 3: Overview of the G-EvoMAC framework. (a) The multimodal configuration evolver generates candidate configurations from historical rollouts; (b) the cross-modal synergy performance predictor estimates each candidate’s score and disagreement; (c) the disagreement-aware selector picks the most informative configuration to evaluate, and the observed score is used to update the predictor for the next round.

Configuration	Navigation		Interaction		Localization		Average
Configuration	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
w/o Visual	34.44 $\pm$ 5.94	6.86 $\pm$ 1.19	42.67 $\pm$ 5.23	7.13 $\pm$ 1.98	41.82 $\pm$ 5.32	5.30 $\pm$ 1.01	39.82 $\pm$ 3.18	6.52 $\pm$ 0.91
w/o Strategy	39.44 $\pm$ 6.25	8.55 $\pm$ 1.32	48.00 $\pm$ 4.73	10.19 $\pm$ 2.19	53.94 $\pm$ 5.47	7.31 $\pm$ 1.46	47.02 $\pm$ 3.16	8.84 $\pm$ 1.05
w/o Action	42.78 $\pm$ 5.75	8.26 $\pm$ 1.79	55.11 $\pm$ 4.76	7.63 $\pm$ 1.65	52.12 $\pm$ 5.83	7.57 $\pm$ 1.39	50.35 $\pm$ 3.12	7.81 $\pm$ 0.94
G-EvoMAC	46.11 $\pm$ 6.11	5.72 $\pm$ 1.01	68.00 $\pm$ 3.83	5.76 $\pm$ 0.89	64.85 $\pm$ 4.27	5.07 $\pm$ 0.96	60.18 $\pm$ 2.87	5.55 $\pm$ 0.54

Table 2: Ablation study on the three evolved modalities. Each row excludes one modality from the full configuration; complete results are provided in the appendix.

Model	Venues	Navigation		Interaction		Localization		Average
Model	Venues	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
Zero-shot VLMs
Qwen3.6-35B	–	18.89 $\pm$ 4.43	42.25 $\pm$ 3.45	14.67 $\pm$ 3.20	23.38 $\pm$ 2.50	20.00 $\pm$ 3.59	29.94 $\pm$ 3.62	17.54 $\pm$ 2.15	31.24 $\pm$ 1.94
Claude-Sonnet-4.6	–	22.78 $\pm$ 5.41	20.97 $\pm$ 2.90	16.00 $\pm$ 3.28	15.54 $\pm$ 1.83	17.58 $\pm$ 5.13	15.76 $\pm$ 2.42	18.60 $\pm$ 2.60	17.32 $\pm$ 1.37
GPT-5.4	–	31.11 $\pm$ 5.88	17.93 $\pm$ 2.59	38.22 $\pm$ 4.66	8.74 $\pm$ 1.34	42.42 $\pm$ 4.06	15.05 $\pm$ 2.87	37.19 $\pm$ 2.87	13.47 $\pm$ 1.32
Game Agent Frameworks

Table 3: Main results on PokeGym, measured by Success Rate (SR) and Ineffective Move Rate (IMR). Except for the zero-shot VLMs, all other agents use GPT-5.4 as the backbone. Values are reported as mean with standard error shown in smaller font, and bold values indicate the best performance.

Figure 4: Cross-modal attention heatmap of the GNN predictor. Each cell shows the normalized attention weight. Higher values indicate stronger cross-modal dependency.

Predictor	Navigation		Interaction		Localization		Average
Predictor	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
LLM-as-Judge	30.00 $\pm$ 5.66	7.92 $\pm$ 1.53	36.44 $\pm$ 4.68	8.42 $\pm$ 1.70	38.18 $\pm$ 5.67	9.83 $\pm$ 1.87	34.91 $\pm$ 3.04	8.67 $\pm$ 0.98
MLP	38.33 $\pm$ 6.29	9.07 $\pm$ 1.54	51.56 $\pm$ 4.94	6.09 $\pm$ 1.10	56.36 $\pm$ 4.48	11.69 $\pm$ 2.48	48.77 $\pm$ 3.12	8.65 $\pm$ 0.99
GNN	46.11 $\pm$ 6.11	5.72 $\pm$ 1.01	68.00 $\pm$ 3.83	5.76 $\pm$ 0.89	64.85 $\pm$ 4.27	5.07 $\pm$ 0.96	60.18 $\pm$ 2.87	5.55 $\pm$ 0.54

Table 4: Comparison of different predictors.

Method	Navigation		Interaction		Localization		Average
Method	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
w/o Predictor	41.11 $\pm$ 5.96	9.66 $\pm$ 1.32	51.56 $\pm$ 4.52	7.75 $\pm$ 1.80	56.36 $\pm$ 4.48	11.92 $\pm$ 2.49	49.65 $\pm$ 2.93	9.56 $\pm$ 1.10
GNN Predictor	46.11 $\pm$ 6.11	5.72 $\pm$ 1.01	68.00 $\pm$ 3.83	5.76 $\pm$ 0.89	64.85 $\pm$ 4.27	5.07 $\pm$ 0.96	60.18 $\pm$ 2.87	5.55 $\pm$ 0.54

Table 5: Comparison with and without the GNN predictor.

Figure 5: Calibration of the GNN predictor. Left: predicted uncertainty ( $\sigma$ ) versus mean absolute error. Right: mean predicted score versus mean true score.

Selector	Navigation		Interaction		Localization		Average
Selector	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
Explore-only	29.44 $\pm$ 6.25	8.57 $\pm$ 1.30	40.44 $\pm$ 4.82	8.94 $\pm$ 1.81	35.15 $\pm$ 5.58	8.50 $\pm$ 1.83	35.44 $\pm$ 3.18	8.70 $\pm$ 0.97
Exploit-only	38.89 $\pm$ 5.85	5.95 $\pm$ 1.11	62.22 $\pm$ 3.93	6.59 $\pm$ 1.00	52.12 $\pm$ 5.36	6.50 $\pm$ 1.21	51.93 $\pm$ 2.99	6.36 $\pm$ 0.63
Thompson Sampling	35.00 $\pm$ 6.19	6.05 $\pm$ 1.05	45.33 $\pm$ 5.10	7.69 $\pm$ 1.99	41.21 $\pm$ 5.57	7.08 $\pm$ 1.36	40.88 $\pm$ 3.23	7.00 $\pm$ 0.93
Deep Ensembles	46.11 $\pm$ 6.27	9.54 $\pm$ 1.53	60.44 $\pm$ 4.14	6.89 $\pm$ 1.18	56.36 $\pm$ 4.22	10.88 $\pm$ 2.49	54.74 $\pm$ 2.87	8.88 $\pm$ 0.99
LinUCB	41.11 $\pm$ 6.22	7.23 $\pm$ 1.14	58.22 $\pm$ 4.75	7.12 $\pm$ 1.29	52.12 $\pm$ 3.89	10.75 $\pm$ 2.41	51.05 $\pm$ 2.99	8.21 $\pm$ 0.94

Table 6: Ablation on the configuration selection strategy.

Model	Navigation		Interaction		Localization		Average
Model	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
Zero-shot	18.89 $\pm$ 4.43	42.25 $\pm$ 3.45	14.67 $\pm$ 3.20	23.38 $\pm$ 2.50	20.00 $\pm$ 3.59	29.94 $\pm$ 3.62	17.54 $\pm$ 2.15	31.24 $\pm$ 1.94
PromptBreeder	22.78 $\pm$ 4.45	33.41 $\pm$ 2.94	17.78 $\pm$ 3.19	17.55 $\pm$ 2.35	29.09 $\pm$ 4.61	24.55 $\pm$ 3.32	22.63 $\pm$ 2.33	24.59 $\pm$ 1.73
Voyager	26.67 $\pm$ 4.44	18.08 $\pm$ 2.25	21.78 $\pm$ 3.75	15.29 $\pm$ 1.61	28.48 $\pm$ 4.27	19.73 $\pm$ 3.44	25.26 $\pm$ 2.38	17.46 $\pm$ 1.38
G-EvoMAC	31.11 $\pm$ 4.88	17.96 $\pm$ 2.17	32.44 $\pm$ 4.13	13.70 $\pm$ 1.47	36.36 $\pm$ 4.22	15.64 $\pm$ 2.35	33.16 $\pm$ 2.54	15.61 $\pm$ 1.13

Table 7: Performance comparison using Qwen3.6-35B.

Model	State		Inspection		Spatial		Average
Model	GCS $\uparrow$	PW $\uparrow$	GCS $\uparrow$	PW $\uparrow$	GCS $\uparrow$	PW $\uparrow$	GCS $\uparrow$	PW $\uparrow$
Zero-shot	12.22	10.34	10.00	6.55	0.00	0.00	8.05 $\pm$ 2.02	6.48 $\pm$ 1.72
Reflexion	12.22	6.79	5.00	3.64	0.00	0.00	7.18 $\pm$ 1.88	4.14 $\pm$ 1.29
EvoTest	11.11	9.24	20.00	8.74	0.00	0.00	9.20 $\pm$ 2.21	6.28 $\pm$ 1.67
Voyager	11.11	7.94	30.00	25.00	0.00	0.00	10.92 $\pm$ 2.41	8.42 $\pm$ 2.04
G-EvoMAC	12.22	12.22	45.00	43.91	5.56	4.93	15.80 $\pm$ 3.26	15.42 $\pm$ 3.15

Table 8: Results on ALFRED using the Qwen3.6-35B. We report goal-condition success (GCS) and path-weighted goal-condition success (PW-GCS, abbreviated as PW).

Figure 6: Trajectory comparison between G-EvoMAC and the other baselines.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Platform	Emulator Version	Game Version	Hit Rate
Windows 10	v1.3.2	v1.0.0	100%
Windows 11	v1.3.2	v1.0.0	100%
Windows 11	v1.3.3	v1.0.0	100%
Windows 11	v1.3.2	v2.0.2	100%
Ubuntu 22.04	v1.3.2	v1.0.0	100%

Appendix

Table 9: Robustness of the automatic evaluator when supplied with signatures discovered by our AOB-scanning scripts. Hit Rate denotes successful first-try detection across different operating systems, emulator versions, and game versions.

Figure 7: Validation of the GNN predictor. Left: predicted versus true configuration scores. Right: Top-1 selection and pairwise ranking accuracy of the score-based acquisition function.

Method	Navigation		Interaction		Localization		Average
Method	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
V-GPS	42.78 $\pm$ 5.92	12.99 $\pm$ 1.92	54.67 $\pm$ 4.98	7.32 $\pm$ 0.92	49.09 $\pm$ 4.09	13.65 $\pm$ 2.47	49.30 $\pm$ 2.97	10.94 $\pm$ 1.03
G-EvoMAC (Ours)	46.11 $\pm$ 6.11	5.72 $\pm$ 1.01	68.00 $\pm$ 3.83	5.76 $\pm$ 0.89	64.85 $\pm$ 4.27	5.07 $\pm$ 0.96	60.18 $\pm$ 2.87	5.55 $\pm$ 0.54

Appendix

Table 10: Comparison with test-time parameter adaptation. V-GPS updates a value function’s parameters at test time and selects candidate actions by learned value.

Modality			Navigation		Interaction		Localization		Average
Visual	Strategy	Action	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
	✓		42.22 $\pm$ 6.00	7.02 $\pm$ 1.54	41.33 $\pm$ 4.56	6.95 $\pm$ 1.52	49.09 $\pm$ 6.10	7.33 $\pm$ 1.49	43.86 $\pm$ 3.14	7.09 $\pm$ 0.88
✓			36.67 $\pm$ 5.99	19.80 $\pm$ 2.39	42.67 $\pm$ 4.20	11.83 $\pm$ 1.72	43.64 $\pm$ 4.80	16.07 $\pm$ 2.89	41.05 $\pm$ 2.86	15.58 $\pm$ 1.34
		✓	37.22 $\pm$ 6.02	8.00 $\pm$ 1.41	37.33 $\pm$ 4.52	9.66 $\pm$ 1.96	38.79 $\pm$ 5.43	8.01 $\pm$ 1.81	37.72 $\pm$ 3.02	8.66 $\pm$ 1.03
	✓	✓	34.44 $\pm$ 5.94	6.86 $\pm$ 1.19	42.67 $\pm$ 5.23	7.13 $\pm$ 1.98	41.82 $\pm$ 5.32	5.30 $\pm$ 1.01	39.82 $\pm$ 3.18	6.52 $\pm$ 0.91
✓	✓		42.78 $\pm$ 5.75	8.26 $\pm$ 1.79	55.11 $\pm$ 4.76	7.63 $\pm$ 1.65	52.12 $\pm$ 5.83	7.57 $\pm$ 1.39	50.35 $\pm$ 3.12	7.81 $\pm$ 0.94

Appendix

Table 11: Complete modality ablation study. The first three columns indicate which modalities (Visual, Strategy, Action) are included in each configuration.

$\boldsymbol{\lambda}$	Navigation		Interaction		Localization		Average
$\boldsymbol{\lambda}$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
0	38.89 $\pm$ 5.85	5.95 $\pm$ 1.11	62.22 $\pm$ 3.93	6.59 $\pm$ 1.00	52.12 $\pm$ 5.36	6.50 $\pm$ 1.21	51.93 $\pm$ 2.99	6.36 $\pm$ 0.63
1	46.11 $\pm$ 6.11	5.72 $\pm$ 1.01	68.00 $\pm$ 3.83	5.76 $\pm$ 0.89	64.85 $\pm$ 4.27	5.07 $\pm$ 0.96	60.18 $\pm$ 2.87	5.55 $\pm$ 0.54
2	46.11 $\pm$ 6.01	7.46 $\pm$ 1.12	56.00 $\pm$ 4.52	7.05 $\pm$ 1.23	54.55 $\pm$ 4.71	8.86 $\pm$ 1.85	52.46 $\pm$ 2.94	7.71 $\pm$ 0.80

Appendix

Table 12: Sensitivity of the disagreement-aware selector to the exploration-exploitation trade-off weight

\lambda

, measured by Success Rate (SR) and Ineffective Move Rate (IMR). Values are reported as mean with standard error shown in smaller font.

\lambda=0

corresponds to pure exploitation,

\lambda=2

to a more exploratory setting, and

\lambda=1

to the default G-EvoMAC.

$\boldsymbol{T}$	Navigation		Interaction		Localization		Average
$\boldsymbol{T}$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
5	41.11 $\pm$ 5.91	9.55 $\pm$ 1.74	66.67 $\pm$ 3.92	6.36 $\pm$ 0.88	56.36 $\pm$ 5.60	8.62 $\pm$ 1.93	55.61 $\pm$ 3.06	8.02 $\pm$ 0.86
10	46.11 $\pm$ 6.11	5.72 $\pm$ 1.01	68.00 $\pm$ 3.83	5.76 $\pm$ 0.89	64.85 $\pm$ 4.27	5.07 $\pm$ 0.96	60.18 $\pm$ 2.87	5.55 $\pm$ 0.54

Appendix

Table 13: Sensitivity of the disagreement-aware selector to the number of MC-Dropout samples

T

, measured by Success Rate (SR) and Ineffective Move Rate (IMR).

T=10

is the default G-EvoMAC setting.

Score	Navigation		Interaction		Localization		Average
Score	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$	SR $\uparrow$	IMR $\downarrow$
w/o $r$	46.67 $\pm$ 5.69	5.24 $\pm$ 0.88	57.33 $\pm$ 3.95	9.81 $\pm$ 1.81	55.76 $\pm$ 4.98	10.64 $\pm$ 2.38	53.51 $\pm$ 2.79	8.60 $\pm$ 1.05
Default	46.11 $\pm$ 6.11	5.72 $\pm$ 1.01	68.00 $\pm$ 3.83	5.76 $\pm$ 0.89	64.85 $\pm$ 4.27	5.07 $\pm$ 0.96	60.18 $\pm$ 2.87	5.55 $\pm$ 0.54

Appendix

Table 14: Sensitivity of the GNN predictor to the score definition used for training. The default score combines task success

v

, recovery rate

r

, and ineffective move rate

I

0.5v+0.5[0.5r+0.5(1-I)]

. The “w/o

r

” variant removes the recovery term, using

0.5v+0.5(1-I)

Figure 8: Average score across evolution rounds for G-EvoMAC with two backbone VLMs on the three task types.

Figure 9: Representative failure cases in PokeGym: (left) visual-spatial understanding failures, (middle) visual-perception and action-control failures, and (right) planning and reasoning failures.

Navigation

Interaction

Localization

Overall

Model

Out

Total

Out

Total

Out

Total

Out

Total

Cost

Zero-shot VLMs

Qwen3.6-35B

51k

24k

75k

42k

18k

60k

53k

24k

77k

48k

22k

70k

–

Claude-Sonnet-4.6

148k

20k

167k

125k

16k

141k

156k

21k

178k

141k

19k

160k

$0.708

GPT-5.4

64k

12k

76k

49k

58k

66k

12k

78k

59k

11k

69k

$0.307

Game Agent Frameworks

Appendix

Table 15: Token consumption and API cost per run. Input, output, and total token counts are reported in thousands (per episode). Costs are shown for proprietary closed-source models and omitted for the open-source model.

Figure 10: Examples of visual enhancements evolved by G-EvoMAC (part 1). Each figure shows original frames in the top row and their enhanced counterparts in the bottom row after per-frame operations.

Figure 11: Examples of visual enhancements evolved by G-EvoMAC (part 2).

Panel of Figure 12 — Figure 12: Example of an evolved language strategy for NPC interaction tasks.

Panel of Figure 13 — Figure 13: Example of an evolved language strategy for navigation tasks with vertical traversal.

Panel of Figure 14 — Figure 14: Example of an evolved language strategy for constrained corridor navigation.

Panel of Figure 15 — Figure 15: Example of evolved macro actions for NPC interaction and collision recovery.

Panel of Figure 16 — Figure 16: Example of evolved macro actions for vertical traversal and corridor escape.

Panel of Figure 17 — Figure 17: Prompt of Planning

Panel of Figure 18 — Figure 18: Prompt of Trajectory Summarization

Panel of Figure 19 — Figure 19: Prompt of Cross-Trajectory Evolution

Explore similar work

May 28, 2026

PTCG-Bench: Can LLM Agents Master Pokémon Trading Card Game?

Given a strategically complex board game, human players can quickly learn to devise strategies after playing a few rounds. Autonomous agents require similar capabilities in realistic interactive environments, yet existing agent benchmarks often fail to fully capture such strategic and evolving decision-making scenarios. We present PTCG-Bench, a benchmark built on the Pok'{e}mon Trading Card Game (PTCG) that evaluates LLM agents at two complementary levels: (1) their decision-making performance within a single complex environment, and (2) their ability to self-evolving through accumulated experience. We further include a modular harness ablation to better interpret agent performance without conflating it with model capability. Our experiments show that, although LLM agents can achieve non-trivial gameplay performance, sustained and stable self-evolution remains challenging, and performance is sensitive to harness design. We hope that PTCG-Bench will facilitate future research on harness-aware and self-evolving agents in realistic interactive environments.

Dongdong Hua, Yifei Sun, Renhong Huang +3

Zhejiang University · FinVolution Group

Game-Playing Agents LLM Agent Evaluation AI Agent Benchmarks Self-Evolving Agents

Ask PDF

Feb 11, 2026

GameDevBench: Evaluating Agentic Capabilities Through Game Development

Despite rapid progress on coding agents, progress on their multimodal counterparts has lagged behind. A key challenge is the scarcity of evaluation testbeds that combine the complexity of software development with the need for deep multimodal understanding. In game development, agents must navigate large, dense codebases while manipulating intrinsically multimodal assets such as shaders, sprites, and animations within a visual game scene. We present GameDevBench, the first benchmark for evaluating agents on game development tasks. GameDevBench consists of 333 tasks derived from web and video tutorials. Tasks require significant multimodal understanding and are complex: the average solution requires over three times the lines of code and file changes compared to prior software development benchmarks. Agents struggle with game development, with the best agent and method solving only 53.8% of tasks. We find a strong correlation between perceived task difficulty and multimodal complexity, with average success rate dropping from 51.4% on gameplay-oriented tasks to 33.0% on 2D graphics tasks. To improve multimodal capability, we introduce two simple image- and video-based feedback mechanisms for agents. Despite their simplicity, these methods consistently improve performance, increasing GPT-5.4's performance from 41.1% to 52.0% when given visual feedback.

Wayne Chi, Yixiong Fang, Arnav Yayavaram +8

Carnegie Mellon University, Pittsburgh, PA, USA · Princeton University, Princeton, NJ, USA

Software Engineering Agents Agent Evaluation AI Agent Benchmarks

Ask PDF

Jul 2, 2026

EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments

Autonomous agents are increasingly expected to improve executable policies through feedback, yet existing evaluations often collapse this process into a final score or confound it with open-ended software-engineering progress. We introduce Autonomous Policy Evolution, a controlled evaluation setting in which a harness-model agent repeatedly edits an executable policy system under a fixed interaction budget. We instantiate this setting in EvoPolicyGym, a benchmark built from compact interactive RL environments that evaluates how agents iteratively improve explored policies. On the EvoPolicyGym suite, GPT-5.5 achieves the strongest aggregate rank score and top-two performance on all 16 environments. Beyond leaderboard results, EvoPolicyGym also provides trajectory-level diagnostics that distinguish how agents allocate budget, convert feedback into parametric tuning. These analyses show that strong autonomous policy evolution depends not only on isolated task wins, but on discovering task-appropriate mechanisms and refining policies under bounded feedback.

Zhilin Wang, Han Song, Runzhe Zhan +13

University of Science and Technology of China · The Chinese University of Hong Kong · University of Macau +5

RL Benchmarks AI Agent Benchmarks Self-Evolving Agents

Ask PDF