Source-linked AI summary
OmniGameArena: A Unified UE5 Benchmark for VLM Game Agents with Improvement Dynamics
Mingxian Lin, Shengju Qian, Yuqi Liu, Yi-Hua Huang, Yiyu Wang, Wei Huang, Yitang Li, Fan Zhang, Zeyu Hu, Lingting Zhu, Xin Wang, Xiaojuan Qi
TL;DR
Existing game-agent benchmarks often provide only first-attempt scores, emphasize Solo play, and lack matched evaluation across heterogeneous agents. This paper introduces OmniGameArena and the Improvement Dynamics Curve, then finds that reflection improves top agents but that peak performance and held-out transfer can diverge. IDC experiments are limited to two environments, three variants per environment, and four agents.
Problem
Game benchmarks leave repeated-reflection adaptation, adversarial and cooperative play, and heterogeneous-agent comparisons insufficiently measured.
Method
OmniGameArena combines twelve newly built UE5 games with unified interfaces and an IDC harness that refines bounded skill prompts across reflection rounds and tests held-out variants.
Results
All four agents evaluated with IDC improve over cold-start baselines, but peak performance typically occurs before the final round and origin-task gains can diverge from held-out transfer.
Takeaways & Limitations
IDC makes score trajectories and held-out skill behavior observable beyond single-round leaderboard scores.
Takeaways & Limitations
IDC covers only two environments, three held-out variants per environment, and four agents, while asymmetric player-reflector setups remain untested.
Abstract
from arXiv · showhide
Vision-language model (VLM) agents are increasingly deployed in interactive game environments. Yet game benchmarks for VLM agents typically report a single first-attempt score per (agent, game) pair, focus on single-agent Solo play, and lack unified protocols for evaluating heterogeneous agent classes (commercial VLMs, open-weight VLMs, and specialized game policies) on the same footing. We address these gaps with OmniGameArena, a real-time benchmark of twelve newly built Unreal Engine 5 games spanning Solo (7), PvP (3), and Coop (2) with unified action interfaces, and the Improvement Dynamics Curve (IDC), an agentic-reflection harness in which a tool-using reflector LLM autonomously refines a bounded skill prompt across multiple rounds. Beyond cold-start leaderboard scores, IDC exposes two additional observables for each (agent, game) pair: how the score evolves across reflection rounds, and how the learned skill behaves on held-out task variants. We report these observables for twelve VLM agents on the cold-start leaderboard and four top agents under IDC.
1 Introduction
OmniGameArena addresses gaps in game-agent evaluation by combining a unified twelve-game UE5 benchmark with the Improvement Dynamics Curve (IDC). The benchmark measures both cold-start performance and how agents improve and transfer skills across reflection rounds.
- Motivation: Existing benchmarks largely report one first-attempt score and underrepresent PvP and Coop settings, leaving repeated-reflection adaptation insufficiently measured.These settings probe capabilities including opponent modeling, role assignment, and recovery from teammate mistakes.
- Benchmark and harness: OmniGameArena provides twelve newly built UE5 games spanning Solo, PvP, and Coop with unified interfaces for heterogeneous agent classes.Commercial VLMs, open-weight VLMs, and specialized policies operate under matched environment conditions through documented adapters.
- Benchmark and harness: IDC runs multiple play-and-reflection rounds, with a tool-using reflector autonomously refining a bounded skill prompt and reporting per-round scores plus held-out transfer.The harness records an Improvement Dynamics Curve for each agent-game instance and evaluates learned skills on held-out task variants.
- Findings: Across twelve cold-start agents, no single VLM dominates, while commercial agents maintain a wide gap over open-weight VLMs and specialized policies.This result establishes leaderboard heterogeneity rather than one universal ordering of agents.
- Findings: Origin-task improvement and held-out transfer can diverge, exposing information that single-round leaderboard scores conceal.The paper identifies this divergence as a central observable of IDC.
2 Related Work
Prior work expanded game-agent evaluation from text and 2D environments to multimodal 3D worlds, but unified multi-regime evaluation and improvement dynamics remain insufficiently covered. OmniGameArena combines capability-diverse games with held-out skill-transfer evaluation to expose differences hidden by single-number metrics.
- Game-agent benchmarks: Earlier benchmarks covered text-only reasoning, 2D-grid spatial tasks, and later 3D open-world interaction with long-horizon planning from pixels.Examples include BALROG, LVLM-Playground, V-MAGE, Cradle, and VideoGameBench.
- Game-agent benchmarks: Existing related suites do not target multi-regime, real-time game interaction across Solo, PvP, and Coop settings.The cited discussion frames this as a limitation motivating the proposed benchmark.
- Game-agent benchmarks: Game-TARS, NitroGen, and Lumine demonstrate increasingly capable visual game agents but outpace standardized cross-game evaluation infrastructure.These systems operate across diverse 3D worlds or long real-time missions, motivating a common evaluation setting.
- Reflection and self-improvement: Reflection methods improve language-model behavior without weight updates through verbal critiques, iterative rewriting, extracted insights, or reusable skills.Examples include Reflexion, Self-Refine, ExpeL, and Voyager.
- Reflection and self-improvement: OmniGameArena evaluates resulting skills on held-out task variants, revealing skill-style differences that single-number metrics hide.This extends reflection evaluation beyond the original task score.
3 OmniGameArena
OmniGameArena is a twelve-game UE5 suite designed to isolate embodied capabilities across Solo, PvP, and cooperative regimes with normalized continuous progress metrics. Its construction uses new game scenarios, contamination audits, and standardized protocols to support cross-task evaluation.
- 3 OmniGameArena: OmniGameArena contains twelve custom UE5 games spanning Solo, PvP, and Coop regimes.The suite is designed to evaluate distinct capability axes of vision-based game agents.
- 3 OmniGameArena: Game progress is normalized to the continuous range [0, 1] across visually rich and physically complex environments.This provides a consistent metric across diverse tasks.
- Contamination avoidance: The benchmark uses a web-exposure audit and entirely new UE5 game designs to reduce memorization of names, rules, and scoring events.Although visual assets may include marketplace components, level geometry, script order, and success criteria are uniquely designed.
- Game suite: The suite includes Solo tasks for navigation, survival, shooting, puzzles, and logistics; PvP tasks for combat and competitive logistics; and Coop tasks for shared delivery and role-based handoffs.The game descriptions specify distinct task mechanics, including obstacle avoidance, resource orders, crystal defense, and restricted-area passing.
- Contamination avoidance: Contamination analysis tests whether models recognize games from screenshots and describe their mechanics from visual inputs.Table 2 compares OmniGameArena with existing benchmarks on recognition and mechanics-description rates.
4 Game Agent Harness
The harness combines a per-episode visual-action loop with a reflective outer loop that refines skills across rounds. Persistent experience, validated skills, and score curves support autonomous improvement while rollback limits catastrophic drift.
- 4.1 Per-Episode Loop: The per-episode loop exposes timestamped RGB observations and executes chunked keyboard-mouse actions until termination.VLM agents receive a bounded history of observation-response pairs alongside the current frame and optional skill prompt.
- 4.2 Improvement Dynamics Curve: The reflective outer loop runs K episodes under a fixed skill, then uses trajectories and persistent state to produce a refined skill.Reflection uses a small fixed tool surface while allowing the reflector to choose what to inspect, how many calls to make, and when to terminate.
- 4.2 Improvement Dynamics Curve: Round 0 uses an empty skill as the cold-start baseline, after which autonomous Explore, Diagnose, Validate, and Distill stages refine the prompt.The reflector reads new trajectories, the experience notebook, and the prior skill before producing the next skill.
- 4.2 Improvement Dynamics Curve: Persistent state includes a factual experience notebook, validated skills, and the best-skill record that carries information across rounds.The notebook is capped at 2000 tokens, while player-visible skill prompts are capped at 1200 tokens.
- 4.2 Improvement Dynamics Curve: If a round falls below half the best prior score, the harness restores the best skill before continuing reflection.This rollback rule is designed to guard against catastrophic skill drift.
- 4.2 Improvement Dynamics Curve: After R rounds, the score sequence [S0, S1, . . . , SR] forms the Improvement Dynamics Curve rather than reducing performance to a single round.The curve captures differences such as early versus late convergence and monotone versus oscillating improvement.
5 Experiments
The experiments evaluate heterogeneous agents across Solo, PvP, and Coop games, then measure reflection-driven improvement and transfer to held-out variants. Results show task-specific leadership, coordination headroom, mid-curve peaks, and a dissociation between origin-task gains and transferability.
- Experimental protocol: The protocol evaluates twelve agents across seven Solo, three PvP, and two Coop games, using cold-start play and IDC reflection with held-out variants.IDC characterizes both original-task improvement and transfer to three unseen variants per game.
- Cold-start leaderboard: No single model dominates Solo games: GPT-5.5 leads four of seven, while Claude Opus 4.6 and Gemini 3.1 Pro lead other tasks.Capability ranking is task-specific rather than monotone in release order.
- Cold-start leaderboard: Coop widens the commercial/open-weight gap: both Qwen3.5 checkpoints score exactly 0.000 on both games, while the strongest model reaches 0.368 and 0.184.The results leave substantial headroom for LLM-LLM coordination.
- Improvement on the original task: All four IDC agents improve over Round 0; LastStand best-round gains range from +0.54 to +0.70, but peaks typically occur before Round 10.The two Opus models lose 0.40 to 0.52 between their best and final rounds, motivating best-skill rollback.
- Transfer to held-out variants: SharedFloor transfers universally across held-out variants, with 16/16 positive gains, while LastStand transfer depends on skill style rather than origin gain magnitude.GPT-5.5 transfers positively on all three LastStand variants despite the smallest origin gain, whereas Opus 4.7 transfers negatively on every variant.
6 Conclusion
OmniGameArena provides a twelve-game UE5 benchmark spanning Solo, PvP, and Coop, while IDC produces multi-round self-improvement trajectories. Together, they expose both score evolution across reflection rounds and learned-skill behavior on held-out variants.
- Benchmark: OmniGameArena contains twelve newly built UE5 real-time games spanning Solo, PvP, and Coop.The games use unified evaluation across these interaction regimes.
- Improvement Dynamics Curve: IDC is an agentic-reflection harness that produces multi-round self-improvement trajectories.It extends evaluation beyond a single leaderboard score.
- Improvement Dynamics Curve: IDC exposes score evolution across reflection rounds and learned-skill behavior on held-out task variants for each agent-game pair.These are the two additional observables beyond single-round scores.
Limitations
The IDC evaluation is limited in environment, variant, model, and skill-library scope, and it uses the same underlying model for both player and reflector.
- Scope: IDC covers only two environments, three held-out variants per environment, and four agents because of compute constraints.Scaling to additional games, variants, and models is identified as future extension.
- Skill format: The reflector replaces one bounded skill prompt each round instead of maintaining a growing skill library.Library-based extensions are described as orthogonal to this round-by-round refinement.
- Model pairing: Each agent uses the same underlying model as both player and reflector, leaving asymmetric player–reflector setups untested.The paper gives a smaller player paired with a stronger reflector as an example of an untested asymmetric setup.
A Latency-Controlled Real-Time Eval
The latency-controlled evaluation injects model decision time into the simulated game timeline, revealing task-dependent effects that are not captured by paused-decision scores.
- Protocol: LCRT pauses inference, then injects the model’s measured decision latency before executing the action; PDQ executes immediately after inference.LCRT uses model-side timing and excludes network round-trip noise, while PDQ treats decision time as free.
- Protocol: LCRT is reported for four models with usable model-side timing signals and selected Solo or cooperative tasks where latency affects state or time budget.Other agents are excluded when available timings include client-side overheads; the selected Solo tasks are LastStand, MonsterShoot, and SoloCraft.
- Results: About 18 actions per game under LCRT versus approximately 42 under PDQ compresses matches into low-scoring, frequently drawn games.The reduction follows from charging decision latency against the game clock.
- Results: GPT-5.5 drops from 0.368 under PDQ to 0.048 under LCRT (∆= −0.320) on SharedFloor, yet ties Opus 4.6 for the best LCRT score.GPT-5.5 is the slowest model and completes the fewest interactions in the latency-controlled setting.
- Interpretation: Latency effects are task-dependent: they reduce throughput, compress symmetric play, and can benefit near-stationary survival strategies.The paper concludes that PDQ margins do not transfer directly to real-time deployment.
B Qualitative Comparison with IDC
IDC improves behavior in both evaluated environments by making survival decisions safer and cooperative behavior more effective and task-aligned.
- Behavioral changes: After IDC, the survival agent maintains safer positions and achieves a higher score, while cooperative agents show stronger coordination and complete more objectives.Before IDC, survival positioning is unstable and cooperative agents complete fewer objectives.
C Skill Inspection
The inspected IDC skills specialize by environment: LastStand favors conservative survival, whereas SharedFloor encodes coordinated role and station management.
- LastStand: LastStand skill prompts converge on conservative tile-survival behavior.The appendix limits this inspection to LastStand and SharedFloor and excludes ObstacleRun3D.
- SharedFloor: SharedFloor prompts emphasize cooperative division of labor, station alignment, and order-refresh handling.These prompt contents provide the qualitative basis for the divergence discussed in the main text.
- Comparisons: The qualitative comparisons use GPT-5.5 on Last Stand and Gemini-3.1-Pro on Shared Floor.The figure captions identify the model used for each comparison.
D Visualization
This section presents representative trajectory visualizations across all twelve games and lists best skill prompts for LastStand and SharedFloor across four models.
- Trajectory visualizations: LastStand visualizations cover one model per row across five sampled trajectory frames.The figures include LastStand, MonsterShoot, ObstacleRun2D, ObstacleRun3D, SceneEscape, and SoloCraft among the displayed games.
- Trajectory visualizations: Cooperative visualizations show one model pair per row for HandoffRun and SharedFloor, with five sampled frames from each episode.These rows represent cooperative episodes rather than single-model trajectories.
- Trajectory visualizations: PvP visualizations show one representative matchup per row for CrystalGuard, MidlineClash, and SkyDuel, using five sampled frames per match.The figures depict representative match progressions for the three PvP games.