At ReBlink, we’ve always believed that games are one of the best proving grounds for real AI. Strategy games, in particular, push decision-making, planning, and adaptability in ways most benchmarks can’t. That’s why we built ARBO: Arena Tactics — not just as a game, but as a platform to test and advance game agent intelligence.

Today, I want to share a bit about what we’ve built, why it matters, and where we’re heading next.

Introducing the Prompt Battler: A New Way to Train Game Agents

We recently released a research paper titled Elevating Game Agent Intelligence — a detailed look at how we created Prompt Battler, a multi-modal imitation learning model trained on expert gameplay in ARBO

The challenge: Arena Tactics isn’t simple. It’s a PvP turn-based strategy game with:

• Dozens of hero classes and unit types
• Deck-building mechanics
• Randomized Protocol cards
• Dynamic maps with fog of war, traps, and summons

We captured 78+ hours of expert matches, totaling over 12,000 state-action pairs. Then, we trained an agent to mimic expert-level play — not through reinforcement learning (which struggles with sparse rewards in strategy games), but via imitation learning.

‍

What Makes Prompt Battler Different

The Prompt Battler doesn’t just take in a game state and output an action. It processes:

• Visual inputs: minimaps and spatial layouts
• Structured vectors: unit stats, card features, operation metadata
• Natural language: full text descriptions of Protocols and Operations

We built custom encoders for each modality, plus a spatial-aware attention system that uses Gaussian maps over unit positions to influence action selection.

The model predicts:

1. What kind of action to take (move, attack, play card, end turn)
2. Which card or skill to use
3. Which tile(s) to target

This architecture has become a robust benchmark — and one we hope other studios will use to advance game AI in tactical and turn-based genres.

Going Beyond: LLMs and Promptable Auto-Battlers

Where it gets really interesting is what we built on top.

Using LLMs, we’ve created a system where players can prompt the game agents mid-match. Imagine typing:

“Focus fire on the backline healer”
or
“Hold this hill and stall for time.”

The agent interprets the intent and adjusts its strategy accordingly.

We took this further by creating a new game mode — a next-gen twist on auto-battlers where players issue high-level strategic prompts instead of micromanaging every move. It’s more expressive, more interactive, and still deeply tactical.

Why This Matters for Game Studios

If you’re a studio building a mid-core or hardcore game and need:

• Smarter single-player agents to fill in early
• Realistic QA bots to test builds
• New mechanics powered by AI (like strategy prompting)

—we’ve built the systems to help.

Members of our team have worked on projects with Meta AI (EgoExo4D, CVPR 2024), collaborated with Google researchers, and published at CVPR, WACV, MICCAI, and more. We’re applying that same R&D depth to solve real game dev problems.

What’s Next

We’re open to collaborating — whether it’s bringing Prompt Battler into your game, co-developing AI-driven features, or helping you stand up data pipelines for gameplay imitation learning.

If any of this resonates, reach out. Happy to demo it live!

‍

‍