PUBG Ally: A Conversational Embodied Agent as an AI Teammate

PUBG Ally. A Conversational Embodied Agent as an AI Teammate. By Beomsoo Kim and twenty-three others, from Krafton AI. Published on the twenty-fourth of September. No code repository is linked to it.

A teammate inside a live game is a hard thing to build, because two difficult capabilities have to live in one agent. It must perceive a game world that never stops changing, and respond to it under strict latency constraints. At the same time it has to interact naturally with a human player, talking with them, and keep what it says synchronized with what it does. An agent that speaks well but acts late is no teammate, and neither is one that plays well in silence. There is a second difficulty underneath the first. The player's speech and actions and the agent's speech and actions continually shape each other, and together they shape the course of the match. Because of that loop, training such an agent requires data from actual gameplay, with real players on the other side of it.

PUBG Ally is an embodied agent for PUBG: BATTLEGROUNDS that can reason, act autonomously, and play alongside players as a voice-enabled teammate. Its answer to the two capabilities is to split them across two layers. Agentic tool use handles the thinking and talking. Real-time game control handles the moment to moment play. A language-model agent sits on top, and a faster control layer sits beneath it, taking direction rather than taking over.

The language-model agent works through a controlled interface. Through it the agent inspects game information, interprets player speech, maintains context, decides what to say, and issues high-level action choices. Those choices steer the faster control layer, which handles movement, combat, and recovery.

Training draws on the live loop itself. The authors collect data across nearly thirty-nine thousand sessions in which real players play alongside Ally. Each record holds gameplay, player speech, agent decisions, tool use, actions, and player feedback, and these records feed iterative training. Evaluation is treated as something to be corrected too. Player feedback and preference comparisons are used to find where offline evaluations disagree with what players actually prefer, and the evaluation criteria are refined in rounds. Running in a live service adds its own demands: low-latency execution on the player's device, and safeguards for anything the agent says to a player. Five measures answer them. Model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction.

The evidence comes from the live service rather than from a benchmark. Players were surveyed in one hundred and forty-one countries. The analysis keeps only respondents whose play with Ally was confirmed in game records. Among them, asked whether they would recommend Ally, positive responses exceeded negative responses by twenty-five point one percentage points. The descriptions players gave matter as much as the margin. They described Ally not only as a tool, but also as a teammate or a companion.

Read end to end, this is less a single model than a production loop. A slow, conversational agent directs a fast controller. Real matches with real players become the training data. Player preference keeps correcting the offline yardstick, and compression and guardrails make the whole thing fit on a device and safe to speak aloud. Each part answers one constraint of the setting, and the survey figure is the measure the authors chose for whether the parts add up to someone a player would want beside them.