PUBG Ally paper reports +25.1pp recommend net in 141 countries

Surveys of players in 141 countries during live service showed positive responses to a voice-enabled AI teammate in PUBG: BATTLEGROUNDS exceeded negative by 25.1 percentage points when asked about recommending it, according to a new arXiv paper introducing the system as 'PUBG Ally.'

The paper describes Ally as an embodied agent that 'combines agentic tool use with real-time game control.' A language-model agent inspects game information, interprets player speech, maintains context and issues high-level action choices; a faster control layer handles movement, combat and recovery. The authors argue the split is required because the system 'must perceive and respond to a constantly changing game world under strict latency constraints while interacting naturally with players, keeping its speech synchronized with its actions.'

Training relied on live play rather than synthetic rollouts. The team collected data 'across nearly 39k sessions in which real players play alongside Ally,' logging speech, agent decisions, tool use, actions and feedback, then reconciled offline metrics against player preference comparisons to 'identify gaps between offline evaluations and player preferences.'

For deployment, the paper lists a stack the product team had to layer in: 'model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction.' It does not publish latency numbers, the compressed model's size, or the raw positive-versus-negative split sitting behind the 25.1-point gap.

Among surveyed players whose sessions were verified in game records, respondents described Ally 'not only as a tool but also as a teammate or companion.'