The research team explored implementing a more interactive LLM system by modifying the model’s inference state (KV-cache). This approach was previously used in previous laboratory papers “Hogwild! Inference” and “AsyncReasoning”. Recent progress shows that the Qwen3.8-27B agent successfully applied similar techniques for interactive gaming in the DOOM environment. The team believes that model inference/run-time design is a new dimension of enhancing agent capabilities that deserves further exploration, in addition to models and tool sets.
The research team explored implementing a more interactive LLM system by modifying the model inference state (KV-cache). This approach was used in previous laboratory papers “Hogwild! Inference” and “AsyncReasoning,” and future work was previewed in the latest blog, where the Qwen3.8-27B agent used similar techniques for interactive gaming in the DOOM environment. The team believes that model inference/runtime design is a new dimension of enhancing agent capabilities worth exploring, in addition to models and toolkits.