Latent Bridge: Continuous Slow-Fast Channel for Game Agents

Bojie Li and Noah Shi submitted a paper to arXiv on 23 Jun 2026 proposing the Latent Bridge, a learned continuous communication channel that connects a slow reasoning visual-language model and a fast reactive model for real-time game agents. They evaluate the design across 7 Atari games and a driving domain (MetaDrive) and release replay recordings and reproducible pipelines.

What did the authors build?

They coupled two frozen models (a 9B reactive model and an 8B reasoning model) and made the communication link the only trainable component, comparing a standard Text Bridge with a learned continuous Latent Bridge. The Text Bridge has the slow model write a suffix the fast model reads; the Latent Bridge projects the slow model's residuals into the fast model's input-embedding space in a LLaVA-style manner, avoiding a text round-trip.

The paper frames the problem as a latency-quality tradeoff: the reasoning VLM (Qwen3-VL-8B-Thinking) deliberates but requires ~1.5 s per response, too slow for a 15 Hz control loop, while a reactive VLM (MiniCPM-o 4.5) acts in milliseconds but underperforms on planning-heavy tasks. To compare channels fairly they keep both models frozen, tune the action decoder per channel on held-out seeds, and test across 7 Atari games plus MetaDrive.

How did the Latent Bridge perform versus the Text Bridge?

Across the 7 Atari games and MetaDrive, the Latent Bridge matched or beat the Text Bridge in every domain, producing large improvements in two games: MsPacman (+57%) and RoadRunner (+28%), and behaving as a safe drop-in elsewhere. The MetaDrive domain served as a controlled negative: the Latent Bridge was inert there because the Text Bridge added no value.

The authors also report destructive interference when both channels are combined: in RoadRunner the combination produced a -96% effect, so they conclude only one channel should be used. The benefit is highly predictable: the bridge helps iff slow reasoning already beats fast reaction (T > F), and the Latent and Text gains over Fast-Only move together with a correlation of r = 0.93. The experiments used tuning of the action decoder per channel on held-out seeds and compared against a Fast-Only baseline.

Why it matters

Real-time interactive agents must act in tens of milliseconds while also planning over seconds. The Latent Bridge offers a concrete way to preserve deliberative capabilities without forcing a text round-trip, and the paper shows measurable, domain-specific gains (for example MsPacman +57%). The result reframes the engineering tradeoff: if a slow reasoning model already improves performance over a reactive model, then a learned continuous projection can carry that reasoning into a fast control loop with predictable benefit.

What to watch

Check the authors' released replay recordings and reproducible pipelines to confirm the reported +57% and +28% gains and the -96% interference case. Also watch whether the same pattern (help only when T > F, and a correlation r = 0.93 between Latent and Text gains) holds across more environments beyond the 7 Atari games and MetaDrive.