Multics Scales Four-Player Training to Ten-Player Inference

Featured

01 Video Gen: Not Every Player Needs a Direct Link

Scaling multiplayer world models is not only a compute problem. Full connectivity makes attention costs grow quadratically with player count. It can also let irrelevant events interfere with each other. Centralized representations risk mixing those events together.

Multics replaces that setup with three communication layers: player, group, and global. It dynamically groups players from recent camera trajectories, then caps communication to limit unnecessary information flow. Players judged relevant are grouped for direct local interaction, while global tokens still support communication across groups. Group parameters are shared, so adding players does not require more group-specific parameters.

The abstract reports better cross-player consistency, action and camera control, and resistance to irrelevant events in four- and six-player settings. The model also scales from four-player training to six, eight, and ten players without retraining. Yet systematic evaluation still centers on four to six players. The abstract does not report detailed ten-player performance, so the strength of that extrapolation needs confirmation from the full paper.

Key takeaways:

• Communication topology may matter more than simply enlarging a multiplayer world model.
• Dynamic grouping can reduce both compute and interference from unrelated events.
• Scaling from four to ten players shows promise, but does not constitute thorough ten-player validation.

Source: Multics: Generative Multi-Player World Model with Hierarchical Communication

02 Image Gen: Attention Should Shrink Over Time

AReA traces two high-resolution generation problems to the same design choice: global attention throughout denoising. Those problems are structural distortion and quadratic compute.

Early denoising needs a global view to establish structure. Later steps only need local focus to refine details. AReA first uses downsampled keys and values to capture the overall composition. It then gradually narrows the attention window.

This approach reduces full attention’s quadratic cost. It also limits exposure to token counts and relative distances absent from training. AReA adds no parameters and requires no retraining. It beats existing training-free baselines across different models and target resolutions. Exact speedups and image-quality gains still need confirmation from the full paper.

Key takeaways:

• High-resolution distortion and high compute may share the same cause.
• Shrinking the receptive field during denoising makes more sense than keeping attention global throughout.
• Existing DiTs can adopt AReA without retraining, but deployment value depends on measured speed and quality gains.

Source: AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers

03 Agent: Isolate Inference State Between Clients

Persistent agents can carry one client’s goals into another client’s decisions through a shared KV cache. The findings therefore highlight the importance of relationship-specific state isolation.

Across 20 controlled market scenarios with conflicting two-client goals, OLMo-3-7B-Instruct departed from the current client’s goal in 22% of eligible decisions. Researchers used linear probes to locate specific attention heads in middle layers. They then ablated the competing client’s goal representation at those heads.

This intervention pulled decisions toward the current client more strongly than random ablation or removing non-goal tokens. That result provides causal evidence. Deployment teams should treat caches and other inference state as part of client-level isolation. The 22% rate comes from one 7B model and controlled scenarios. More models and tasks are needed before estimating production risk.

Key takeaways:

• Multi-tenant agents need relationship-specific isolation of the KV cache.
• Shared inference state can directly alter decisions, so this is not merely a privacy-leakage problem.
• A 22% deviation rate warrants testing, but cannot be projected directly onto production systems.

Source: When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions