Runtime Control Moves Into the Open

Featured

01 WorldTrace Makes Long Video Memory Addressable

Interactive video models can lose old scenes once generation runs beyond the time horizons seen in training. The frames may still be stored, but the model cannot find them. WorldTrace reframes long-term consistency as a memory-addressing problem, not merely a context-length problem.

The KV cache keeps growing, but RoPE’s temporal offsets move outside the training distribution. Attention then struggles to retrieve the right frames. Directly compressing a position-rotated cache makes matters worse by averaging content with incompatible phases.

WorldTrace needs no retraining. It assigns each compressed memory slot a distinct virtual position within the training distribution. One variant compresses continuous history for temporal coherence. Another preserves landmarks at scene transitions for later revisits. On LoopBench, the variants improve temporal consistency by 15.5% and episodic recall by 19.5%. Before expanding cache capacity, test whether position encoding and compression have made old memories inaccessible. Broader interactions and longer horizons still need validation.

Key takeaways:

• Memory addressing may limit long-video consistency more than cache capacity.
• Cache compression should avoid mixing incompatible positional phases.
• Long-horizon world models need explicit tests for scene recall after detours.

Source: Addressable Memory for Video World Models

02 Stop Rebuilding TTT From Scratch

Test-time training has many variants, but each often arrives as a hard-coded architecture. That makes it difficult to identify which design choice produced the result. Modular TTT represents the inner learner as a directed acyclic graph.

Fast weights, loss functions, and learning rates become replaceable components with separate attribution. The reported ablations favor lower initial learning rates, weight decay, and a single-layer nonlinearity. Deeper fast-weight networks and normalization can hurt by driving activations too high.

Models with 410M and 1.45B parameters, trained on 100B tokens, match Gated DeltaNet in training loss and benchmark performance. The useful contribution is a reusable runtime-learning interface that reduces duplicate implementation. Its compositional range and engineering cost still require closer inspection.

Key takeaways:

• Evaluate fast weights separately from their internal learning rules.
• Deeper inner networks may underperform when activations grow too large.
• A shared modular representation speeds up ablations and reuse, not just benchmark gains.

Source: Modular TTT: Rethinking Test-Time Training as Composable Modules

03 ReASearch Puts Search Inside the Agent Loop

Optimizing prompts, programs, and ML workflows usually requires an outer-loop algorithm. It decides what to try, how to revise candidates, and when to stop. ReASearch assigns those decisions to the same tool-using agent.

The agent evaluates results, diagnoses failures, allocates budget, and adapts its strategy through persistent memory. Across 14 tasks, one shared loop beats specialized optimization systems in most cases. Gains over strong baselines range from 2% to 40%. Some results also exceed the best previously known human-designed solutions.

The open question is whether this unified interface transfers reliably across tasks. Proposal, evaluation, and stopping authority all sit with one agent. The full experimental design must show that validation and budget controls remain dependable.

Key takeaways:

• An agent can internalize search strategy instead of merely generating candidates.
• A reusable loop can reduce the cost of building task-specific outer algorithms.
• Centralized editing and evaluation demand strict validation and budget controls.

Source: The Optimizer Is the Agent: Reasoning-Driven Search across Prompts, Programs, and ML Workflows