Tilde Research Launches 'One Layer Deeper' Competition to Explore Deeper Serial Computation in AI Models
AashaySachdeva · x · 2026-08-23
Tilde Research published a blog post discussing limitations of current reasoning models that scale sequential computation by generating more tokens. This approach forces some intermediate computation through generated tokens rather than entirely within the model's latent state, potentially serializing computations that could otherwise be performed in parallel. The blog proposes alternative architectures that could scale computation through additional latent depth, but notes these models have proven harder to train. The team launched 'One Layer Deeper' competition (concluding August 31, 2026) to study whether architectures, objectives and optimizers can be co-designed to learn deeper serial computation and productively extrapolate beyond reasoning depths seen during training. The competition uses repeated modular squaring as the public task.
More from Research
- Article: Ontology evolution from semantics to AI agents — adnan_hashmi · 2026-08-23
- Vero includes formal audit mechanism for machine-checked spec errors — dawnsongtweets · 2026-08-23
- Vero shows the gap: agents rarely build reusable lemma libraries (74% of proof lines) — dawnsongtweets · 2026-08-23
- Vero composition: 43 Lean 4 repos, 743 APIs, 2,705 formal specs — dawnsongtweets · 2026-08-23
- Vero: First Benchmark for Repository-Scale Formal Verification by AI Agents — dawnsongtweets · 2026-08-23
- CMU Blog: Forking-Sequences for efficient multi-horizon forecasting — rsalakhu · 2026-08-23