Trend 8 — Recursive architectures (vs. recursive training)

A separate branch: building models whose forward pass is itself iterative.

Papers

• spotlight · yIDgVx3OoN — From Growing to Looping: A Unified View of Iterative Computation in LLMs — (spotlight) From Growing to Looping — depth-grown LIDAS and looped UTs share the same depth-wise mechanistic signature. You can loop a depth-grown model at inference (no retraining) for ~2× gains on reasoning primitives.
• poster · Vxu6kcIjwV — Generative Recursive Reasoning Models — GRAM — stochastic latent recursion via amortized variational inference; strong on Sudoku-Extreme and ARC-AGI; outperforms HRM/TRM on multi-solution problems.
• short_paper · 1zDG1o16xB — Depth vs Recursion: Outperforming Transformers in Jigsaw Reconstruction — Depth vs Recursion — on jigsaw 5×5: encoder-only Transformer collapses to 57.6%, TRM holds at 94.2% with same 0.7M params.
• spotlight · aY5kmaNrwB — Tiny Autoregressive Recursive Models — (spotlight) Tiny Autoregressive Recursive Models — negative result: under autoregressive decoding, token-internal recursion doesn't help.
• poster · rIzREKws05 — Unrolled Policy Iteration for Tiny Recursive Models — Unrolled Policy Iteration for TRMs — formalizes TRM training as policy iteration with Lipschitz-bounded inner unrolling.
• poster · fWrhAOQZ5A — Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space — Reasoning Within the Mind — confidence-driven gradient updates on latent think-tokens.
• poster · 5UbKomO77O — Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning — Emergent Temporal Abstractions — discover RL options inside the residual stream of a pretrained autoregressive model.

Synthesis

The picture is that recursion in the forward pass is real but fragile: latent recursion crushes depth on visual grids (jigsaw) but disappears under causal autoregressive decoding. The "loop a depth-grown model at inference for free" finding is the most actionable.

Related

• Trend 7 — Failure Modes Catalogued (Tiny Autoregressive RM negative)
• Trend 4 — Test-Time Training at Scale (latent updates at inference)
• Home