Refine an idea over several passes.

Internal computation does not have to be written in words

Neural networks already transform continuous hidden vectors before producing output tokens. A recurrent system applies an update repeatedly to an internal state, potentially giving a difficult problem more sequential computation:

$h_{k+1}=F_\theta(h_k,x)$

Here x is the problem, h is the current latent state, and the same update operator F is reused. Generating intermediate text is another way to add sequential work, with different constraints and opportunities for inspection. Neither approach guarantees useful reasoning merely by running longer.

A visible example: distance waves

Our grid stores one distance estimate per square. The goal starts at zero; other squares start unknown. For a non-goal square with legal neighbors, the programmed update is:

$d_{k+1}(s)=1+\min_{s'\in\mathrm{neighbors}(s)}d_k(s')$

The goal stays at zero. The implementation caps unknown distances at 26 and leaves walls unknown. With at most 25 squares, any simple shortest path uses at most 24 moves, so the cap is larger than any finite shortest distance in this world.

Updates are synchronous: every new value uses the previous pass, not values just written during the current pass. The first pass informs neighbors of the goal. The second can inform squares two moves away. This explains why a distant start needs multiple passes, even though the same rule is reused each time.

The source exercise asks you to write this Bellman-style rule and explain the need for repeated passes. The displayed calculation supplies a concrete answer. It is an exact programmed procedure under its assumptions, not a learned neural update and not evidence that a neural system has discovered the algorithm.

Learning the update is a separate problem

A neural recurrent model must learn an operator that makes useful progress. Backpropagation through time treats repeated uses of the operator as an unrolled computation graph. A final loss can send gradients through those uses, accumulating contributions to the shared parameters.

Long unrolls can require substantial memory and can produce vanishing or exploding gradients. If useful intermediate targets are known, intermediate supervision can guide the process. Outcome-only feedback is more flexible but makes it harder to identify which internal steps helped or hurt. These are training choices, not guarantees of success.

Our lab changes no weights when it runs a pass. A neural experiment would need separate training data, a defined loss, parameter updates, and evaluation against the programmed baseline before making claims about learned reasoning.

Stopping is not the same as being right

Possible stopping rules include a fixed budget, small change in internal state, a learned halting probability, or an external task verifier. Convergence means the update approaches a stable state; it does not establish that the state represents the right world or solution.

Discard the walls and the distance calculation can stabilize for an imaginary open grid. The decoded route may cross a real obstacle. Exact replay rejects that route even though the internal numbers stopped changing. Conversely, an unreachable goal cannot be made reachable by adding passes.

An exact task check is strong evidence of satisfying the checked constraints. It does not by itself prove optimality, efficiency, or transfer to other tasks. Latent reasoning remains a research topic; this demonstration explains mechanisms and failure modes without claiming that repetition produces general intelligence.