Search papers, labs, and topics across Lattice.
This study rigorously investigates the phenomenon of "compositional ignition" in a 30M-parameter recurrent-depth reasoner, assessing whether it represents genuine computation or is merely an artifact of training data. By independently recreating the model and employing a pre-registered whole-signature gate, the authors demonstrate that the ignition effect is indeed real, characterized by a significant increase in decision margin during problem-solving. The findings reveal that the model's readout mechanism is critical to this ignition, with sharp resolution and consistent performance across divergent training paths, challenging previous assumptions about the nature of latent reasoning.
Compositional ignition in latent-reasoning models is not an illusion; it鈥檚 a robust computational phenomenon that dramatically enhances decision-making efficiency.
We test whether the"compositional ignition"reported in latent-reasoning models is real computation, an instrument artifact, or inherited from verbal training data. We grow an independent realization of a published 30M-parameter recurrent-depth reasoner from scratch (same recipe and seed), film its development, certify fidelity through a pre-registered whole-signature gate, and measure resolution in two channels at once: the vocabulary readout and the hidden state. The ignition is real and lives at the readout: arrival time rises lawfully with problem depth, resolution is sharp and holds, and the signature reproduces across two same-seed realizations with divergent training trajectories. At commitment the decision margin jumps 5.8-8.0 logits in one iteration, exceeding the 90th percentile of near-threshold non-event steps in 96% of cases; the signed margin's zero-crossing there is definitional and carries no evidential weight, so the evidence is that conditioned magnitude. The hidden-state direction snaps in raw geometry, meeting its pre-registered criterion (in the decoder's LayerNorm coordinates it attenuates just below our bar, so the composite decoder-coordinate claim is not confirmed), and then freezes in both (descriptively so in decoder coordinates; angular steps 52.9 to 1.2 degrees over eight iterations), while subsequent displacement is predominantly radial (0.961 of squared-norm) and readout-null to a measured bound (radial logit effect<=5.7e-6). An earlier velocity-trough claim is withdrawn: pre-registered normalization controls showed it coordinate-dependent. Intermediates were never recoverable through the tied readout (relay 0.00). All criteria were frozen before their data; the predictions ledger, including this paper's own withdrawn headline, ships in the companion repository.
