Latent reasoning that decides how long to think, and gets it exactly right.
Coconut feeds the last hidden state back in as the next input embedding. Reverie supervises every thought to decode to its gold reasoning step, and lets a PonderNet halt pick how many thoughts each problem gets. One loss, one stage, no RL.
L = Σₙ pₙ · CE(answer, W yₙ) # PonderNet answer loss
+ α · Σᵢ CE(path[i], W yᵢ) # every thought decodes to its gold step
+ γ · (−log p_m) # halt at the teacher's hop count
+ β · KL(p ‖ Geometric(λ_prior)) # don't collapse
γ is the point. The generator knows how many hops each problem needs, so train the halt against that. One MLE term.
It works. 2 hops gets 2.0 steps, 3 gets 3.0, 4 gets 4.0, ρ = +1.00. Not "correlated". Exact, across 606 seeds now, not just one. Kill γ and the halt does not pin to max depth, it pins to zero: the model stops thinking entirely, and at that scale that's a real accuracy win over full Reverie (8 sigma), not a wash. Calibration costs something, and dropping both α and γ turns out to be a trade, not a fix: it closes Reverie's two-layer loss to nocot completely (7-12 sigma, every width) but gives back most of Reverie's one-layer win over nocot too, sometimes turning it into a loss. See docs/multiseed.md for the numbers and what's queued to find out which term the one-layer edge actually needs.
No accuracy edge at the default 2-layer width. No-CoT gets 0.847 on the search task with zero reasoning steps, Reverie gets 0.850. I figured it was cheating off component membership and added cross-edges to force real directed reachability (--connect). No-CoT went up: 0.847 → 0.882 → 0.940 as cross-edges went 0 → 12 → 24. A 2-layer transformer resolves reachability on graphs this small in one pass, so nothing here needs to think twice, and the multi-seed sweep confirms it: Reverie loses to No-CoT at every width tested, 2 layers deep.
Shrink the backbone to one layer and the story flips. A 1-layer transformer can't one-shot this task, and Reverie beats No-CoT at every width past 64, growing from noise at d=96 to 44 sigma at d=192. The extra latent passes only pay for themselves where the backbone actually needs them, and it's real: a beta sweep at two layers shows accuracy there is flat whether the halt is collapsed or maxed out, so the one-shot backbone genuinely doesn't care how much latent compute gets spent, forced or not. Full numbers, and why dropping the forcing terms costs Reverie most of this one-layer edge instead of just fixing the two-layer loss, in docs/multiseed.md.
make install
make demo # ~8 min CPU
make phase0 # the numbers above
make test0.43M params, from scratch. Binary choice, so chance is 0.5. Rust generates the graphs, JAX and Equinox do the rest. Method in docs/DESIGN.md, numbers in docs/paper.md.