1. The benefit of the simplex state is not isolated. All baselines in Table 1 are quoted from Movahedi et al. (2026) and Jolicoeur-Martineau (2025), not re-run, and the training recipes differ (for example, GFPR uses dihedral augmentation on Maze-Hard and FPRM does not, and the S5 model uses a causal convolution that the paper notes is not part of the original TRM). The paper acknowledges that Table 1 "does not by itself isolate the effect of the state space" (Sec. 4.2). The natural control is mentioned in the paper itself: Sec. 3.1 notes that a bounded latent z∈[−1,1]d updated by tanh(fθ(z;x)) also has a fixed point. A comparison with this bounded-latent variant, using the same network, training and inference budget, is the experiment needed to show that the gains come from iterating beliefs rather than from the training recipe. It is not reported.
2. The theoretical guarantee is limited. The fixed-point guarantee is existence via Brouwer's theorem, which holds for any continuous self-map of a compact convex set and says nothing about uniqueness, reachability or correctness. Whether the iteration reaches the correct fixed point depends entirely on training, and the paper states that "none of this certifies convergence at test time" (Sec. 3.2). This matters because training rollouts are about 32 steps on Sudoku while test budgets reach 35,000 steps, and some puzzles settle to a stable but incorrect grid. For Sudoku and Maze, the convex constraints of Sec. 3.4 are applied only as a terminal readout, so the existence argument for the constrained recurrent state applies only to S5.
3. The empirical validation is incomplete in several places.
- The fixed-point claims rest on very little evidence. The contribution list states that GFPR "settles to a fixed point while FPRM typically halts at a state that is not one". For GFPR, the only evidence is the single puzzle in Fig. 2. The paper reports no population-level residual statistics for GFPR, and does not say how many puzzles end in a stable but wrong grid. For FPRM, the evidence is the same single puzzle plus a power-iteration probe on 128 puzzles that, as stated in Sec. 2, has not been cross-checked with Arnoldi.
- The convex constraints are only partly supported. They help on Maze-Hard (92.0% to 96.0%) but not on Sudoku (94.99% vs. 94.95%), and the Sudoku number is on a 23,680-puzzle sample rather than the full test set used for the headline result.
- The language-model experiment does not support the claims made about it. It is not compute-matched: each Picard step is a full pass over the sequence, so 24 steps cost roughly 24 times a GPT-2 forward pass. The training data also differ: GFPR-LM is trained on FineWeb-Edu, an education-filtered corpus known to raise scores on knowledge-oriented benchmarks such as ARC and SciQ, while GPT-2 was trained on WebText. The appropriate baseline, a standard Transformer of the same size trained on the same 6B tokens, is not reported. Even so, GFPR-LM trails GPT-2 medium on three of four tasks, and App. F reports that GPT-2 large is higher on all four but omits it from the main table. The evaluation uses only multiple-choice likelihood; the wavefront decoder of Sec. 3.3 is never evaluated on generation. App. F also states that these scores "do not show that the simplex state has reached a fixed point", so the paper's central property is not verified in this setting.
Beyond these points, from my own perspective, what matters most about looped models is that they can iterate a free hidden state that carries whatever intermediate computation a problem needs. In GFPR, the only quantity carried from one step to the next is the belief state p (Eq. 2): each step recomputes the Transformer's hidden activations from Winp+Enc(x) and then discards them after the softmax, so all intermediate computation must pass through per-site probabilities plus a few auxiliary coordinates. To me this is a step back in expressiveness rather than an improvement. The evaluated tasks, small puzzles with a fixed set of choices per position, are where this restriction matters least, and the paper does not ablate the number of auxiliary coordinates or test a task that needs richer state. For a design that deliberately narrows what the state can represent, I would expect either a fundamental argument for why this is the right restriction, or evidence beyond toy-scale settings.
Overall, I cannot lean towards acceptance in the paper's current form.