When Recursive Models Finish Computing
Hare Krishna, Shubham Singh, Stephen Ebert, Hao-Yu Sun
cs.LG · cs.AI · cs.NE
Quick verdict
TRM nominal-budget Sudoku failures are often unfinished (59.2%→87.5% at 512 steps); anisotropic-stability claim well-evidenced but Sudoku-scoped.
Summary
The paper asks what changes in a recursive model's latent state when a computation finishes, as opposed to being truncated by an inference budget. Using two public Tiny Recursive Model checkpoints — attention-based (6.83M params) and MLP-based (5.03M params) — on 1,000 hard Sudoku puzzles, the authors extend the nominal 16-step recurrence to 512 steps. Cumulative exact-solve accuracy rises from 59.2% to 87.5% (attention) and 74.4% to 91.9% (MLP); in both primary checkpoints roughly two-thirds of nominal-budget failures turn out to be unfinished computations. Puzzles are split into EARLY, LATE, and PERSISTENT completion groups; aligned to each puzzle's first solve step, latent motion drops sharply after completion, and the solve step shows an unusually large ZH update. Jacobian and perturbation analyses show completed states contract along the trajectory direction (median γnat ≈ 0.25–0.47 at step 512) while the same Jacobians retain expanding directions (σmax > 1 at every probed state), a phenomenon the authors call trajectory-conditioned anisotropic stability. The pattern replicates across both architectures and a third attention checkpoint, though the fate of injected expanding-direction perturbations differs (absorbed within 16 steps in attention, still 1.91× in MLP at k=16). Limitations: the finite 512-step horizon, Sudoku-dominated evidence, subset-based probes, and no adaptive stopping criterion for exploiting late computation in practice.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Extending TRM recurrence from 16 to 512 steps on 1,000 hard Sudoku puzzles raises cumulative exact-solve accuracy from 59.2% to 87.5% (attention) and 74.4% to 91.9% (MLP), so about two-thirds of nominal-budget failures in both primary checkpoints are unfinished computations rather than failures.
- When trajectories are aligned to each puzzle's first exact solve, latent motion drops sharply after completion, EARLY and LATE solvers converge to the same low-motion regime, and the solving step is the largest-ever ZH update for ~99% of LATE solvers.
- Completed states are locally contractive along the trajectory direction (median γnat 0.25–0.47 at step 512) while the same Jacobians retain strongly expanding directions (σmax > 1 at every probed state) — trajectory-conditioned anisotropic stability.
- The anisotropic pattern replicates across attention and MLP architectures and a second attention checkpoint, but the fate of v1 perturbations differs: absorbed within 16 steps in attention (0.10×) versus persistent in MLP (1.91× at k=16), consistent with higher direction transport (T1 = 4.9×10⁻⁵ vs 1.8×10⁻³).
Strengths
Solve-time alignment is a genuinely informative methodological choice
Aligning every trajectory to its own first-solve step (r = t − τi, §2.2, Fig. 2) rather than fixed wall-clock steps exposes the EARLY/LATE convergence that fixed-step plots cannot show, and reveals the solving step as the largest-ever ZH update for 99.3% of LATE attention solvers (§3.2).
Replication across checkpoints and architectures with quantified differences
The full pipeline is repeated on Attention-A, MLP, and Attention-B (Table 5), plus Easy Sudoku and Maze-Hard controls (Apps J, K). The anisotropic pattern replicates qualitatively (Pr[γnat<1<σmax] = 0.621 / 0.604 / 0.624) while magnitude differences are honestly reported (σmax ≈ 27 attention vs 6.4–6.7 MLP at completed states).
Above-average statistical hygiene
Exceptional reproducibility apparatus
Negative controls and honest disclosures
Weaknesses
Sudoku-dominated evidence base
Finite horizon underdetermines the headline architecture contrast
No deployable stopping criterion
Subset-based probing with a one-step linearization
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a precise question: when a recursive model trained with a fixed recurrence budget produces a wrong answer, is the computation unfinished or has it entered a persistently unsuccessful regime? Using two public Tiny Recursive Model checkpoints — Attention-A (6.83M params) and an MLP variant (5.03M params) — on 1,000 hard Sudoku puzzles from the Sudoku-Extreme test split, the authors extend the nominal 16 outer steps to 512. Three claims carry the paper: (1) nominal-budget failure is largely unfinished computation (cumulative exact solve 59.2%→87.5% attention, 74.4%→91.9% MLP; 283/408 and 175/256 of step-16 failures later solved, i.e., 69.4% and 68.4%); (2) latent motion drops sharply after each puzzle's first exact solve, with a distinctive ZH reorganization spike at the solving step (largest-ever ZH update for 99.3% of LATE attention solvers, §3.2); (3) completed states are contractive along their own trajectory direction (median γnat 0.25–0.47 at step 512, Table 1) while the same Jacobians retain strong expansion (σmax > 1 at every one of 750 probed states, Fig. 3c) — trajectory-conditioned anisotropic stability. The claims are scoped honestly in §5.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does trajectory-conditioned anisotropic stability mean?
At states where the puzzle is already solved, the recurrent map shrinks perturbations along the direction its own state trajectory is traveling (one-step gain γnat < 1), while the same Jacobian still contains strongly amplifying directions (σmax > 1). Completion is therefore stable only along the computation's own path, not across the whole state space — the two conditions coexist at about 62% of probed states.
Did extending the recurrence from 16 to 512 steps actually fix the models?
On 1,000 hard Sudoku puzzles, exact-solve accuracy rose from 59.2% to 87.5% for the attention model and 74.4% to 91.9% for the MLP model, meaning roughly two-thirds of the apparent 16-step failures were unfinished computations. However, 12.5% stayed unsolved even at 512 steps and the paper offers no adaptive stopping rule, so this is a diagnostic finding rather than a turnkey accuracy improvement.
Is the code and data available for this paper?
Yes. The paper states supplementary code including a CPU-only analysis pipeline that regenerates every reported figure from puzzle-level CSVs in about a minute, plus a documented GPU path to rerun from the original checkpoints (Appendix A). The checkpoints are public on Hugging Face and the task data comes from the public Sudoku-Extreme test split.
Do the findings generalize beyond Sudoku?
Only partially so far. A second, independently trained attention checkpoint replicates the pattern on Sudoku, and Easy Sudoku and Maze-Hard controls show the early-solver regime. But the full extended-recurrence, Jacobian, and perturbation analyses are confined to Sudoku, and the authors themselves state that broader claims require more checkpoints, architectures, and tasks.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as caption text in the extraction; visual details such as band widths and distribution shapes could not be verified.
- The extraction appears complete through Appendices A–K, but referenced external artifacts (e.g., ICLR_RESULT_INVENTORY.md, Hugging Face checkpoints) were not independently checked.
- Several supporting citations are 2026 preprints whose claims were taken at face value.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.