MOSAIC-SR: Transformer-Guided Symbolic Regression for Scientific Equation Recovery
Peiyi Zheng, Yanming Kang, Hans De Sterck, Giang Tran
cs.LG
Quick verdict
Claims transformer-guided sketch search + repair tops symbolic recovery on every benchmark tested; strong, but baselines mix cited vs rerun numbers, no code.
Summary
MOSAIC-SR is a hybrid symbolic regression method coupling a pretrained Transformer with explicit search. Stage one contrastively aligns (bidirectional InfoNCE) a Set Transformer numeric encoder over 200 signed-log points with normalized synthetic expressions; stage two trains an autoregressive decoder to emit operator 'sketches' with variable and constant placeholders. At inference, stochastic decoding proposes diverse sketches, conditional MCTS assigns observed inputs to variable tokens, and the top 24 complete equations seed 24 parallel local searches of 300 mutation steps each, alternating structural repair with scale-aware Levenberg-Marquardt constant fitting from log-spaced starts; selection uses validation NMSE with a simplicity tiebreak, plus simplification and physical-constant snapping. On SRSD-Feynman it reports symbolic solution rates of 84.7/55.0/34.8% (Easy/Medium/Hard) and accuracy 96.7/90.0/53.2% (R2>0.999), retaining 74.0/54.5/31.2% symbolic recovery with dummy variables. Across six classic benchmarks (Nguyen, Korns, Keijzer, Vladislavleva, Strogatz, Livermore) it posts the highest symbolic rate everywhere (75.0% on Nguyen, 64.3% on Strogatz) while running 1.3-2.2x faster than PySR. Ablations show repair is decisive: decoder-only reaches 30.0-43.3% symbolic recovery on Easy versus 83.3-90.0% for the full pipeline (3 seeds). Limitations: baseline tables mix values cited from the SRSD-Feynman paper with the authors' own reruns; no significance tests; no code release stated; recovery degrades on high-dimensional targets.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 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
- MOSAIC-SR achieves the highest symbolic solution rate on every dataset evaluated (SRSD-Feynman with and without dummy variables plus six classic benchmarks) while ranking among the top two in predictive accuracy.
- A recovery-oriented pretraining scheme that emits normalized structural sketches, deferring variable assignment and constants to inference, preserves algebraic structure better than full-expression prediction.
- Local symbolic repair with scale-aware constant fitting is the decisive component: decoder-only recovers 30.0-43.3% symbolically on SRSD-Feynman Easy versus 83.3-90.0% for the full system (Table 4).
- The advantage persists under 1-3 irrelevant dummy inputs, with symbolic rates of 74.0/54.5/31.2% versus 40.0/36.0/17.2% for the strongest baselines (Table 2).
Strengths
Dual-metric evaluation exposes the fit-vs-recovery gap
Section 4.1 reports symbolic solution rate separately from R2-based accuracy using the SRBench verifier; this cleanly shows, e.g., SBP-GP at 87.3% accuracy but only 8.0% symbolic recovery on SRSD-Feynman Easy (Table 1), the exact failure mode the paper targets.
Component ablation quantifies each stage's contribution
Table 4 (decoder-only vs +MCTS vs full on Easy, 3 seeds: 30.0-43.3% vs 83.3-90.0% Sym) and Table 5's paths-by-steps grid (1 to 48 paths, 100 to 500 steps; Sym 56.7% to 87.8% at 500 steps) substantiate the claim that repair, not the prior alone, drives recovery.
Robustness to irrelevant variables is demonstrated, not just asserted
Runtime transparency on cross-benchmarks
Useful diagnostic on sketch size
Weaknesses
Baseline tables mix cited and re-evaluated numbers
Neural baselines collapse suspiciously in Table 3
No statistical inference; small problem counts
Simplification steps interact with the symbolic verifier
No code or pretraining-corpus release stated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MOSAIC-SR (arXiv 2609.20997, University of Waterloo) is a method paper for symbolic regression that positions itself against two families: search-based systems (PySR, Operon, GP-GOMEA) that start from random trees, and pretrained neural generators (NeSymReS, E2E-SR, SNIP-SR) that decode expressions in one shot. The thesis, laid out in Sections 1-2, is that high test R2 does not imply algebraic correctness, so evaluation must separate predictive accuracy from symbolic solution rate. Three claims are tested: (1) a pretrained Transformer proposing multiple structural sketches concentrates search in promising regions; (2) parallel local repair with scale-aware constant fitting converts good numerical fits into symbolically exact equations; (3) the resulting system attains the highest symbolic solution rate on SRSD-Feynman (with and without dummy variables) and six classic benchmarks while staying top-two in accuracy (Tables 1-3). The motivating evidence for the fit/recovery gap is concrete: SBP-GP reaches 87.3% accuracy but 8.0% symbolic recovery on the Easy split (Table 1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 is MOSAIC-SR and how does it work?
MOSAIC-SR is a symbolic regression system that recovers exact mathematical equations from data. A pretrained Transformer first proposes several 'sketches' - equation structures with variable and constant placeholders. A Monte Carlo tree search assigns the observed input columns to variable slots, and 24 parallel local searches then repair the structure (swapping operators, variables, or subtrees) while refitting constants with Levenberg-Marquardt from logarithmically spaced starting points, alternating until the validation error stops improving. The final expression is chosen by validation error with a preference for simpler forms once the fit is near exact.
How does MOSAIC-SR compare to PySR or NeSymReS?
The paper reports 84.7% symbolic solution rate on SRSD-Feynman Easy versus a cited 60.0% for PySR, and on six classic benchmarks (Nguyen, Korns, Keijzer, Vladislavleva, Strogatz, Livermore) the highest symbolic recovery everywhere - for example 75.0% on Nguyen-12 versus PySR's 8.3% - while running 1.3-2.2x faster than PySR. Unlike NeSymReS, which uses the neural model's output directly, MOSAIC-SR treats neural output only as starting points for explicit search and repair.
Does MOSAIC-SR still work when the data contains irrelevant variables?
Yes, according to the paper's dummy-variable experiments, which add 1-3 irrelevant input columns to SRSD-Feynman. Symbolic solution rates drop only modestly, from 84.7/55.0/34.8% to 74.0/54.5/31.2% on Easy/Medium/Hard, remaining 34.0/18.5/14.0 percentage points above the strongest competing method on each split.
Is MOSAIC-SR code available?
The paper text does not mention a code or data release. Reproducing it currently requires a full reimplementation, including roughly 650,000 pretraining steps (a ~17.4M-parameter model) on a GPU plus the MCTS and local-repair search stack, and the synthetic training-data distributions are only qualitatively described.
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 2-11 (rank plots, NMSE trajectories) are only summarized in the extracted text; visual claims about seed spread are taken on trust.
- Exact synthetic pretraining sampling distributions and the physical-constant snapping whitelist are not enumerated in the text; reproducibility is assessed on stated details only.
- Symbolic verifier specifics are deferred to the SRBench citation; its tolerance and sampling scheme are not restated in this text.
- Tables were parsed from linearized text with minor OCR artifacts in equations; values were cross-checked against prose where possible.
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.