Compressed Recurrent Feedback in Tsetlin Machines: A Reproducible Boolean-FSM Study
Ankit Kumar, Utkarsh Raj, Rishad Shafik, Sudip Roy
cs.LG
Quick verdict
XOR-folded 96-bit RTM feedback matches raw feedback on synthetic FSMs but its no-feedback control does slightly better; honest, reproducible, value unproven.
Summary
The paper replaces the raw clause-activation feedback of a Recurrent Tsetlin Machine (RTM) with a fixed-width compressed summary: 480 clause outputs are XOR-folded into 48 bits per timescale, retained by fast (alpha=0.3) and slow (alpha=0.9) exponential moving averages, and thresholded back to a 96-bit binary state, shrinking the per-member recurrent interface from 960 to 96 bits. Evaluation uses a reproducible Boolean finite-state-machine benchmark (two families, six circuits, exported transition tables and seeds) with 144 RTM runs and 90 neural baseline runs. The compressed model reaches 61.47±6.74% (6-bit) and 62.94±9.92% (8-bit) mean accuracy; raw feedback adds only +0.72/+0.58 pp with confidence intervals including zero, while costing 4.38×/3.71× more host time. GRU/LSTM baselines remain more accurate, and a no-feedback control retaining only two observed symbols and the previous prediction slightly outperforms clause feedback in both families, so clause-derived recurrence adds nothing on these tasks. A nearest-centroid probe shows compressed signatures fall below random reference at recovering hidden FSM state. Limitations: only three circuit and three training seeds, unadjusted t-intervals across seven ablations, no formal equivalence test, synthetic noiseless tasks only, and host-time rather than hardware measurements. The artifact (runners, arrays, hashes) is openly released.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
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
- XOR folding plus two-timescale EMA thresholding compresses 480 clause activations into a 96-bit recurrent state (182-bit input) with mean accuracy within about one percentage point of raw 960-bit clause feedback on the Boolean-FSM benchmark.
- Compression cuts measured host run time by 4.38× (6-bit) and 3.71× (8-bit) versus raw feedback (16.6→72.7 and 19.7→73.2 minutes, Table V).
- A no-feedback control retaining only two observed symbols and the previous prediction slightly outperforms both compressed and raw clause feedback, so clause-derived recurrence is unnecessary on these deterministic FSM tasks.
- Gated neural baselines (GRU/LSTM) remain more accurate at the family level, while the tested Transformer configuration underperforms the RTM by 4.58/5.50 pp.
- Compressed feedback signatures fail to recover hidden FSM state (nearest-centroid probe below matched random references, Table VI), so predictive utility does not imply state representation.
Strengths
Unusually honest negative-result reporting
The abstract, Table IV and Discussion state outright that removing clause feedback improves accuracy (+0.43 pp [0.09, 0.78] on 6-bit, +0.89 pp [0.34, 1.44] on 8-bit), and Section IX.A explicitly disclaims equivalence from CIs containing zero — no spin on the central comparison.
Strong reproducibility artifact
Section VIII exports experiment runners, all 144 RTM records, 90 neural records, 144 probe records, per-circuit manifests, complete transition tables, split arrays and SHA-256 hashes, with resume-from-run-key and an analysis mode regenerating all tables/figures from archived CSVs in seconds.
Careful matched, paired design
Causality and leakage discipline
Predictive utility separated from state representation
Weaknesses
The proposed mechanism shows no benefit on the only benchmark tested
Small seed sample with no multiplicity control
Synthetic, noiseless, stationary tasks only
Efficiency claims limited to host software
Neural comparison is descriptive, not controlled
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.06133, cs.LG) proposes a fixed-width compression of the recurrent feedback path in a Recurrent Tsetlin Machine (RTM). Instead of returning all 480 clause activations (240 per class) to the next step, clause outputs are XOR-folded into 48 bins (Eq. 3), tracked by two exponential moving averages (α = 0.3 and 0.9, Eq. 4), and thresholded at 0.5 into a 96-bit binary state (Eq. 5) — shrinking the per-member recurrent interface from 960 to 96 bits and total input from 1046 to 182 bits. Five independently seeded members vote. The vehicle is a synthetic Boolean-FSM benchmark: two families (6- and 8-bit hidden state), three circuit seeds each, label = parity of the next hidden state, 1000/300/500 train/val/test sequences per circuit. The claims are stated with unusual care, including the negative ones: folding tracks raw feedback within <1 pp at 10× narrower interface and 4.38×/3.71× faster host time; gated neural models stay ahead; a no-feedback control slightly beats clause feedback; compressed signatures do not encode hidden state.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 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 is a Recurrent Tsetlin Machine and why compress its feedback?
A Tsetlin Machine learns Boolean classification clauses with simple finite-state automata; the recurrent variant feeds clause outputs from one time step back in as inputs to the next, giving it sequence memory. The raw feedback vector grows with the clause bank (480 bits in this paper), which is costly on small devices, so the paper compresses it to 96 bits using XOR folding and exponential moving averages.
Does the compressed recurrent Tsetlin machine beat LSTM or GRU?
No. On the paper's Boolean finite-state-machine benchmark, GRU and LSTM baselines remain more accurate at the family level (GRU leads the RTM by 5.02 and 6.58 percentage points), while the tested Transformer configuration performs worse than the RTM. The compressed RTM's advantages are a 10× narrower feedback interface and 3.7–4.4× lower host execution time, not accuracy.
Did the compression hurt accuracy?
Only marginally and not significantly: raw clause feedback raised mean accuracy by +0.72 and +0.58 percentage points over the folded version on the two task families, with 95% confidence intervals that include zero, while costing 4.38×/3.71× more host execution time. The authors explicitly stop short of claiming statistical equivalence.
Is clause feedback useful at all in this study?
Surprisingly, no. A control model with no clause feedback — keeping only the last two observed symbols and the previous prediction — scored slightly higher than both the compressed and raw feedback variants (Table IV). The authors present this as evidence that sequence-model benchmarks need no-feedback recurrence controls.
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 1-4 appear in the extraction only as captions; their graphical content could not be inspected, though all six tables' data are present in the text.
- The GitHub URL in Section VIII renders with an internal space ('Ankitkumar1062/LBP2026 Ankit-Utkarsh'), likely a line-break extraction artifact; the exact repository path is unverified.
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.