- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes molecularization: transforming pretrained neural modules into “resource-aware molecular programs” through a five-stage pipeline — replace operators with molecularly representable primitives, adapt on a calibration split, generate structural candidates via molecular architecture search (MAS) under compiler-estimated budgets, select between MAS and a matched saliency candidate on an independent split, hard-commit, then run post-commit recovery — followed by lowering to CRN/DSD/DNA for autonomous one-pot execution.
Experiments cover four questions. (Q1) Naive molecularization costs 1.36–2.82 NLL across SmolLM2-135M and Qwen3-0.6B/1.7B/4B; adaptation recovers 37.4–79.4%. (Q2) Across a dense 9-point retained-channel sweep on three Qwen3 scales, the independent selector matches a post-hoc evaluation oracle at 27/27 points, switching to saliency at the two tightest Qwen3-1.7B budgets; a frozen-choice LAMBADA probe matches at 6/6. (Q3) Matched SmolLM2 ablations attribute 0.146/0.301/0.986 NLL to post-commit recovery at 75/50/25% reaction budgets; Mamba shows the same adapt–commit–recover ordering helps. (Q4) An 8-channel SwiGLU block is lowered to a one-pot program (64 species, 488 reactions, 4 external boundaries removed), and a sequence-resolved DSD model of the clock reaches 5.92 phase-progress cycles with 23/24 bounded perturbations passing.
Critical strengths
- The question is genuinely new and worth posing.
Molecular-computing work designs the molecular algorithm from scratch (Qian et al., 2011; Cherry & Qian, 2018), and neural compression optimizes digital objectives. Asking how a pretrained modern block should change when the target substrate is non-ideal, resource-limited and must execute autonomously is a real gap, and the four-way failure-mode decomposition — substrate mismatch, candidate misranking, post-commitment function shift, control chemistry — is a sensible way to organize it. Appendix B’s motivation is careful, explicitly declining to make joule-level efficiency claims.
Why it is consequential: This is the only reason my recommendation is “reject this version” rather than “discourage this direction.” It makes me want to see a revision. But a framing is not a result, and nothing in the paper’s execution follows from it, so it cannot carry acceptance on its own.
- Scope discipline and honest disclosure.
Resource axes are never merged across model families. Appendix J states that the 27 cells are operating-point breadth, not independent replications; that the LAMBADA probe shows no broad generalization; and that there is no wet-lab validation. Table 16 states plainly that a 50% budget removes only 11.1–14.3% of decoder-MLP hidden-path weights. Table 22 reports 0/27 externally calibrated displacement interfaces and that no external enumerator was run.
Why it is consequential: It raises my confidence in the assessment — but note which direction. These disclosures are what let me establish the weaknesses below: G.3 derives the resource model’s collapse to width, M.2 gives the 6/24/8 step counts, Table 22 gives the missing validation. The honesty is real and I credit it, yet here it functions as confirmation of the defects rather than as support for acceptance.
Critical weaknesses
- “Molecular” does no work in any experiment — the term is load-bearing for the novelty claim and inoperative at all three places it should act.
Substrate: undefined. App. F.1 says only that operators are replaced by “SiLU, molecular linear maps, or resource-constrained factorizations.” No functional form, encoding convention, dynamic range, error model, or leak/fuel coupling to the forward pass. Every number in the paper is a function of this unspecified map, so the 1.36–2.82 NLL penalty cannot be attributed to chemistry rather than to an implementation artifact. Resource model: provably equals retained width. F.3 states channels have uniform marginal cost; G.3 then derives C(ρ) = C₀ + ρhc. Table 5 confirms it (159,805,440 × 0.75 = 119,854,080 exactly). So the 27-point sweep optimizes a width budget, and the claimed differentiator over parameter/FLOP counting is absent. Motif search: explicitly not done. Table 13 marks the CRN-motif axis “Conceptual design space,” yet Sec. 3 and App. D.1 present it as what distinguishes MAS from pruning. The two halves never meet. Molecularized SmolLM2-135M needs 529,920 species and 159.8M reactions; the autonomous lowering is demonstrated on a block with 64 species, against a few hundred strands in the largest experimental DNA network the paper itself cites.
Why it is consequential: This is my main reason for rejection. Delete the word “molecular” and what remains is distillation plus gated channel pruning on 8 of 28 MLPs, with a validation-set choice between two candidates. It is also not a rebuttal-scale fix: specifying the substrate changes the meaning of every reported number, and making the resource model non-trivial requires a different experimental setup.
- Every stage comparison is confounded with optimization steps, and the headline selector result is near-automatic.
App. M.2 fixes distillation/MAS/recovery steps 6/24/8 on 2,048 calibration tokens — substrate adaptation is six gradient steps. All three central claims (adaptation repairs mismatch, recovery repairs commitment, the ordering transfers) take the form “more steps at stage k helps,” with no equal-total-step control. The tables exhibit the confound: Qwen3-4B at full budget after 6 steps is 3.0456 (Table 2), while the 90%-compressed model after 38 steps is 2.8624 (Table 19) — fewer channels, better NLL. Separately, MAS-versus-saliency gaps are 0.067–1.857 NLL, so choosing the better of two candidates on 2,048 tokens is near-certain; only one of 27 cells (4B at 90%, +0.0671) is within plausible selection noise. At 10% budget the pipeline is worse than naive molecularization at all three scales (Recovered = −14.1%/−112.9%/−8.6%).
Why it is consequential: It removes the empirical support for all three of the paper’s causal claims simultaneously, leaving the pipeline’s stage structure untested. Combined with Weakness 1, nothing substantive survives: the method’s defining component is undefined, and the evidence for its stages does not isolate them.
Questions for authors
- Please give the exact definition of the molecularized operator: the functional form substituted for Eq. 33, the concentration/encoding convention (dual-rail or otherwise), any dynamic-range or saturation limits, the approximation-error model, and whether leak or finite fuel affect the forward pass used for NLL. Which of these is responsible for the 1.36–2.82 NLL penalty? An ablation isolating each would greatly help.
- Given F.3 (uniform marginal cost) and G.3 (affine cost in width), does the Qwen3 MAS objective differ from structured pruning under a width budget by anything other than the gate parametrization? If so, where?
- Can you report equal-total-step controls: adaptation-only at 38 steps, and a no-commit baseline matched on total optimization steps? Relatedly, why does the 90%-budget Qwen3-4B model (2.8624, Table 19) beat the uncompressed adapted model (3.0456, Table 2)?
- How many sequential phases does one forward pass of a molecularized SwiGLU MLP require, and how does that compare to the 5.92 sustained clock cycles? What total wall-clock time and fuel inventory would one token of SmolLM2-135M imply at 529,920 species and 159.8M reactions?
- Would you run an external domain-level enumerator (Peppercorn/Multistrand/KinDA) on the released kernel and report agreement with your effective kinetics? Table 22 suggests this was planned.
- What happens if candidates are made deliberately close in quality — does the selector still match the oracle, and how does agreement degrade as the selection set shrinks below 2,048 tokens?
- At 10% budget the pipeline is worse than naive molecularization at all three Qwen3 scales. What is the mechanism, and does it bound the useful compression range?
- Figures 1, 4, 6, 7 are described as AI-generated conceptual illustrations. Were their contents (arrows, stage labels, motif equations in Fig. 6) verified against the implementation? I could not inspect the images.
2 comments
Q4. Sequential phases, autonomous clocking, latency, and fuel.
The native molecular SwiGLU compiler uses four functional phases: parallel gate/up projections, SiLU activation, multiplication, and down projection.
The autonomous implementation adds molecular clocking to coordinate execution. The reported 5.92 phase-progress cycles characterize the sustained operation of the clock subsystem.
Our native 8×8×8 autonomous block contains 120 formal species and 1,896 reactions before frozen-sparse specialization. Elementary-reaction lowering and logical DSD compilation further expand the program.
These resource measurements provide a basis for estimating implementation requirements. Physical token-latency and fuel-inventory estimates additionally require calibrated reaction rates, molecular concentrations, reaction volume, throughput, and control-chemistry consumption. We will state these requirements explicitly when discussing full-model scaling.
Q5. External DSD enumeration and kinetic validation.
We have extended the native compilation experiments to elementary CRNs and logical DSD programs, with complete reaction mapping for the tested sparse SwiGLU instances.
Independent domain-level enumeration and kinetic equivalence assessment using tools such as Peppercorn, Multistrand, or KinDA are the next validation steps. We will preserve the reported external-calibration coverage of 0/27 interfaces in the manuscript and clearly identify the additional verification needed for DNA sequence-level implementation.
Q6. Selector reliability with near-tie candidates.
We agree that the original 27/27 selector-oracle agreement should be interpreted together with candidate gaps and selection-set size.
We conducted a near-tie study using 41 synthetic candidate pairs whose evaluation-loss differences were at most 0.01 normalized MSE.
| Selection examples | Near-tie oracle agreement |
|---|---|
| 16 | 51.9% |
| 64 | 53.7% |
| 256 | 58.5% |
Larger selection sets improve reliability, but distinguishing closely matched candidates remains difficult.
We also performed a real SmolLM2 clean-split comparison. Across five calibration seeds and three retained-width budgets, validation selected the lower-test-NLL candidate in 11 of 15 comparisons. In the compiler-cost-aware reaction-budget study, agreement was 3 of 4.
We will report selection accuracy, candidate separation, and regret alongside the original oracle-agreement results, providing a more informative characterization of candidate-selection reliability.
Q7. Performance degradation at the 10% reaction budget.
The negative recovery results at the most aggressive Qwen3 compression level are an important part of the empirical trade-off.
They indicate that the useful compression range depends on retained capacity, operator approximation, and recovery optimization. We will emphasize the observed transition in performance as the reaction budget decreases and report the 10% results as a practical boundary of the current pipeline.
A targeted analysis of capacity loss and recovery dynamics would help distinguish the underlying mechanisms.
Q8. AI-assisted conceptual figures.
Figures 1, 4, 6, and 7 illustrate the computational pipeline and molecular design concepts.
We will review the illustrations against the implementation, including stage ordering, arrows, operator labels, and molecular motifs. In particular, we will identify the Figure 6 motif alternatives as an architectural design space and distinguish them from the motifs evaluated in the reported experiments. The AI-use statement will describe the role of AI assistance in figure preparation.
Summary.
The additional studies substantially extend the technical evidence in three directions:
- Explicit molecular operators: Pretrained-model ablations quantify approximation fidelity and the influence of chemical dynamics.
- Native molecular compilation: Real pretrained-weight SwiGLU programs are lowered into phased and autonomous CRNs, elementary reactions, and logical DSD constructions.
- Compiler-aware architecture optimization: Joint channel-and-edge sparsity improves SmolLM2 NLL under matched reaction budgets, while exposing the associated species-cost trade-off.
We will incorporate these results with precise experimental protocols and strengthen the presentation of optimization controls, physical resource accounting, and molecular execution requirements.
We thank the reviewer for the thorough technical assessment and the concrete suggestions. We have conducted additional experiments using pretrained SmolLM2-135M, independent WikiText-2 data splits, and the native MolNova compiler. These studies address molecular operator definitions, compiler-derived resource costs, optimization controls, candidate selection, and chemical execution.
We respond to the eight questions below.
Q1. Molecular operator definition and substrate mismatch.
We now provide an explicit molecular SiLU formulation. Signed inputs use dual-rail encoding:
x = x_+ - x_-, wherex_+ = max(x,0)andx_- = max(-x,0).Define the truncated exponential approximation:
E_P(t) = sum_{p=1}^P t^p/p!.The competitive-gate approximation is
g_P(x) = [1 + E_P(x_+)] / [2 + E_P(x_+) + E_P(x_-)],with molecular SiLU represented by
SiLU_P(x) = x g_P(x).The implementation supports approximation order, concentration scaling, finite reaction time, and kinetic parameters. We evaluate operator approximation and chemical dynamics separately to identify their effects on model accuracy.
In a follow-up experiment, we replaced SiLU across all 30 SmolLM2 MLP layers and measured whole-model NLL on a separate WikiText-2 test split.
The difference is only +0.00023 NLL, indicating that the order-6 analytical approximation closely preserves the pretrained model's computation under this protocol.
We also measured the actual gate-input distribution and found that 0.291% of preactivations exceed ±4. Separate kinetic experiments demonstrate additional sensitivity to finite reaction time and leakage.
We will document the precise submission-time operator and its configuration, and distinguish the contributions of approximation error, input range, and chemical kinetics to the observed substrate mismatch.
Q2. Molecular MAS versus width-constrained pruning.
The reviewer correctly identifies that the submitted Qwen3 search uses uniform marginal channel costs. Consequently, that experiment evaluates architecture selection under a retained-width constraint.
We have extended the search to jointly allocate hidden channels and frozen-weight reactions using compiler-calibrated chemical resources.
For the tested frozen-sparse MolNova topology, with
kretained hidden channels andenonzero signed weights, the native-verified structural cost model is:Reactions: R(k,e) = 8 + 44k + 4eSpecies: S(k) = 2312 + 10kThe coefficients were checked against native compilations of pretrained SmolLM2 fragments containing one and two full-input/output hidden channels.
We then evaluated joint channel-and-edge optimization across four complete SmolLM2 MLPs.
Each comparison uses the same calculated reaction budget. The hybrid method achieves the lowest mean NLL at both budgets while retaining fewer species than full-width edge pruning.
These two-seed results demonstrate a promising accuracy–reaction–species trade-off. They also motivate multiobjective architecture search with explicit limits on both reactions and species.
Q3. Equal-step adaptation and commitment controls.
We agree that optimization budgets must be matched to establish the effect of each pipeline stage.
The submitted Qwen3 experiments allocate six weight-adaptation steps, 24 architecture-gate steps, and eight recovery steps. Thus, the Qwen3-4B values of 3.0456 NLL after initial adaptation and 2.8624 NLL after the full pipeline correspond to different optimization budgets.
We performed follow-up controls on four selected SmolLM2 MLPs across five calibration seeds.
The stronger saliency baseline achieves lower NLL at every retained width. This result shows the importance of matching weight-training capacity when assessing the staged procedure.
We will clarify the optimization accounting in the manuscript and use the Qwen3 matched-update comparison as the appropriate control for causal claims about the original pipeline.