Écrire un commentaire

Does the introduction explain the objective of the research presented in the preprint?

Yes

The introduction explains the objective by setting up a gap and then posing a question. It moves in 4 steps: 1. Drug discovery means searching a huge chemical space for molecules meeting competing objectives. Scoring oracles range from cheap and low-fidelity to expensive and high-fidelity, and that cost limits how much space gets searched. 2. LLMs are a good fit, but existing RLVR work trains only on cheap oracles — and training directly on expensive ones is impractical because online RL becomes too slow or produces stale rollouts. 3. So the question: can LLMs be post-trained for expensive, high-fidelity objectives by scaling training on cheaper synthetic tasks instead? 4. The objective is phrased as a question ("we investigate whether") rather than an aim. The key phrase is "too expensive to train on" — it names the constraint that makes the question worth asking and points straight at the synthetic-task solution.

Are the methods well-suited for this research?

Somewhat appropriate

1. The frontier comparison isn't fair. Trained models had 500 steps of practice with the exact prompt format. Frontier models got none. Table A3 shows Claude Opus 4.8 failed formatting in 39 of 200 rollouts, GLM-5.2 in 71 — and every failure counts as zero improvement. Part of the reported gap is formatting, not chemistry. No few-shot baseline separates the two. 2. The error bars are too wide for the rankings. The top result is +3.88 ± 1.98 kcal/mol. Twelve recipes are compared, each with one training seed, 40 systems, five rollouts. Change the reward weights and the ordering reshuffles (Figures 3, A3, A4). The broad claim holds; specific claims like "T2 beats Mix at 300 steps" don't. 3. One base model, one metric. Everything starts from Qwen3.6-35B-A3B, so the curriculum effect may not generalize to other models. And SQM2.20 is the only measure of binding — no FEP or experimental check. Trained models strip anionic groups for a mean 26.1 kcal/mol gain, which could be real chemistry or an artifact of the scoring function. Nothing distinguishes them. 4. Small asymmetry. The scaffold-checking tool is given at evaluation but withheld during training.

Are the conclusions supported by the data?

Somewhat supported

The main claim is solid: cheap synthetic tasks transfer to an expensive held-out objective. Every curriculum model beats T1(500), and the T2(500) collapse shows the curriculum is doing real work. The efficiency gains (7.7 turns vs 13–18) are large enough that noise doesn't threaten them. The authors are also honest where it costs them — they state plainly that no scaling law appears, and they flag T1(250)–Mix(250) as an exception to their own mixing conclusion. Where they reach past the data: 1. "Surpasses frontier models." Their own Table A3 shows Opus 4.8 failing formatting in 39/200 rollouts, GLM-5.2 in 71, each scored as zero. The data is reported but never qualifies the claim. 2. Recipe rankings stated too firmly. "No clear benefit to mixing" rests on ±1.98 kcal/mol error bars, one training seed each, and an ordering that reshuffles under different reward weights. 3. Chemistry read as strategy. Exploring worse intermediates is called "encouraging" exploration; it's equally consistent with noisy search. Anionic-group removal worth 26.1 kcal/mol is treated as a learned principle, but could be an SQM scoring artifact. Nothing tests which. 4. Wet-lab extension. The conclusion suggests the approach may transfer to experimental tasks. Hedged, but this study is entirely in silico. Net: reasonable interpretations throughout, with a few claims carrying more confidence than single seeds, wide error bars, and one scoring function can support.

Are the data presentations, including visualizations, well-suited to represent the data?

Somewhat appropriate and clear

What works. Figure 1 carries the paper. The three-tier schematic with real prompt text makes the task hierarchy immediately legible, which matters most since task design is the contribution. Table 2 pairs the headline metric with turns, SQM calls, and constraint satisfaction, so efficiency and affinity can be read together. Table A3's failure breakdown is transparent about where every rollout went. Figure A6 uses a log scale for a wide range and marks unavailable closed-source reasoning as "∼" rather than zero. Error bars appear with their definition stated. What holds it back. 1. Sign convention. Lower SQM means tighter binding, but "relative BA improvement" is built so higher is better. Figure 5 then labels an improvement "ΔSQM = +5.5 kcal/mol." Every caption has to restate "higher is better" — a fix, but also a sign the convention isn't self-evident. 2. Figure 3, right panel. Twelve points, error bars omitted "for clarity." Since the y-differences are about the size of the SEs shown in the left panel, this makes separation look cleaner than it is. 3. Figure 2, right panel. Five overlapping curve pairs in similar greens and oranges, distinguished by dashed vs solid. Telling T1(350)–T2(150) from T1(350)–Mix(150) is genuinely difficult. 4. Color accessibility. The green/orange trained-vs-zero-shot split (Figures 3, A3, A4, A8) is the standard deuteranopia problem. Figure 7 uses a red-green diverging heatmap — the least accessible diverging choice — and suppresses values below 10%, so most cells are readable only by color. 5. Crowding. Figure 6's leader lines converge on small labels near the origin. Figure A10's rotated labels are hard to read and ordered by value rather than recipe family. Net: the tables are strong and the schematic is excellent, but the color palettes, the sign convention, and Figure 3's missing error bars add friction and slightly oversell the pattern.

How clearly do the authors discuss, explain, and interpret their findings and potential next steps for the research?

Somewhat clearly

Clear where they explain mechanism: The T2(500) collapse gets its own appendix with reward-decomposition curves showing the policy saturating constraints while making zero binding progress — that explains why the curriculum is needed, not just that it works. Design choices come with reasons: the tool-penalty warmup is justified by what failed without it, and withholding the scaffold tool during training by the mode collapse it caused. They also state plainly that no scaling law appears, and flag T1(250)–Mix(250) as an exception to their own conclusion. Next steps are concrete: turn-level rewards, supervised finetuning, on-policy distillation. Less clear where they should consolidate: The conclusion is one substantive paragraph and mostly recaps. It never revisits the paper's own complications — the formatting failures in Table A3, the reward-weighting sensitivity in C.2, the missing scaling law. Those sit in appendices and stay there. There is no limitations section. Single base model, single scoring function, one seed per recipe — none acknowledged as bounding the claims. A few readings are asserted rather than argued. Worse intermediate designs are called "encouraging" exploration, with no alternative considered. Anionic removal is treated as a learned principle without asking whether it exploits an SQM bias, and no next step proposes checking. Net: strong mechanistic explanation, honest about the negative result, but the discussion doesn't fold its own caveats back into what it concludes.

Is the preprint likely to advance academic knowledge?

Somewhat likely

The advancements are real: It addresses a genuine bottleneck. Chemically meaningful oracles are too slow for online RL — SQM takes tens of minutes per evaluation, FEP takes GPU-hours — which has confined RLVR molecular design to cheap, low-fidelity rewards. Showing that a curriculum over cheap synthetic tasks transfers to an expensive held-out objective offers a concrete way around that, and the pattern should apply wherever the reward you care about is too costly to train on directly. The T2(500) collapse is the most valuable finding. Training directly on the harder task produces a degenerate policy that saturates constraints and never improves binding — with reward decomposition showing exactly how. That converts "curriculum helps" into "curriculum is necessary here," which is mechanism, not just a benchmark number. The efficiency result deserves more attention than it gets: a 35B model competitive at 7.7 turns versus 13–18 for frontier models, using fewer tokens. Cost per design matters in practice. The task construction is reusable — 200,000 verifiable Tier-1 tasks from 116 descriptors, plus the structural textification approach. Why not "highly likely." Curriculum learning and synthetic data scaling are established ideas. The novelty is the application and the specific low-to-high fidelity transfer in chemistry — a meaningful instance of a known principle, not a new principle. The frontier comparison carries much of the paper's rhetorical weight but is confounded by formatting failures the authors report and never reconcile. One base model, one scoring function, one seed per recipe, wide error bars: the finer rankings won't replicate unchanged. And SQM2.20 is a proxy with no orthogonal check. Whether the models learned chemistry or learned SQM's biases is unresolved — and it's the question that determines whether any of this transfers to real campaigns.

Would it benefit from language editing?

No

The writing is clean and professional. Sentence structure is controlled, technical terminology is defined at first use (Tier-1, Tier-2, SQM, the recipe notation), and the transitions between sections are logical. Notation is consistent across the methods and appendices. Nothing in the prose obscures the argument. The one point worth flagging isn't a language problem but a convention problem: lower SQM means tighter binding, while "relative BA improvement" is defined so higher is better. That mixed polarity forces the authors to restate "higher is better" in nearly every figure caption. It's handled correctly throughout — no errors — but a reader has to hold two opposite directions in mind. That's a presentation choice an editor might question, not a grammatical issue.

Would you recommend this preprint to others?

Yes, but it needs to be improved

Worth recommending to anyone working on RL for scientific design, molecular generation, or curriculum construction for expensive objectives. The core finding is useful and the T2(500) collapse — with its reward decomposition — is the kind of mechanistic negative result that saves other groups from a dead end. The task-construction recipe is reusable, and the appendices are transparent enough that a careful reader can audit the claims. What would need addressing before it's "high quality": 1. A fair frontier baseline. Some format adaptation or few-shot prompting, so the 39/200 and 71/200 formatting failures stop inflating the margin the paper's headline rests on. 2. Multiple seeds. One training run per recipe with ±1.98 kcal/mol error bars can't support the pairwise recipe rankings currently asserted. 3. An orthogonal validation of SQM. Even a small FEP subset would help settle whether the anionic-stripping behavior is chemistry or scoring-function exploitation. 4. A limitations section, and a conclusion that folds in the reward-weighting sensitivity and formatting confound rather than leaving them in appendices.

Is it ready for attention from an editor, publisher or broader audience?

No, it needs a major revision

The judgment turns on whether the fixes are editorial or require new experiments. Two of them require new experiments. Needs new work: 1. A fair frontier baseline. The headline claim — a 35B model surpassing much larger frontier models — rests on a comparison where trained models had 500 steps of format-specific practice and frontier models had none. Opus 4.8 failed formatting in 39/200 rollouts, GLM-5.2 in 71, each scored as zero improvement. Re-running the frontier evaluations with format adaptation or few-shot prompting is a substantive addition, not a revision of text. 2. Multiple training seeds. One run per recipe with ±1.98 kcal/mol error bars cannot support the pairwise rankings the paper asserts. Even two or three seeds on a subset of recipes would change what can be claimed. Editorial, but downstream of the above: 1. A limitations section. 2. A conclusion that integrates the reward-weighting sensitivity and formatting confound rather than leaving them in appendices. 3. Softening the recipe-level comparisons to match the statistical resolution. An orthogonal check on SQM — even a small FEP subset — would strengthen the anionic-removal interpretation considerably, though it's more "would substantially improve" than "must have." None of this reflects on the underlying research, which is well designed and honestly reported. The transfer result and the T2(500) collapse would survive any of these revisions. But the paper's most quotable claim isn't currently supported at the level it's stated, and fixing that means going back to the compute, not the manuscript.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they used generative AI to come up with new ideas for their review.