{"id":"dce4c2f6-b0e4-4215-aca5-5cdce388a70f","arxiv_id":"2607.27251","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Depth Recursive transformer, which injects the depth index as a state and uses per-step losses, gives the best accuracy-per-FLOP trade-off among three recursive weight-sharing designs on small engineering surrogate benchmarks.","lead":"The paper compares three recursive weight-sharing transformer designs as cheap surrogates for semiconductor package stress/warpage and capacitor field simulation. It argues that treating depth as a recursive input and giving each recursive step a loss gives the best accuracy-per-compute trade-off, but provides no numeric tables or code to back the Pareto comparison.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim cites a per-step BPTT loss that never appears in the architecture, training, or experiment sections; no ablation isolates depth injection or loss structure.","rationale":"The reader's REJECT verdict focuses on the uncontrolled architecture comparison (SIMPLE vs TRM vs DEPTH jointly varying block count, normalization, recursion depth, latent states, and loss structure). That is valid and also central. My stress-test sharpens one specific internal inconsistency: the Conclusion asserts a per-step loss term during BPTT as part of the causal mechanism, but no such term appears in the model definition or experiments. This makes the central claim internally incomplete. The two concerns are related—both bear on whether 'depth injection + per-step loss' is truly the causal factor—but they are distinct: the reader emphasizes confounded baselines; I emphasize that one of the two named mechanisms is entirely undefined. Thus agreement is partial.\n\nThe strongest practical test is a controlled ablation that removes per-step losses from M7. If performance persists, the stated mechanism is not load-bearing. Complementarily, matching T between SIMPLE and DEPTH isolates the depth-state injection from recursion depth. Without such tests, neither the paper's central claim nor the reader's rejection can be resolved. The verdict remains REJECT: the preprint does not currently substantiate its headline causal attribution, and the missing loss definition is an internal gap, not merely a matter of external comparison. I am not claiming the DEPTH architecture is wrong or that the authors are misleading; the evidence presented is simply insufficient for the strength of the conclusion.","tokens_in":5479,"tokens_out":3207,"duration_ms":36970,"concrete_test":"Re-run M7 (DEPTH, 2-block RMS, T=16) with the per-step loss terms removed, training only on the final recursive output L(H(T)), keeping all other hyperparameters and the depth-state injection fixed. If MRR/Recall does not materially change, the claimed 'individual loss term during BPTT' is not the primary factor. As a second arm, run M3 (SIMPLE, 2-block RMS) at T=16 without depth injection; if M3 matches M7, the reported advantage is due to recursion depth or parameters rather than depth-state injection.","verdict_should_be":"REJECT","load_bearing_attack":"The Conclusion (§IV) states that DEPTH's advantage is due to 'explicitly uses depth as a recursively injected input in which each recursive output contributes an individual loss term during BPTT.' The depth injection is defined in Eq. 4, but the per-step loss is nowhere defined or referenced before the conclusion: Eqs. (2)–(5) give only forward recursions; §III-B describes datasets; and §III-C reports MRR/Recall Pareto results. No loss function, BPTT objective, or training procedure is stated. Thus a core causal mechanism of the headline claim is absent from the paper's own technical content, making the conclusion internally unsupported.\n\nEven if one grants a per-step loss, the architecture comparison cannot support 'primary factor' attribution. The closest controlled pair is M3 (SIMPLE, 2-block RMS, T=3) vs M7 (DEPTH, 2-block RMS, T=16), which differ simultaneously in recursion depth T and the injected depth-state S(t)W_s. M6 (1-block DEPTH) vs M2/M3 varies block count as well. No ablation varies depth injection while holding T, block count, and loss fixed, and no ablation removes per-step losses. Consequently, the paper does not show that depth injection or per-step losses—rather than simply more recursion steps or the added S(t)W_s parameters—drive the reported Pareto gains. Additionally, MRR/Recall retrieval scores are an indirect proxy; direct regression error on stress/warpage fields is never reported, so even the outcome measure is not fully aligned with the engineering surrogate task.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three recursive weight-sharing transformer families—Tiny Recursive Model (TRM), a proposed Depth Recursive Model (DEPTH), and a Simple Recursive Model (SIMPLE)—for small-data engineering surrogate modeling, specifically thermo-mechanical reliability of semiconductor packages (Stress10k, Warpage10k) and a PINN-based capacitor electrostatic field dataset. Seven configurations (M1–M7) are compared using Pareto analyses of MRR versus FLOPs and MRR versus parameter count. The authors claim that DEPTH, which injects a depth state S(t) into each recursive step, achieves the most favorable accuracy-per-parameter and accuracy-per-FLOP trade-offs, and the conclusion attributes this primarily to the recursively injected depth input and a per-step loss during backpropagation-through-time that is never defined in the paper.","tokens_in":5778,"tokens_out":3122,"duration_ms":34550,"significance":"If properly validated, the DEPTH architecture would offer a useful design guideline for resource-constrained transformer surrogates on low-dimensional engineering datasets. The paper's weight-sharing motivation is reasonable, and the proposed forward-recursion equations (Eqs. 1–5) are clear. However, the supporting evidence is almost entirely qualitative: no numerical MRR/Recall values, parameter counts, FLOPs, training hyperparameters, error bars, or repetitions are reported, and the central causal claim is not supported by a controlled ablation. The paper also uses retrieval-style metrics (MRR/Recall) rather than direct regression error on the engineering outputs, so the reported trade-offs may not reflect the actual surrogate quality needed for stress/warpage prediction. These gaps are load-bearing rather than cosmetic.","major_comments":[{"comment":"The conclusion states that DEPTH's advantage comes from \"explicitly uses depth as a recursively injected input in which each recursive output contributes an individual loss term during BPTT.\" However, no loss function, BPTT objective, or per-step loss is defined anywhere in the paper. Eqs. (2)–(5) present only forward computations, and Section III describes datasets and Pareto curves. This per-step loss is a core component of the attributed causal mechanism; its complete absence means the architecture cannot be reproduced and the main conclusion is internally unsupported.","section":"Section IV and Eqs. (2)–(5)"},{"comment":"The architecture comparison is confounded, so the \"primary factor\" attribution is not justified. The closest controlled pair, M3 (SIMPLE, 2-block RMS, T=3) versus M7 (DEPTH, 2-block RMS, T=16), differs simultaneously in recursion depth T and in the injected depth-state term S(t)W_s. M6 versus M2/M3 varies block count as well. No ablation varies depth injection while holding T, block count, and loss fixed, and no ablation removes per-step losses. Therefore the reported Pareto gains could equally be due to the larger recurrence depth, the extra S(t)W_s parameters, or other confounds.","section":"Section III-A (baselines)"},{"comment":"No quantitative results are reported. The reader is asked to rely on Figures 4 and 5 without access to the underlying MRR/Recall values, parameter counts, FLOP calculations, training hyperparameters, random seeds, or error bars. None of the Pareto-frontier claims (e.g., \"DEPTH models form the exclusive frontier\" for the parameter trade-off) can be verified or compared across datasets. For a hardware-aware evaluation, exact complexity and accuracy numbers are essential.","section":"Section III-C (Pareto analysis)"},{"comment":"MRR and Recall are retrieval metrics; direct regression errors on stress and warpage fields are never reported. For an engineering surrogate, the primary objective is accurate prediction of the response field, not ranking of candidate designs. The paper's central claim about \"accuracy-per-parameter and accuracy-per-FLOP\" is therefore evaluated on a proxy metric that may not align with the downstream engineering use, and no evidence connects retrieval performance to regression fidelity.","section":"Section III-C (outcome metric)"}],"minor_comments":[{"comment":"Typo: \"capactior\" should be \"capacitor.\" Also, the abstract mentions \"Recall, Mean Reciprocal Rank\" while the body uses \"Recall@K\"; clarify the metric name.","section":"Abstract"},{"comment":"The notation [e_1,...,e_L, z^(n-1), y^(n-1)] is ambiguous: it is unclear whether these are concatenated along the token dimension, fused additively, or combined by another operation. Dimensions are not specified, making the forward pass only partially defined.","section":"Eq. (2)"},{"comment":"The superscript/subscript expression \"ESt⊈E\" is not standard and is not defined. Please clarify how the initial embedding E and the depth states S(t) are combined.","section":"Eq. (4)"},{"comment":"Reference [16] \"Tiny recursive models\" lacks an arXiv identifier or venue details; reference [3] is marked as a 2026 preprint. These should be completed or removed.","section":"References"},{"comment":"The paper does not include any discussion of limitations, reproducibility (code/data availability), or potential overfitting to the author's own prior datasets. A reproducibility statement would strengthen the submission.","section":"General"}],"recommendation":"reject","confidential_remarks":"This is a very short empirical report that, despite an interesting architectural idea, does not meet the standards of a journal publication in its current form. The main conclusion is not supported by the evidence: the per-step loss is never defined, the comparison is confounded, and no quantitative results are reported. The evaluation uses the author's own datasets from prior papers, and no external validation or code is provided. Fixing these issues would require substantial additional experiments and rewriting, beyond a routine revision. I would encourage the authors to resubmit a version that includes the missing training objective, a proper ablation isolating depth injection and per-step loss, and full numerical results with standard errors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The architecture idea is real and modestly new: DEPTH injects a depth state S(t) as an additive bias at each recursion step, which is not in the cited recursive/looped transformer work. The equations for TRM, DEPTH, and SIMPLE are clear, and the problem — small-data surrogate modeling for semiconductor stress/warpage — is a legitimate one. That part is worth credit.\n\nThe evidence, though, does not hold up. The conclusion attributes the Pareto advantage to \"each recursive output contributing an individual loss term during BPTT,\" but no such loss is ever defined. Equations (2)–(5) give only forward recursions; there is no training objective, no BPTT description, no per-step loss anywhere in the methods or experiments. This is a load-bearing mechanism that appears only in the conclusion. That is not a minor omission; it undercuts the paper's headline claim.\n\nThe comparison is also confounded. M3 vs M7 differ in both recursion depth (T=3 vs 16) and the injected S(t)W_s term; M6 vs M7 also change block count. No ablation isolates depth injection or loss structure, so \"primary factor\" cannot be attributed. And all quantitative results live in prose descriptions of Figures 4 and 5 — no tables with MRR, Recall, FLOPs, parameters, or error bars, no training details, no code or data. The reader cannot verify the Pareto claims. The choice of MRR/Recall rather than direct regression error on stress/warpage fields is defensible but not justified.\n\nThe reader's stress-test note is correct here: the per-step loss never appears in the technical sections. This is the central flaw.\n\nStill, the paper deserves a serious referee. The architecture is clearly specified, the subfield could benefit from a rigorous comparison of recursive weight-sharing variants, and the idea is plausible. It needs major revision: report actual numbers, add ablations, define the training loss, and either add a controlled baseline or soften the causality claim. I would not cite it in its current form.","headline":"A clearly specified recursive transformer variant, but the supporting evidence is too thin and the central causal claim is unsupported by the paper's own technical content.","tokens_in":6276,"tokens_out":3340,"would_cite":false,"duration_ms":37415,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A depth-recursive transformer that injects the depth coordinate as a state input and applies a loss at every recursion step achieves the most favorable accuracy-per-parameter and accuracy-per-FLOP trade-offs for small-data thermo-mechanical","keywords":["recursive transformer","weight sharing","surrogate model","thermo-mechanical reliability","semiconductor packaging","small data","parameter efficiency","FLOPs"],"falsifier":"A controlled experiment that matches SIMPLE and DEPTH on everything except the depth-state injection and per-step losses (same block count, normalization, and recursion depth) and measures whether DEPTH still wins; or measuring direct regression error on stress and warpage fields instead of retrieval scores.","tokens_in":5291,"feed_emoji":"🔁","tokens_out":3289,"duration_ms":30882,"temperature":0.7,"pith_summary":"The paper argues that for small, low-dimensional engineering datasets—where generating simulation data is expensive—a recursive transformer that injects depth as an input state and applies a loss at each recursion step delivers the best accuracy per parameter and per FLOP among three recursive designs. It compares a tiny recursive model, a simple recursive transformer, and the proposed depth recursive model on two semiconductor package reliability tasks (stress, warpage) and a capacitor field prediction task. If correct, this suggests a practical rule: spend compute on recursion depth rather than on parameter capacity or latent state machinery when building surrogates for resource-constrained design.","feed_headline":"Depth-recursive transformer wins on accuracy per parameter","feed_subtitle":"For stress, warpage, and capacitance prediction, injecting depth as a recursive state gives the best accuracy-per-parameter trade-off.","key_machinery":"The central object is the Depth Recursive transformer (DEPTH), which reuses a single transformer block RECUR at every step t and adds a depth-dependent state S(t) scaled by a shared weight Ws to the previous hidden state before the block. With T steps, this creates a recursion whose outputs H(1)...H(T) each receive an individual loss term during backpropagation through time. The architecture keeps parameters constant as depth grows, trading additional compute for better accuracy on small datasets.","core_discovery":"The paper claims that the Depth Recursive model (M6/M7) achieves the strongest or near-strongest retrieval accuracy while using among the fewest parameters and lowest FLOPs evaluated, and that its design choices—explicitly injecting the depth coordinate as a recursively added state and giving each recursive output its own loss term during backpropagation through time—are the primary factor behind those trade-offs. On the Stress10k, Warpage10k, and PINN datasets, the Depth models sit on or near the Pareto frontier for both FLOPs and parameter count, while the dual-latent-state Tiny Recursive model lies off the frontier and the simple recursion is FLOP-efficient but accuracy-limited.","pith_inferences":["Because the compared models differ simultaneously in block count, normalization, recursion depth, and per-step loss, the paper's causal attribution of the gains to depth-state injection and per-step loss is not actually isolated; a matched ablation is needed to confirm.","The evaluation uses retrieval metrics (MRR/Recall) rather than direct regression error on the stress/warpage fields; a direct error comparison could change which architecture ranks best for engineering reliability.","The same depth-injection principle might extend to other PDE or multi-physics surrogate tasks where field quantities vary along a physical coordinate (e.g., through-thickness position), which the depth state S(t) represents."],"forward_implications":["On small engineering datasets, prefer architectures that deepen compute via weight-shared recursion over adding parameters.","The Depth Recursive model is a viable surrogate for package-level stress and warpage prediction and for Laplace-type field problems, with a compact memory footprint suitable for on-chip or edge inference.","Per-step loss terms during recursive training appear to improve learning in this low-data regime, a design choice other recursive transformers could adopt.","The Pareto analysis gives concrete guidance: for memory-bound settings, DEPTH dominates; for compute-bound settings, simple recursion can be competitive."],"fun_headline_variants":["Depth-recursive transformer tops accuracy-per-parameter","Recursive depth beats larger models on small data","Best accuracy per parameter: recursive depth wins","Small-data transformer: depth recursion is key","Weight-sharing transformer cuts FLOPs without losing accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim that depth-state injection and per-step loss are the primary cause of the accuracy gains assumes the compared architectures differ only in those aspects; in fact they also differ in block count, normalization, recursion depth, and input masking, so the causal attribution is not isolated.","fun_headline_variants_meta":{"raw":{"variants":["Depth-recursive transformer tops accuracy-per-parameter","Recursive depth beats larger models on small data","Best accuracy per parameter: recursive depth wins","Small-data transformer: depth recursion is key","Weight-sharing transformer cuts FLOPs without losing accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":996,"prompt_tokens":771,"completion_tokens":225,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":165}},"tokens_in":515,"tokens_out":225,"duration_ms":2914,"temperature":1.0,"reasoning_tokens":165,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:36:51.323866+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that matches SIMPLE and DEPTH on everything except the depth-state injection and per-step losses (same block count, normalization, and recursion depth) and measures whether DEPTH still wins; or measuring direct regression error on stress and warpage fields instead of retrieval scores.","supporting_citations":[],"review_version":1}