{"id":"89af1744-08b2-4571-b4fa-c2f80ceb3321","arxiv_id":"2505.16060","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-loop reverse-model method for test-time input optimization is claimed to find semiconductor etching recipes in five iterations, but the supporting evidence is simulation-only and not reproducible.","lead":"This paper proposes Model Feedback Learning, a method that optimizes inputs to a fixed machine or model by training a lightweight reverse model, without retraining the machine. The authors claim it generates semiconductor etching recipes in five iterations, beating Bayesian optimization and human experts, but all results are simulated and no code or data is provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed '5 iterations' contradicts the paper's own experimental setting (τ=200 in Table 10), undermining the central few-shot efficiency result.","rationale":"The reader's verdict identifies the Jacobian availability as the weakest assumption. That is a real deployment concern: Eq. (4) requires ∂M/∂x, and no gradient interface for physical hardware is described. However, within the paper's simulation setting M may be a differentiable emulator, so the Jacobian issue does not by itself refute the reported simulation numbers. The τ=200 vs '5 iterations' contradiction is more load-bearing because it attacks the truth of the reported efficiency result itself. The paper's own experimental parameters imply 200 machine-loop iterations, which is not 'few-shot' relative to the 84-iteration human baseline and directly contradicts the abstract and Sections 5.1-5.2. This is an internal inconsistency, not a matter of external consensus. The synthetic Gaussian dataset and borrowed human-expert numbers (from [18]) further weaken the empirical support, but the τ discrepancy alone is sufficient to reject the central claim in its current form. A code check or reproduction would settle it.","tokens_in":17656,"tokens_out":7941,"duration_ms":64353,"concrete_test":"Obtain the authors' code or run a reproduction using the exact parameters in Table 10 (T=1200, T0=1150, τ=200, τ0=150, α1=0.01, α2=0.0099) on the described Gaussian-sampled emulator, and record the Loop B iteration at which the outputs in Table 1 are first reached. If the effective machine-loop count is approximately 200 rather than 5, the 'five iterations' claim is refuted. If early stopping was used, the stopping criterion and per-experiment iteration counts must be reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MFL generates recipes in only 5 machine-loop iterations (Sec 5.1: 'only 5 iterations within the machine loop'; Sec 5.2: 'MFL requires only five iterations to generate recipes'). Algorithm 1's Loop B (lines 15-26) executes a fixed τ iterations with no early-stopping condition, and Appendix E Table 10 specifies τ=200, τ0=150, T=1200. Thus the configuration used to produce the reported results appears to run 200 machine-loop iterations, not 5. If τ=200, the comparison with the 84-iteration senior-engineer baseline is inverted (200 > 84), and the headline few-shot advantage disappears. The paper does not explain how '5 iterations' was measured or whether early stopping (not present in Algorithm 1) was applied. This is a direct internal contradiction in the evidence supporting the paper's primary contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces Model Feedback Learning (MFL), a test-time optimization framework that uses a lightweight reverse model R_theta to iteratively search for input values for a deployed machine or pre-trained model, without retraining the forward model. The method has two loops: Loop A pre-trains R_theta against an emulator E, and Loop B refines it against the machine M using gradients that include the machine's Jacobian. Conservative learning rates based on model sensitivity are proposed for stability. Experiments are reported on semiconductor plasma etching recipe generation, chemical vapor deposition, and wire bonding, with the headline claim that MFL generates etching recipes in only five machine-loop iterations, outperforming Bayesian optimization and human engineers (claimed 20 and 84 iterations, respectively). The paper also compares against supervised learning and a random-search baseline, and reports robustness and ablation studies.","tokens_in":17824,"tokens_out":3572,"duration_ms":32821,"significance":"If the empirical claims were fully supported, MFL could be a practically valuable tool for input-recipe optimization in settings where deployed equipment cannot be modified. The paper has real strengths: the convergence result in Appendix B is a standard and correct gradient-descent argument under an L-smoothness assumption, and the algorithm is coherent when run against a differentiable emulator. The appendices provide detailed experimental settings for the etching, CVD, and wire-bonding cases, and the comparison against the LSRS-LR baseline (Algorithm 2 and Table 7) is a concrete, reproducible-looking experiment. However, the paper's central claimed advantage, the 'five iterations' efficiency result, is not supported by the experimental configuration: Algorithm 1 runs a fixed tau iterations in Loop B with no early-stopping condition, and Appendix E sets tau=200. In addition, the deployment mechanism relies on the Jacobian of the machine model without specifying how it is obtained. These are load-bearing issues, not presentation concerns.","major_comments":[{"comment":"The central claim that 'MFL requires only five iterations to generate recipes' is contradicted by the paper's own experimental configuration. Algorithm 1's Loop B runs a fixed τ iterations with no early-stopping condition, and Table 10 sets τ=200 (with τ0=150 and T=1200). The text nowhere explains how '5 iterations' is measured, whether early stopping is applied (it is not in Algorithm 1), or why the reported results would correspond to 5 instead of 200 machine-loop iterations. If Loop B actually ran 200 iterations, the comparison with the 84-iteration senior-engineer baseline in Section 5.2 is inverted (200 > 84), and the headline few-shot efficiency advantage disappears. This needs a concrete, quantitative explanation: a modified algorithm with an explicit stopping rule, a reporting of the actual number of machine evaluations used for Tables 1–4, or a clear distinction between 'iterations to convergence' and 'fixed loop budget.' As written, the paper's primary empirical claim is not derivable from the presented algorithm and settings.","section":"§5.1, §5.2 vs. Appendix E (Table 10) and Algorithm 1 (Lines 15–26)"},{"comment":"Loop B updates the reverse model using gradients that contain the Jacobian ∂M(x)/∂x of the deployed machine model. The paper does not describe how this Jacobian is obtained for real hardware: there is no finite-difference scheme, no surrogate-gradient procedure, no sensor interface, and no discussion of noisy or black-box machine access. If a real plasma etcher cannot supply ∂M(x)/∂x, the machine loop as written cannot run, and the 'five iterations within the machine loop' deployment claim has no mechanism. The authors should either specify a practical way to compute or approximate these Jacobians (e.g., by calibrated finite differences on the physical system, with a cost/iteration budget) or replace the machine Jacobian in Loop B with a computable surrogate and justify that substitution.","section":"§4.1, Eq. (4), and Algorithm 1 (Lines 16–25)"},{"comment":"The claimed superiority over Bayesian optimization is not supported by any reported experiment. The text states that 'Bayesian optimization without human efforts... is unable to generate correct recipes,' but gives no data, no BO configuration, no number of trials, and no evaluation protocol. The Lam Research comparison is taken from [18] rather than measured under the same setup, so the '5 vs. 20 vs. 84 iterations' comparison is not a controlled study. To support the efficiency claim, the paper needs a BO baseline run on the same emulator/machine with the same target ranges and a reported distribution of iteration counts, not just a sentence saying BO fails.","section":"§5.2 (Bayesian optimization comparison)"},{"comment":"The accuracy results are presented as a single set of output values with no standard deviation, no number of seeds, and no repetition. Given that Figure 5 presents MFL as a curve over epochs and the robustness section claims low error under noise, the paper should report mean ± std over multiple seeds for both the final recipe error and the number of iterations required. Without this, 'achieving accurate results' is an anecdotal observation from one run, and it is impossible to assess whether the method reliably meets the semiconductor targets.","section":"§5.1, Table 1, and §5.4"}],"minor_comments":[{"comment":"The 'Process time [s]' row contains two conflicting values, '5.05' and '144.5516', which appears to be a formatting error; please correct it.","section":"Table 4"},{"comment":"The statement 'requiring only 5 iterations to converge for CVD and 9 for wire bonding' is not accompanied by a formal convergence criterion; define explicitly what counts as an iteration and what stopping threshold is used.","section":"§5.5 and abstract"},{"comment":"The notation is confusing: z' is used both for the target random variable and for the vector of targets in Eq. (2); please use consistent notation (e.g., Z' for the random variable and z' for its realization).","section":"§3 and Eq. (2)"},{"comment":"The y-axis label 'Error' has no units; indicate whether this is mean-squared error, L2 norm, or normalized error.","section":"Figure 9"},{"comment":"The 'supervised learning approach' baseline is attributed to [24, 34], but [34] is a plasma etching review rather than a supervised-learning recipe-generation method; please cite a more appropriate baseline or clarify the relationship.","section":"§5.3"},{"comment":"Line 13 says 'Update x with learning rate η' but does not specify the optimizer; the text mentions Adam, so either state that in the algorithm or use a concrete update rule.","section":"Algorithm 2"}],"recommendation":"reject","confidential_remarks":"The stress-test concern about τ=200 vs. the claimed '5 iterations' lands squarely: it is an internal contradiction between the algorithm, the experimental settings, and the headline claim. The Jacobian-availability issue is also real and structural. I would not send this back for minor fixes; the authors would need to either supply a genuine early-stopping protocol with actual machine-loop counts, or acknowledge that Loop B runs 200 iterations, which would invalidate the central efficiency comparison. The convergence appendix is sound and the emulator-based framework is coherent, but the load-bearing empirical claims are not supported as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the interesting thing here is the two-loop scheme: pre-train a reverse model on an emulator, then refine on the machine. That is a legitimate extension of inverse neural networks and internal model control, and the framing—test-time optimization of inputs without retraining a deployed model—is real. The conservative learning-rate rule and the sensitivity threshold are reasonable additions, and the convergence appendix is textbook but correct.\n\nBut the paper's central empirical claim does not survive contact with its own appendix. Section 5 claims recipe generation in 'only five iterations' of the machine loop, but Algorithm 1 has no early stopping, and Table 10 sets τ=200. So the configuration that produced the reported results runs 200 machine-loop iterations, not 5. If that is right, the comparison to the 84-iteration senior engineer is inverted, and the few-shot headline collapses. The paper never explains how '5' was measured. That is a direct internal contradiction in the evidence for the main contribution.\n\nBeyond that, the method assumes access to ∂M/∂x for the deployed machine (Eq. 4 with E replaced by M). No finite-difference scheme, surrogate gradient, or sensor interface is described. If a real etcher cannot supply gradients, Loop B cannot run as written. That is a structural gap, not a detail.\n\nThe experiments are all on a synthetic Gaussian-sampled emulator; no code or data are shipped; the human-expert baseline is borrowed from Kanarik et al. So the empirical support is thin, and the claims go beyond what is shown.\n\nCredit where due: the idea is plausible, the formulation is clear, and the convergence proof is fine for what it is. But the load-bearing efficiency claim is contradicted by the paper's own settings, and the deployment story has a missing Jacobian. This looks like a promising method in search of an honest evaluation. With real-machine or properly benchmarked simulation results, it could be a solid contribution. As it stands, the paper needs major revision before I'd trust the headline numbers.\n\nRecommendation: I would not take the current version seriously as a citation, but I think it deserves a serious referee—the flaws are concrete and fixable, and the core idea is worth engaging. If it comes to you, send it to review with the explicit request to check the iteration count and the gradient assumption.","headline":"A plausible two-loop inverse-model idea whose central '5 iterations' claim is directly contradicted by its own Table 10 (τ=200), and whose machine-loop gradient assumption is unaddressed.","tokens_in":18362,"tokens_out":1847,"would_cite":false,"duration_ms":16046,"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 five-iteration feedback loop tunes a fixed machine's inputs to new targets without retraining.","keywords":["model feedback learning","test-time optimization","recipe generation","plasma etching","semiconductor manufacturing","inverse model","few-shot adaptation","chemical vapor deposition"],"falsifier":"A real test would be to run Loop B on a physical plasma etcher that returns only scalar measurements and provides no input-output gradient. If the recipe cannot reach the target ranges in about five iterations under those conditions, the paper's core efficiency claim collapses.","tokens_in":17433,"feed_emoji":"⚙️","tokens_out":7429,"duration_ms":61762,"temperature":0.7,"pith_summary":"The paper proposes Model Feedback Learning (MFL), a way to make a deployed machine or pre-trained model hit new output targets by changing only its inputs, never its parameters. MFL trains a small reverse network that maps a desired target to the input that produces it, using feedback from a cheap emulator first and from the real machine second. In semiconductor plasma-etching recipe generation the method reaches the target ranges in five machine-loop iterations, compared with at least twenty for a published human-machine method and eighty-four for a senior engineer. It also reports five iterations for chemical vapor deposition and nine for wire bonding. If correct, MFL offers a retraining-free route to re-target costly deployed equipment when process requirements change.","feed_headline":"Five machine queries tune a semiconductor recipe to spec","feed_subtitle":"A lightweight reverse model adapts fixed machine inputs to new targets, beating 20-step and 84-step human baselines.","key_machinery":"The central object is the reverse model $R_\\theta$, a small neural network (about 7 kB) that maps a desired target output $z'$ to a candidate input $x' = R_\\theta(z')$. The mechanism is gradient feedback through a fixed forward model: each iteration computes the output mismatch $z' - E(R_\\theta(z'))$ or $z' - M(R_\\theta(z'))$ and back-propagates it through the forward model's Jacobian $\\partial E/\\partial x$ (Loop A) or $\\partial M/\\partial x$ (Loop B) to update $\\theta$. The stability-aware part is a sensitivity check $s_E(x)$ or $s_M(x)$ that switches the learning rate from $\\alpha_1$ to the smaller $\\alpha_2$ when the forward model is too sensitive at the current input. This division of labor makes few-shot adaptation possible: Loop A does the bulk learning cheaply, and Loop B needs only a handful of real-machine evaluations.","core_discovery":"The central claim is that input optimization can be solved by learning the inverse map $R_\\theta$ from targets to inputs and refining it in two loops. Loop A pre-trains $R_\\theta$ against an emulator $E$, a supervised approximation of the machine, so that most of the learning happens cheaply. Loop B then refines $R_\\theta$ against the real machine model $M$, using the gradient of $M$'s outputs with respect to its inputs; a conservative learning-rate rule that switches to a smaller step when model sensitivity is high keeps the updates stable. The paper's empirical headline is that in plasma etching, five machine-loop iterations produce a recipe with etch depth 2255.55 nm, etch rate 109.9 nm/min, mask remaining 358.95 nm, top CD 198.80 nm, $\\Delta$CD 10.04, and bow CD 198.52 nm, all inside the target windows. The same few-shot pattern appears in CVD (5 iterations) and wire bonding (9 iterations).","pith_inferences":["Editorial inference: the method only needs a differentiable forward model, so any simulator or digital twin with accessible gradients could be re-targeted this way; the paper demonstrates this on a simulator, not on live fab tools.","Editorial inference: the five-iteration count covers Loop B only; the emulator still requires supervised training data, so the full deployment cost depends on how that emulator is built and validated.","Editorial inference: a natural stress test is to withhold the machine Jacobian and see whether finite-difference or learned-gradient variants preserve the five-iteration behavior; the paper does not address that regime."],"forward_implications":["Deployed semiconductor equipment could be re-targeted to new specifications in roughly five machine evaluations, without retraining or hardware changes.","Process development would need far fewer physical trials than the current human-driven practice, reducing time and cost while keeping recipes inside domain constraints.","The same two-loop inverse-learning procedure transfers to other manufacturing steps, with five iterations reported for chemical vapor deposition and nine for wire bonding.","Stability-aware conservative updates make the optimization robust to target shifts and attack noise, which is relevant when process conditions drift in production."],"supporting_citations":[{"why":"Supplies the semiconductor etching dataset, target windows, human input ranges, and the published human-machine baseline (20 iterations) and senior-engineer baseline (84 iterations) that MFL is measured against.","marker":"[18]"},{"why":"Defines Bayesian optimization, the standard black-box optimizer whose iteration count MFL is compared with.","marker":"[9]"},{"why":"Positions supervised learning as the standard data-driven recipe-generation approach that MFL is compared against.","marker":"[8]"},{"why":"Is the inverse-neural-network method MFL is designed to improve on by avoiding retraining of the original model.","marker":"[3]"},{"why":"Is internal model control, the control framework MFL distinguishes itself from because it needs no exact forward model.","marker":"[11]"},{"why":"Is the test-time training baseline that MFL contrasts with because it typically needs many samples rather than few-shot feedback.","marker":"[45]"},{"why":"Supplies the chemical vapor deposition process background used to define the second application domain.","marker":"[16]"},{"why":"Supplies the wire bonding process background used to define the third application domain.","marker":"[6]"}],"fun_headline_variants":["No retraining, five queries tune recipe to spec","Five test-time queries beat human experts on etch recipe","Reverse model fine-tunes inputs in five machine loops","Few-shot input search hits semiconductor targets fast","Model feedback learning nails recipe in five iterations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The five-iteration result depends on getting a gradient of the real machine's output with respect to its input settings during Loop B; the paper assumes that gradient exists and is accessible, but does not explain how a physical etcher would provide it.","fun_headline_variants_meta":{"raw":{"variants":["No retraining, five queries tune recipe to spec","Five test-time queries beat human experts on etch recipe","Reverse model fine-tunes inputs in five machine loops","Few-shot input search hits semiconductor targets fast","Model feedback learning nails recipe in five iterations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1387,"prompt_tokens":958,"completion_tokens":429,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":357}},"tokens_in":574,"tokens_out":429,"duration_ms":4606,"temperature":1.0,"reasoning_tokens":357,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:07:43.237980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A real test would be to run Loop B on a physical plasma etcher that returns only scalar measurements and provides no input-output gradient. If the recipe cannot reach the target ranges in about five iterations under those conditions, the paper's core efficiency claim collapses.","supporting_citations":[{"cited_title":"Human–machine collaboration for improving semiconductor process development.Nature, 616(7958):707–711, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the semiconductor etching dataset, target windows, human input ranges, and the published human-machine baseline (20 iterations) and senior-engineer baseline (84 iterations) that MFL is measured against."},{"cited_title":"Exploring machine learning for semiconductor process optimization: a systematic review.IEEE Transactions on Artificial Intelligence, 2024","cited_arxiv_id":null,"evidence_quote":"Positions supervised learning as the standard data-driven recipe-generation approach that MFL is compared against."},{"cited_title":"Analyzing inverse problems with invertible neural networks","cited_arxiv_id":null,"evidence_quote":"Is the inverse-neural-network method MFL is designed to improve on by avoiding retraining of the original model."},{"cited_title":"Internal model control","cited_arxiv_id":null,"evidence_quote":"Is internal model control, the control framework MFL distinguishes itself from because it needs no exact forward model."},{"cited_title":"Royal society of chemistry, 2009","cited_arxiv_id":null,"evidence_quote":"Supplies the chemical vapor deposition process background used to define the second application domain."},{"cited_title":"Springer, 2014","cited_arxiv_id":null,"evidence_quote":"Supplies the wire bonding process background used to define the third application domain."}],"review_version":1}