{"id":"10fe3924-1ca9-4163-bc38-3a6fd4817f5d","arxiv_id":"2607.15313","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper argues that probabilistic scaling alone cannot fix the validity gap in quantum circuit generation, so quantum code assistants must build verification into generation rather than filter outputs after the fact.","lead":"This position paper argues that scaling up large language models cannot reliably produce correct quantum circuits, because valid circuits are exponentially rare and post-hoc filtering is too expensive. It proposes shifting from human-centered copilots to verifier-centered generation, where constraints are enforced during generation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim is gated by an assumed-into-existence bound on Cθ; without evidence that scaling cannot shrink Cθ for OOD targets, the exponential gap of Eq. (1) is a conditional conjecture, not a demonstrated barrier.","rationale":"The reader's weakest_assumption correctly identifies the Cθ non-pathology condition as a key weak point. The paper itself flags this in footnote 1, explicitly conceding that in-distribution targets can cancel the exponential decay. This is the single most load-bearing concern because Eq. (1) is the quantitative foundation for the coupled exponential gap in Eq. (2), and without a bound on Cθ that holds uniformly with model scale, the claim that 'scale alone cannot bridge the validity gap' is not established. The Proposition B.1 amplification error is a real technical flaw but secondary: even without the FTQC structural-sparsity term, the NISQ functional-sparsity argument would still imply exponential cost if Eq. (1) were valid. The Proposition C.2 no-op-padding count is also a proof gap, but it may be fixable and does not by itself undercut the qualitative thesis. The paper should be read as a well-motivated position statement with a conditional, not unconditional, impossibility claim. I recommend keeping the reader's CONDITIONAL verdict unchanged, with conditions focused on scoping the Cθ assumption and correcting the cited proofs.","tokens_in":23665,"tokens_out":8132,"duration_ms":74639,"concrete_test":"Empirical scaling check: Train a family of decoder-only transformers of varying sizes (e.g., 10M–1B parameters) on a fixed corpus of verified QASM circuits for n=10–20 qubits. For a held-out set of target unitaries generated from a different distribution (e.g., random circuits with distinct gate sets) at n=20,30,40, estimate the conditional success rate after structural filtering, and compute Cθ = P_success · e^{γn} using the paper's γ. If Cθ decreases with model size or success rate does not decay as e^{-γn}, the Cθ=O(poly(n)) assumption fails and the central claim is undermined; if Cθ stays polynomially bounded with scale, the non-pathology assumption is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Appendix C.2 / Eq. (1)'s non-pathology assumption Cθ=O(poly(n)). The paper's footnote 1 states the bound is 'informative only when Cθ=O(poly(n)), which holds by definition of generalization for out-of-distribution targets.' This is definitional, not demonstrated. If a larger model places more probability mass near the target unitary for OOD tasks, Cθ = ||dπθ/dμ||∞ can be much larger than poly(n) and can cancel e^{-γn}; the paper explicitly concedes this for in-distribution algorithmic primitives (QFT, Grover, modular arithmetic). Nothing in the argument prevents scaling from producing the same cancellation for OOD targets—the inverse-scaling hypothesis is speculative, and the authors' own empirical support (Song et al., 2026) is a self-cited workshop paper reporting a ceiling, not a scaling-law analysis. POSITION2's conclusion—that 'probabilistic scaling cannot close the coupled exponential gap'—is the paper's central claim, and it depends entirely on Cθ remaining polynomial in n for arbitrary model scale. The impossibility result therefore rests on an untested assumption rather than a proven mathematical theorem. This is a fixable gap for a position paper, but the claim as stated should be scoped as a conditional conjecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper argues that standard probabilistic scaling of large language models is the wrong paradigm for generic quantum circuit synthesis. It identifies a syntax-semantics gap, claims that training on unverified quantum code biases models away from physical validity, and argues that post-hoc verification is intractable because valid circuits are exponentially sparse. The paper formalizes a two-layer validity notion (structural/functional) and derives a coupled exponential cost for generate-then-filter pipelines (Eq. (2)), then proposes verifier-centric agents that enforce validity during generation via hierarchical constraints, validity masks, and trace-based data. Supporting appendices contain formal sparsity propositions for NISQ and FTQC, a structural-hardness claim, and a worked Cuccaro adder example.","tokens_in":23955,"tokens_out":12514,"duration_ms":112829,"significance":"The paper addresses an important and timely question: whether LLM scaling alone can produce correct quantum programs. The qualitative thesis is plausible and aligns with recent evidence in constrained domains such as formal mathematics and planning. The proposal to make verification a generation invariant rather than a post-filter is a useful contribution for quantum software and AI4Science. The paper is also commendably explicit about several assumptions (footnote 1, Appendix C.2 remarks), and the hierarchical multi-level abstraction framework in Section 4 is concrete. However, the central impossibility claim is currently gated by assumptions that are not established, and the structural-hardness argument contains a technical error. With those repaired or explicitly scoped, the paper could be a strong position statement. No machine-checked proofs or reproducible artifacts are provided; the empirical anchor is a self-cited workshop paper.","major_comments":[{"comment":"The amplification step is incorrect. The proof supposes P(x∈M_struct^FTQC)≥2^{-cn} and claims that 'poly(n) independent repetitions' followed by poly(n)-time verification yields constant success probability, contradicting NP⊈BPP. In fact, poly(n) repetitions of a 2^{-cn} event succeed with probability poly(n)·2^{-cn}=e^{-Ω(n)}, still exponentially small; constant success requires 2^{cn} repetitions. The earlier ETH argument is also wrong: 2^{cn}=2^{Θ(m)} expected trials is not o(2^m) for any constant c>0, so no ETH contradiction follows. The reduction step asserting a poly-time bijection between structurally valid FTQC descriptions and satisfying assignments is asserted without proof. Because Eq. (4) and POSITION2 rely on p_struct^FTQC≤e^{-βn}, this is a load-bearing gap. Please correct the statement (e.g., conditional on a sampler achieving 1/poly(n) mass) or provide a valid reduction/a","section":"Appendix B.3, Proposition B.1"},{"comment":"The functional-sparsity bound is informative only when Cθ=O(poly(n)), which the paper takes as holding 'by definition of generalization for out-of-distribution targets.' This is an assumption, not a demonstrated property, and it is the gate for the paper's central negative claim. The paper concedes that for in-distribution algorithmic primitives Cθ can be large enough to cancel e^{-γn}; nothing in the argument rules out a scaled model moving currently-OOD targets into its effective support and giving Cθ=e^{Ω(n)}. The cited empirical evidence (Song et al., 2026) is a self-cited workshop paper reporting a 79% ceiling, not a scaling-law study showing the ceiling persists with model size. POSITION2 should therefore be stated as a conditional conjecture under the non-memorization condition Cθ=O(poly(n)), or the paper should supply a scaling-dependent bound on Cθ and a concrete falsifiable exp","section":"Section 3.2 / Eq. (1) / footnote 1"},{"comment":"The proof of Proposition C.2 does not establish the claimed upper bound on the number of length-G programs that ε-approximate U_target. The bound poly(n,log(1/ε))·|I|^{G-G*} assumes every such program is a no-op/rewrite extension of a canonical length-G* program; an arbitrary length-G program approximating U_target need not contain such a subprogram, and there can be exponentially many distinct programs approximating the identity that could be appended. Without a rigorous counting argument, the numerator bound and hence γ_FTQC=Ω(d log n) are not proven. This is a second load-bearing gap in the formal core of POSITION2. Please provide a complete argument or label Proposition C.2 as a heuristic scaling estimate.","section":"Appendix C.2.2, Eq. (19)"}],"minor_comments":[{"comment":"The Pauli-stabilizer evolution appears incorrect: starting from ZII and applying CNOT(y,x), CNOT(x,temp), CNOT(y,x) gives ZZZ, not ZIZ as stated. The qualitative conclusion that temp is entangled still holds, but the displayed calculation should be corrected.","section":"Appendix A"},{"comment":"The phrase 'post-hoc filtering is mathematically intractable' is stronger than the formal results, which are conditional on Cθ=O(poly(n)) and on the corrected structural-hardness argument. Consider qualifying the abstract and POSITION2 accordingly.","section":"Abstract / Section 3.3"},{"comment":"The assumption that U_target admits a polynomial-length canonical description is stated only in the setup, not in Proposition C.2. State it in the proposition statement.","section":"Appendix C.2.2"},{"comment":"The key empirical ceiling (79%) is cited to the authors' own workshop paper (Song et al., 2026). Please provide more experimental detail (model sizes, tasks, error bars) or cite independent evidence, since this is a primary motivation for the inverse-scaling hypothesis.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central empirical data point is the authors' own AAAI 2026 workshop paper (Song et al., 2026). For a journal position paper, the 79% ceiling claim should be supported by a peer-reviewed or independently reproduced scaling study; otherwise POSITION2 rests on a self-citation. Also, the paper's scope (generic synthesis vs. modular primitives) should be clearer in the abstract. The formal gaps in Propositions B.1 and C.2 are fixable, but they are load-bearing for the stated impossibility claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this if you care about LLMs for quantum code. The paper argues that scaling LLMs on unverified QASM won't converge to physically valid circuits; you need verifiers in the generation loop. That general message is right and consistent with the constrained-decoding and solver-in-the-loop literature. The new content is the formalization: an MDP-style verifier-centric agent, multi-level abstraction, and sparsity bounds for NISQ/FTQC functional validity.\n\nThe best parts are concrete. Section 4's workflow — modules, decomposition, property-preserving rewrites — is clear, and the Cuccaro adder example actually works. The distinction between cheap structural checking in NISQ and NP-hard structural feasibility in FTQC is useful. The paper also scopes its own result in footnote 1: for in-distribution primitives (QFT, modular arithmetic), Cθ can cancel e^{-γn}, so the hard cases are out-of-distribution targets. That is the right way to read the paper.\n\nThe soft spots are real. Proposition B.1's proof says poly(n) independent repetitions amplify a 2^{-cn} success probability to constant. That is false. You need ~2^{cn} repetitions, so the \"or NP⊆BPP\" conclusion is not supported as written. The expected-cost version can be salvaged with ETH for c<1, but not the unconditional claim. Proposition C.2's numerator bound assumes the number of ε-approximating programs is only poly(n) times no-op paddings; the paper doesn't justify why many genuinely different approximating programs can't exist. And the whole functional-sparsity bound depends on Cθ=O(poly(n)) for OOD targets, which is assumed, not demonstrated. Since scaling could in principle reduce that density ratio, the paper's central statement — \"probabilistic scaling cannot close the coupled exponential gap\" — is a well-motivated conjecture, not a theorem.\n\nThe authors lean on their own workshop paper for the 79% ceiling, which is a weak empirical anchor. But this is a position paper; the lack of new experiments shouldn't be held against it.\n\nMy take: engage with it as a position. Fix B.1, scope C.2, and present POSITION2 as conditional on Cθ. A serious referee should see it. I'd bring it to reading group and would cite it with the caveat.","headline":"A well-argued position for verifier-centric quantum circuit generation; the qualitative thesis is sound, but the formal hardness claims are conditional and one proof is wrong as written.","tokens_in":24471,"tokens_out":2406,"would_cite":true,"duration_ms":24974,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum program generation must prioritize validity over scale: post-hoc filtering cannot close the exponential gap.","keywords":["quantum program generation","scaling hypothesis","inverse scaling","circuit synthesis","verifier-centric generation","syntax-semantics gap","functional correctness","fault-tolerant quantum compilation"],"falsifier":"Benchmark a model on random n-qubit target unitaries (n up to about 20) that are provably absent from training data, and measure the conditional functional pass rate as n grows. The paper predicts exponential decay e^{-γn}; observing polynomial decay or a constant pass rate would refute the central claim. Also estimate the model's probability mass on the exact-solution region: if this mass grows exponentially with n even for random targets, the non-pathology assumption Cθ = O(poly(n)) fails.","tokens_in":23452,"feed_emoji":"⚛️","tokens_out":6278,"duration_ms":60482,"temperature":0.7,"pith_summary":"The paper argues that the scaling hypothesis—more parameters yield emergent reasoning—fails for generic quantum circuit synthesis. Because physically valid circuits are exponentially rare among all syntactically parseable circuits, generating many candidates and then filtering out invalid ones has an expected cost that grows exponentially in qubit count, no matter how large the model is. Training on unverified quantum code teaches syntax without Hilbert-space semantics, so larger models can become more confidently wrong. The paper's proposed remedy is verifier-centric generation: build formal checks for structural and functional validity into the decoding process itself, via a masked policy that only permits actions that preserve correctness. If the argument holds, progress in quantum program generation depends on coupling models to verifiers and curated verified data, not on continued scaling.","feed_headline":"Scaling alone won't close the quantum code validity gap","feed_subtitle":"Valid quantum circuits are exponentially rare, so generation must build verification in from the start.","key_machinery":"The central object is the two-layer validity decomposition into structural validity Mstruct and functional correctness Mfunc, combined in the coupled-exponential post-selection cost identity E[C] = Cstruct/(pstruct·pfunc|struct) + Cfunc/pfunc|struct, where Cfunc = Θ(2^n), Cstruct = Θ(nd), and pfunc|struct ≤ Cθ·e^{-γn}. The companion mechanism is the constrained policy πθ(a|s) ∝ πθ^(0)(a|s)·V(a,s), a validity mask over a Markov decision process over circuit states, which turns the verifier from an output gate into a generation invariant.","core_discovery":"The paper's central claim is that post-hoc verification—generate candidates, then discard those that fail to compile or fail to implement the target unitary—is exponentially intractable for generic synthesis. Validity is split into two layers: structural validity (obeying hardware constraints) and functional correctness (implementing the target unitary within tolerance). The expected cost of obtaining one valid candidate follows a coupled-exponential identity: the structural verification cost is divided by the product of the structural and functional pass rates, while the functional verification cost, which is exponential in qubit count, is divided only by the functional pass rate. Because t","pith_inferences":["A testable extension is to measure the conditional pass rate on a family of random n-qubit target unitaries not in the training set; the paper's bound predicts exponential decay with an exponent tracking depth and log(1/ε) or log n, checkable by state-vector verification up to about 20 qubits.","The coupled-exponential argument generalizes to other engineering-science domains where valid outputs are exponentially rare and functional verification is exponential—protein design, theorem proving, and chip layout are analogous, and the same verifier-in-the-loop recipe should apply there.","The sharpest consequence the paper leaves implicit is that scale investments in quantum-code language models only pay off when paired with verification-aware architectures; absent that, continued scaling is predicted to be actively harmful rather than merely flat.","The framework implies benchmark design should shift from token-level accuracy to verified-synthesis pass rates at a fixed verification budget, a shift that would make the paper's central claim directly measurable across different model families."],"forward_implications":["For generic synthesis, larger models trained on unverified code will not converge to physical validity; at best they produce more fluent but still invalid circuits, a predicted inverse-scaling regime.","Generate-then-verify pipelines have expected cost exponential in the logical qubit count (e^{(ln2 + γ)n} for NISQ, e^{(max(ln2,β)+γ)n} for FTQC), so post-hoc filtering is not a viable route to reliable quantum code.","Verification must be moved inside generation: masks, hierarchical module constraints, and property-preserving rewrites replace sampling-and-discarding.","Verified, trace-based training data becomes a primary design component, not a curation afterthought; the agent's target audience shifts from human reviewer to formal verifier.","The validity gap is independent of physical substrate: NISQ and FTQC both face the same exponential functional-sparsity barrier because the logical Hilbert-space dimension is what matters."],"fun_headline_variants":["Quantum code needs built-in verification, not filtering","Post-hoc quantum filtering: exponentially hopeless","For quantum code, bake verification into generation","Exponentially rare valid circuits: build verification in","Verification must be part of quantum code generation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The exponential-wall argument assumes the model has not memorized the specific target circuit; for textbook primitives like quantum Fourier transform or modular arithmetic that assumption fails, and in those cases scale plus data can still work.","fun_headline_variants_meta":{"raw":{"variants":["Quantum code needs built-in verification, not filtering","Post-hoc quantum filtering: exponentially hopeless","For quantum code, bake verification into generation","Exponentially rare valid circuits: build verification in","Verification must be part of quantum code generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001876,"raw_usage":{"total_tokens":7153,"prompt_tokens":655,"completion_tokens":6498,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":399,"completion_tokens_details":{"reasoning_tokens":6429}},"tokens_in":399,"tokens_out":6498,"duration_ms":40899,"temperature":1.0,"reasoning_tokens":6429,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T02:14:42.768690+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Benchmark a model on random n-qubit target unitaries (n up to about 20) that are provably absent from training data, and measure the conditional functional pass rate as n grows. The paper predicts exponential decay e^{-γn}; observing polynomial decay or a constant pass rate would refute the central claim. Also estimate the model's probability mass on the exact-solution region: if this mass grows exponentially with n even for random targets, the non-pathology assumption Cθ = O(poly(n)) fails.","supporting_citations":[],"review_version":1}