{"id":"6e076900-dd49-401f-83f0-603a5238c268","arxiv_id":"2504.17116","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"OneAdapt introduces a resource-adaptive compilation approach for photonic one-way quantum computing, using dynamic node refresh and skewed temporal edges to reduce hardware size and execution depth.","lead":"A new compiler design helps photonic quantum computers run programs with fewer hardware resources and shorter execution time. It changes how quantum programs are mapped onto the entangled states these machines generate, and reports large efficiency gains in simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The depth/size gains rest on Sec 4.5's claim that Hamming-distance-1 skewed temporal edges are nearly free, but the supporting evidence is a randomized PL-ratio simulation and the authors' own compiler, not a test on the actual skew patterns or hardware.","rationale":"The reader's weakest assumption and my read converge on the same load-bearing point: the new IR's skewed edges are only useful if their physical realization is nearly free. The missing code/data and the 'minimizes' overclaim are real issues, but they are addressable by release and wording changes; they do not attack the central mechanism. The PL-ratio evidence is the only direct support for the key hardware abstraction, and it is not yet a faithful test of the compiler's output: Fig 10(a) randomizes skew edges while Sec 4.4's heuristics produce correlated, direction-dependent patterns, and Fig 10(b) does not exercise the extreme 3x3/D_f=10 regime highlighted in the headline. A simulation on exact compiler outputs settles whether the physical resource claim holds, without requiring hardware. Since this concern reinforces the conditional verdict rather than overturning it, the reader's CONDITIONAL assessment remains appropriate.","tokens_in":22127,"tokens_out":6007,"duration_ms":57287,"concrete_test":"Recompute the Fig 10 PL-ratio simulation using the exact skewed-edge sets emitted by OneAdapt for the benchmark configurations in Tables 1-2 (e.g., VQE-64 at S=3x3 with D_f=10, and QAOA-100 at standard size), with the same 75% fusion success rate over at least 100 trials. If the measured PL ratio exceeds the FlexLattice baseline of ~3.1 by more than about 25%, or the resulting RSL count changes materially, the 'negligible overhead' assumption is unsupported and the physical-resource claims need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"OneAdapt's headline results (3.68x vs OnePerc, 3.56x vs Qiskit, 8x8 to 3x3 at D_f=10) depend on the 2D-bounded temporal routing pass in Sec 3.2/4.4, which is only valid if skewed temporal edges are realizable with negligible overhead. Section 4.5 asserts that realization requires 'little modification' and restricts skew to Hamming distance 1, citing Fig 10. However, Fig 10(a) is a preliminary random-edge experiment: 10-40% of all possible skewed edges are chosen and skewed in random directions. This is not the edge distribution produced by the compiler's heuristics, which minimize distances or maximize free space and can generate dense, directionally correlated skew patterns. Fig 10(b) compiles benchmarks, but only at 6x6, and reports PL ratio without a hardware-level routing scheme. In the advertised extreme regime (3x3, D_f=10), path conflicts among skewed edges are likely more severe than in the randomized experiment, so the PL ratio may exceed the FlexLattice value of ~3.1. Additionally, skewed edges may require extra delay lines or switches that the IR-level abstraction does not model, and no physical demonstration is provided. If the PL ratio or hardware overhead rises significantly, the logical-depth reductions do not translate into the claimed physical resource savings, undermining the central 'resource-constrained' contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces OneAdapt, an MBQC compiler for photonic one-way quantum computing. It proposes an intermediate representation that extends the FlexLattice IR by (i) bounding the length of temporal edges via dynamic node refresh and (ii) permitting skewed temporal edges within Hamming distance 1, exploited by a 2D-bounded temporal routing pass. The evaluation reports, relative to the OnePerc FlexLattice compiler, a 3.68x average reduction in 1D depth with bounded temporal-edge lengths, and relative to a Qiskit-adapted cluster-state baseline a 3.56x reduction; it also reports a reduction of the required 2D size for 64-qubit programs from 8x8 to 3x3 at D_f=10, and a 2.87x depth reduction in a surface-code FTQC setting. The central contribution is a resource-adaptive IR and associated compiler passes, with the physical realizability of skewed edges as the key enabling assumption.","tokens_in":22292,"tokens_out":5944,"duration_ms":54471,"significance":"If the reported numbers hold, OneAdapt would be a meaningful step toward resource-constrained photonic MBQC: the IR-level depth and area reductions are large, the dynamic-refresh design is a clean idea, and the inclusion of an independent Qiskit-based baseline and an ablation isolates the effect of the 2D-bounded routing. The depth numbers in Tables 1 and 2 are measured compiler outputs rather than fitted quantities, and the PL-ratio result reproduces the prior OnePerc value, which indicates internal consistency rather than circularity. The overall significance is, however, conditional on the physical overhead of skewed edges: the claim that Hamming-distance-1 skew is nearly free is supported only by the authors' own randomized simulation, not by a hardware scheme or a test on compiler-generated skew patterns. Reproducibility also limits confidence: no code or data are provided and randomized heuristics are evaluated with single values.","major_comments":[{"comment":"The central physical-realizability claim is not supported by the experiment presented. Fig. 10(a) measures PL ratio by selecting 10-40% of all possible skewed edges at random and skewing them in random directions on a 6x6 lattice; but the compiler's routing heuristics in Sec. 4.4 choose positions to minimize inter-node distances or maximize free space, so actual skew patterns are likely more directionally correlated and can differ substantially from random samples. The claim that PL ratio remains about 3.1 should be verified on the actual compiler-generated output, including the extreme 3x3 / D_f=10 regime of Fig. 11(c), with a distribution over compilation seeds. Without this, the advertised depth and area reductions cannot be translated into physical resource savings.","section":"Sec. 4.5, Fig. 10(a)"},{"comment":"Section 4.5 asserts that realizing skewed temporal edges requires 'little modification' because path searching can connect IR nodes at nearby 2D locations across layers, but it does not provide a concrete physical routing scheme, a resource accounting of additional delay lines or switches, or any hardware demonstration. The IR-level model counts logical nodes and layers only; it does not model whether skewed paths can be implemented with the same fusion devices and delay-line resources, or whether extra switches and delay lines are needed. Please either give an explicit physical construction with its hardware overhead, or soften the conclusion so that the depth and size reductions are claimed at the IR level rather than as physical resource reductions.","section":"Sec. 4.5"},{"comment":"Results are not reproducible from the manuscript. The compilation flow in Sec. 4.1 includes random selection of new nodes and random placement in multiple rounds, yet Tables 1-3 and Fig. 11 report a single value per configuration without seeds, variance, or a statement of the number of runs. Since the improvements are factor-level (3.68x, 3.56x, 2.87x), a randomized heuristic evaluated once per point does not rule out seed-dependent outliers. Please provide the artifact and report the mean and spread (or min/max) over a fixed number of seeds.","section":"Sec. 4.1, Tables 1-3"},{"comment":"The ablation in Table 3 has many '-' entries in the 'w/o 2D-bounded temporal routing' columns, and the text states that compilation 'would get stuck at the level of D_f ~ 75 layers' without stating the criterion for being stuck or explaining each missing entry. Since this table supports the load-bearing claim that the 2D-bounded temporal routing is what makes small D_f attainable, please specify whether missing entries are failed compilations, not attempted, or omitted for space, and give the exact congestion or failure criterion.","section":"Table 3"}],"minor_comments":[{"comment":"The phrase 'compilers on previous IRs lacks the adaptability' contains a subject-verb agreement error and should read 'lack adaptability.'","section":"Abstract and Sec. 1"},{"comment":"The sentence 'the lengths of edge e2, e4, e6 and e6 are also reduced to 4' should read 'e2, e4, e5 and e6'; the duplicate 'e6' appears to be a typo.","section":"Sec. 3.1, Fig. 6(c2)"},{"comment":"The parameter p in Eq. (1) is described as adjustable with a default of 0.4, but no sensitivity analysis is reported; please state how results change with p or justify the fixed choice.","section":"Sec. 4.3, Eq. (1)"},{"comment":"The description 'we restrict the number of dependency layers that can appear on each layer (2 is achievable)' is unclear; please specify the algorithmic mechanism and why 2 is the achievable bound.","section":"Sec. 5.1, Baseline 1"}],"recommendation":"major_revision","confidential_remarks":"The primary baseline is the authors' own OnePerc, but the independent Qiskit baseline and the ablation partly mitigate circularity concerns. The main risk is the Sec. 4.5 physical-overhead claim; I would condition acceptance on either a concrete routing construction or a characterization of compiler-generated skew patterns. The absence of code and data is an additional concern for a compiler paper in this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid systems contribution. The genuinely new pieces are the bounded temporal-edge-length IR and the Hamming-distance-1 skewed temporal edges, plus the two passes (dynamic refresh and 2D-bounded temporal routing) that exploit them. The depth reductions against OnePerc and the Qiskit-derived baseline are large and appear measured, not fitted. The ablation isolating the routing pass is sensible, and the PL-ratio simulation at skew distance 1 reproducing ~3.1 is a reasonable first check. If the hardware hypothesis holds, this directly attacks fusion-device counts and chiplet area, which are real bottlenecks. I buy the core mechanism. Soft spots, in rough order of weight. First, no code or data are shipped, and the compiler uses randomized placement in later rounds while all tables report single numbers without seeds or variance. That makes the exact improvement factors hard to pin down. Second, the abstract and introduction say the compiler \"minimizes\" depth and size, but the methods are heuristics and the paper itself notes possible node congestion; that is an overclaim. Third, the load-bearing assumption in Sec 4.5 that skewed edges require \"little modification\" and negligible overhead is supported only by the authors' own simulation, not by a hardware test or a detailed routing scheme. The stress-test concern is partially fair: Fig 10(a) randomizes edge choices and directions, which is not the same as the compiler's distance-minimizing, possibly correlated skew patterns. But Fig 10(b) does compile actual benchmarks at 6x6, so it is not purely synthetic. The extreme 3x3/D_f=10 regime is not tested for PL ratio, so the physical savings there are less certain. Still, the IR-level depth reductions are measured; even a somewhat higher PL ratio would leave substantial gains. This is not a revolutionary paradigm, and the novelty sits inside the authors' own OnePerc lineage. But it is a real step forward for photonic MBQC compilation, with honest experimentation modulo the missing artifacts. I would send it to peer review and ask for code/data and a more careful treatment of skewed-edge hardware overhead, rather than desk-reject it.","headline":"A credible compiler/IR extension with large measured depth reductions; main caveats are missing code/data, randomized placement without seeds, and a hardware-cost assumption for skewed edges supported only by simulation.","tokens_in":656,"tokens_out":1010,"would_cite":true,"duration_ms":27761,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"OneAdapt claims a resource-adaptive compiler for photonic one-way quantum computing that reduces 1D depth by 3.68x over a flexible-lattice IR and shrinks 2D area from 8x8 to 3x3 for 64-qubit programs.","keywords":["photonic quantum computing","measurement-based quantum computing","one-way quantum computing","intermediate representation","compiler optimization","temporal edges","fusion-based architecture","fault-tolerant quantum computing"],"falsifier":"A hardware experiment would settle the claim: generate a series of 2D layers with Hamming-distance-1 skewed temporal edges under 75% fusion success, and measure the average physical-to-logical layer ratio and success probability of the skewed path searches. If the ratio rises substantially above the roughly 3.1 value reported for straight temporal edges, or if skewed paths fail much more often than straight ones, the claimed depth and area reductions are not attainable at the assumed hardware cost.","tokens_in":21774,"feed_emoji":"⚛️","tokens_out":11037,"duration_ms":94299,"temperature":0.7,"pith_summary":"Photonic one-way quantum computers run a program as a sequence of measurements on a large entangled state generated by fusing small resource states. That entangled state is the intermediate representation (IR) linking program to hardware, and its shape determines required chip area, number of fusion devices, and runtime. This paper claims that a new IR—keeping the lattice structure of a cluster state but bounding temporal-edge lengths and allowing those edges to skew one step within a 2D layer—lets a resource-adaptive compiler reduce the 1D depth of compiled programs by 3.68x on average over a flexible-lattice IR, while respecting a user-given bound on temporal-edge length. Relative to a cluster-state compiler the paper reports a 3.56x depth reduction at length limit 1, and a 2.87x depth reduction in surface-code fault-tolerant settings. These numbers matter because they translate into fewer delay lines and fusion devices on each chiplet, shorter photon storage times, and smaller chip count.","feed_headline":"Photonic one-way QC compiler cuts depth 3.68x","feed_subtitle":"Bounded temporal edges and skewed layer links cut chip area and runtime, such as 8x8 to 3x3 for 64-qubit programs.","key_machinery":"The central object is an extended flexible-lattice IR: a 3D lattice with two finite spatial dimensions and one infinite temporal dimension, in which every temporal edge has length at most $D_f$ and may connect nodes whose 2D coordinates differ by a Hamming distance of 1. The mechanism that makes the claims work is a pair of optimization passes on this IR. Dynamic node refresh selects nodes stored in delay lines, remaps them to the current 2D layer before their age reaches the limit, and tunes the fraction of refreshed positions so that new nodes still have room; 2D-bounded temporal routing chooses each refreshed node's new position within the unit Hamming ball to either minimize distance to already-mapped neighbors or maximize free space. Together they convert the unbounded fusion-device cost of flexible temporal edges into a controllable resource while keeping the physical layer-generation procedure, based on percolation and path searching among successful fusions, essentially unchanged.","core_discovery":"The central claim is that the conflict between easy hardware generation and efficient program mapping in photonic MBQC can be bypassed by an IR that sits between the rigid cluster state and the unbounded graph state. The IR keeps a lattice within each finite 2D layer, allows temporal edges between layers only up to an adaptive length limit $D_f$, and permits temporal edges to connect nodes at Hamming distance 1 in the 2D plane, which the paper argues requires little physical modification. Two passes carry the result: dynamic node refresh remaps nodes before their delay-line storage reaches the limit, removing the dedicated refresh layers of prior work; and 2D-bounded temporal routing uses skewed edges to bring connected nodes together, behaving like a native SWAP in the simplest case and more generally reducing both depth and area. The evaluation claims an average 3.68x 1D-depth reduction over the prior flexible-lattice compiler with all temporal edges within the specified $D_f$ (the baseline overshoots by 89.9% on average), a 3.56x depth reduction at $D_f=1$ against a cluster-state-style compiler, 2D size reduction from 8x8 to 3x3 for 64-qubit programs at $D_f=10$, and a 2.87x resource reduction in surface-code FTQC.","pith_inferences":["The reported 2.01x advantage over a SWAP-strengthened baseline implies that the value of 2D-bounded temporal routing is global position choice, not just qubit swapping, so circuits with heavy routing demand such as QFT should show the largest gains.","If the Hamming-distance-1 skew is physically confirmed, the IR could cut delay-line diversity: instead of building $O(n)$ distinct delay-line lengths for edges up to length $n$, only the lengths actually used by the compiled IR would need to be fabricated.","The refresh bound $b_r$ and its tuning parameter $p$ are heuristics; an online controller that adapts them to observed congestion and delay-line occupancy is a natural extension that could reduce depth further without changing the IR.","Dynamic refresh is fundamentally a delay-line memory management policy, so the idea should transfer to other quantum platforms with bounded intermediate storage, though the paper only evaluates it on photonic fusion hardware."],"forward_implications":["A processor with a fixed set of fusion devices can run any compiled program up to a chosen temporal-edge bound, because every temporal edge is guaranteed to be no longer than $D_f$.","At $D_f = 1$, the compiled IR has zero measurement wait, so one measurement device per chiplet suffices and the compiler becomes a direct competitor to cluster-state mapping.","Allowing temporal edges of length up to 10 lets 64-qubit benchmarks fit on a 3x3 grid instead of 8x8, with the 3D volume roughly halved despite a local depth increase.","The same dynamic refresh that bounds fusion-device needs also cuts surface-code FTQC resource consumption by an average factor of 2.87x, so the benefit is not limited to the NISQ regime."],"supporting_citations":[{"why":"Supplies the main baseline: the flexible-lattice IR compiler with periodic refresh and unbounded temporal edge lengths.","marker":"[24]"},{"why":"Defines the cluster-state IR used as a comparison baseline and the identity-insertion property that node refresh relies on.","marker":"[18]"},{"why":"Provides the fusion-based photonic architecture whose resource-state generators, delay lines, fusion and measurement devices define the hardware cost model.","marker":"[19]"},{"why":"Fixes the 75% fusion success probability used in hardware settings, sensitivity analysis, and PL-ratio estimates.","marker":"[7]"},{"why":"Supports the claim that temporal edges of length 1 to $n$ require $O(n)$ fusion devices with separate delay lines, motivating the bounded-length constraint.","marker":"[26]"},{"why":"Supplies the dependency DAG scheduling that maps graph-state nodes onto later 2D layers in dependency order.","marker":"[22]"},{"why":"Provides the node-mapping and routing algorithm that the new adaptive passes extend.","marker":"[43]"},{"why":"Supplies the circuit-model compiler baseline adapted to $D_f=1$ with minimal 2D size and zero measurement wait.","marker":"[66]"},{"why":"Provides the logical-blocks surface-code FTQC scheme used for the QEC integration and interleaving baseline.","marker":"[53]"}],"fun_headline_variants":["OneAdapt cuts photonic one-way QC depth 3.68x","Adaptive compiler reduces photonic QC depth 3.68x","Resource-adaptive compilation cuts QC depth 3.68x","Photonic one-way QC: adaptive compiler cuts depth 3.68x","OneAdapt: 3.68x depth reduction for photonic 1WQC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that skewed temporal edges—edges that connect nodes one step apart in the 2D plane across nearby layers—can be realized on photonic hardware with only minor modifications and with no meaningful increase in the physical-to-logical layer ratio; the paper's evidence for this is its own simulation, not a hardware demonstration.","fun_headline_variants_meta":{"raw":{"variants":["OneAdapt cuts photonic one-way QC depth 3.68x","Adaptive compiler reduces photonic QC depth 3.68x","Resource-adaptive compilation cuts QC depth 3.68x","Photonic one-way QC: adaptive compiler cuts depth 3.68x","OneAdapt: 3.68x depth reduction for photonic 1WQC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000988,"raw_usage":{"total_tokens":4202,"prompt_tokens":968,"completion_tokens":3234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":3137}},"tokens_in":584,"tokens_out":3234,"duration_ms":20571,"temperature":1.0,"reasoning_tokens":3137,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:48:55.751753+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A hardware experiment would settle the claim: generate a series of 2D layers with Hamming-distance-1 skewed temporal edges under 75% fusion success, and measure the average physical-to-logical layer ratio and success probability of the skewed path searches. If the ratio rises substantially above the roughly 3.1 value reported for straight temporal edges, or if skewed paths fail much more often than straight ones, the claimed depth and area reductions are not attainable at the assumed hardware cost.","supporting_citations":[{"cited_title":"Oneperc: A randomness-aware compiler for photonic quantum computing","cited_arxiv_id":null,"evidence_quote":"Supplies the main baseline: the flexible-lattice IR compiler with periodic refresh and unbounded temporal edge lengths."},{"cited_title":"Measurement-based quantum computation on cluster states","cited_arxiv_id":null,"evidence_quote":"Defines the cluster-state IR used as a comparison baseline and the identity-insertion property that node refresh relies on."},{"cited_title":"3/4-efficient bell measurement with passive linear optics and unentangled ancillae","cited_arxiv_id":null,"evidence_quote":"Fixes the 75% fusion success probability used in hardware settings, sensitivity analysis, and PL-ratio estimates."},{"cited_title":"Logical blocks for fault-tolerant topological quan- tum computation","cited_arxiv_id":null,"evidence_quote":"Provides the logical-blocks surface-code FTQC scheme used for the QEC integration and interleaving baseline."}],"review_version":1}