{"id":"6e83d086-3a7e-4b52-b069-67956513b04e","arxiv_id":"2605.20801","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A quantum-enhanced spiking Q-network is reported to outperform classical, spiking, and quantum-dense baselines in small grid-world navigation, with gains that are small relative to the reported error bars.","lead":"This paper proposes Q-SpiRL, a navigation framework that combines spiking neural networks with a variational quantum circuit, and reports that the quantum-spiking agent reaches up to 99% success in small grid-world obstacle-avoidance tasks. A reader might consult it to see whether quantum feature transforms add value to spiking reinforcement-learning policies, though the reported gains are small and the comparison is not tightly controlled.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"QSNN-vs-SNN comparison does not isolate the quantum layer: hidden neurons (30 vs 35) and all spiking hyperparameters differ, so the central causal claim is confounded.","rationale":"The reader's verdict is REJECT because the comparison does not isolate the quantum layer AND the state representation is non-Markovian. I agree with the rejection, but the most load-bearing single concern is the confounded architecture/hyperparameter comparison, not the state representation. The paper's own design (Sec III.C, Tables I-II) varies hidden neurons and every spiking hyperparameter between QSNN and SNN, so no causal conclusion about quantum enhancement can be drawn. Even if Eq. (7) were fully Markovian, the table would still not support the headline claim. The state representation issue is real but less decisive: since all agents use the same observation function, relative performance could still be meaningful; the missing control is what kills the attribution. A controlled ablation would settle this. Thus the verdict remains REJECT, and my reading does not change the reader's overall conclusion. Agreement is partial because the reader's rationale mentions the confound, though their stated weakest assumption is the state representation.","tokens_in":15775,"tokens_out":4705,"duration_ms":56857,"concrete_test":"Train QSNN and classical SNN with identical architecture except for the quantum layer: same hidden neurons (e.g., 30), same fmax, T, Δt, τ_mem, τ_syn for each environment, and same random seeds. Use the same grid-search protocol to choose shared hyperparameters, and also run an ablation grid over the shared settings. If QSNN's SPL/TR advantage persists across all shared-hyperparameter configurations and seeds, the confound is resolved; if the advantage shrinks, reverses, or falls within standard error, the central claim fails. Repeat for QMLP/MLP.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the variational quantum layer reliably improves spiking RL policies—rests on the QSNN vs SNN comparison, but that comparison is confounded. Table II fixes hidden neurons at 30 for classical models and 35 for quantum models; Table I selects different fmax, T, Δt, τ_mem, and τ_syn for classical vs quantum SNN in every environment (e.g., 40×40: fmax 100 vs 200 Hz, T 20 vs 5, Δt 0.20 vs 0.01, τ_mem 0.04 vs 0.01, τ_syn 0.02 vs 0.005). The parameter counts also differ (SNN: 1,985; QSNN: 1,977), contradicting the 'parameter-matched' claim. Section III.C explicitly tunes classical and quantum settings separately, so the reported QSNN gains in SPL/turn rate could be caused by larger hidden layer, different temporal dynamics, or more favorable hyperparameters rather than by the quantum circuit. Without an ablation that changes only the quantum layer, the paper's central causal attribution is not established. The non-Markovian state in Eq. (7) is an additional concern, but even if the state is kept fixed, the confounded comparison alone prevents accepting the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Q-SpiRL, a framework that combines spiking reinforcement learning with a variational quantum circuit for grid-world robot navigation. It evaluates five agent families (tabular Q-learning, MLP, SNN, QMLP, QSNN) in 20x20, 30x30, and 40x40 environments with static and dynamic obstacles. The central claim is that the QSNN, which processes spike-derived firing rates through a parameterized quantum circuit, achieves the strongest overall trade-off among success rate, SPL, path length, and turn rate, and that the quantum layer consistently improves over the classical SNN. The paper also reports a single-episode feasibility run on IBM quantum hardware.","tokens_in":16107,"tokens_out":7861,"duration_ms":91273,"significance":"If the central comparison were valid, the paper would make a useful empirical contribution to hybrid quantum-spiking reinforcement learning, with a clean deterministic evaluation protocol and a hardware feasibility demonstration. The unified Q-table conversion and the use of multiple navigation metrics are commendable. However, the key causal claim—that the variational quantum layer improves spiking RL policies—is not supported by the current experiments because the QSNN and SNN differ in hidden-neuron count, hyperparameters, and parameter count. In addition, the compact state representation in Eq. (7) is not Markovian, which undermines the validity of the Q-table evaluation as a representation of an optimal policy. These are load-bearing issues that prevent acceptance.","major_comments":[{"comment":"The QSNN-vs-SNN comparison does not isolate the quantum layer. Table II fixes classical hidden neurons at 30 and quantum hidden neurons at 35, and Table I selects different spiking hyperparameters for the classical and quantum settings in every environment (e.g., 40×40: fmax 200 vs 100 Hz, T 5 vs 20, Δt 0.01 vs 0.20). Parameter counts also differ (QSNN 1,977 vs SNN 1,985), contradicting the 'parameter-matched' description. The Table III gains could therefore be due to network capacity or temporal dynamics rather than the variational circuit. A matched ablation that changes only the quantum layer is required for the central claim.","section":"Section III.B.4, Table I, Table II"},{"comment":"The state st=(Ro,Do,RT,AT→o) is not a sufficient statistic for the navigation task: it omits the agent position, the distance to the target, the distance to the nearest obstacle, and the dynamic obstacle's position. The reward in Eq. (9) explicitly depends on Euclidean distances, so the process is partially observable. Consequently, the Q-table conversion in Section III.D over the 2,560 states cannot represent an optimal Markovian policy, and all Table III metrics are for a reactive POMDP policy. The paper must either augment the state with the missing coordinates/distances or provide a formal justification of Markovianity.","section":"Section III.A, Eq. (7)"},{"comment":"No statistical significance testing is reported, and many of the claimed QSNN-vs-SNN differences are within the reported standard errors. For example, in 20×20 the SPL values are 0.8933±0.0192 vs 0.8765±0.0211; in 40×40 the success rates are 0.990±0.0100 vs 0.980±0.0141. The repeated statement that the quantum layer provides 'consistent gains' is not supported without paired tests over the same 100 seeds or confidence intervals for the differences.","section":"Section IV.B, Table III"},{"comment":"The hyperparameter selection protocol is a selection-on-performance procedure: each family is tuned independently, configurations with success rate below 95% are discarded, and the best path length among the survivors is chosen before reporting the final metrics. This creates an optimism bias and, combined with the small evaluation set, can inflate apparent advantages. The paper should separate training, validation, and test splits, and report the selection outcomes for all searched configurations.","section":"Section III.C"}],"minor_comments":[{"comment":"The reward coefficients β1, β2, β3 and the magnitude of the per-step penalty are never specified. Without these values the experiments cannot be reproduced.","section":"Section III.A, Eq. (9)"},{"comment":"The phrase 'one additional training-only diagonal static obstacle' is ambiguous: if this obstacle is absent at test time, the training and test distributions differ; if it is present at test time, 'training-only' is misleading. Please clarify.","section":"Section III.A"},{"comment":"The grid-cell side length Δ is not defined, so the reported path lengths in meters are not reproducible. Please state the physical scale used.","section":"Section III.F, Eq. (23)"},{"comment":"The table lists an initial temperature T0 and temperature decay for exploration, while the text states ε-greedy with ε=0.01. Please clarify which exploration scheme is actually used.","section":"Table II"}],"recommendation":"reject","confidential_remarks":"The paper is well organized and the evaluation pipeline is transparent, but the main experimental comparison is confounded and the state representation is not Markovian. The central causal claim is therefore not supported. A substantially revised version with matched hyperparameters, a properly Markovian state, and appropriate significance testing could be reconsidered, but the current manuscript is not suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is the first, as far as its own survey goes, to put a variational quantum circuit inside a DSQN-style spiking policy for navigation. That combination is genuinely new, though the components are known. The authors also do some things well: the environment and reward are specified precisely, the five policy families share a common evaluation protocol, and the hardware experiment is honestly labeled as a feasibility check, not a benchmark. The writing is clear.\n\nThe problem is the central comparison. The paper's headline claim is that QSNN beats the classical SNN because of the quantum layer. But QSNN uses 35 hidden neurons and the SNN uses 30, and the spiking hyperparameters (T, Δt, τmem, τsyn, fmax) are tuned separately for the two families in every environment (Table I). That means the reported gains in SPL and turn rate could come from a larger hidden layer or different temporal dynamics, not from the quantum circuit. The paper does not run an ablation that changes only the quantum layer. The 'parameter-matched' language is misleading in this context: the matching is within the classical pair and within the quantum pair, not across the pair that is actually compared. The error bars in Table III overlap for the key metrics, and no significance tests are provided, so even the direction of the effect is not established.\n\nThere is a second issue: the state in Eq. (7) is a 4-component discrete vector that omits the agent's position and the distance to the target. The Q-table conversion in Section III.D assumes this vector is a sufficient statistic. It probably isn't for these environments, since obstacles are randomized and dynamic. If the state is not Markovian, the Q-table cannot represent an optimal policy, and the absolute performance numbers become hard to interpret. This is a real concern, though it is secondary to the confounded comparison.\n\nWhat the paper does give you is a plausible experimental template and an honest exploration of a niche question. The authors do not overclaim on the hardware experiment, and the related work is properly surveyed. But the evidence does not support 'the quantum layer provides consistent gains.' A serious revision would need a matched-architecture ablation (same hidden neurons, same hyperparameters, only the VQC toggled), a more observable state, and repeated-seed significance testing. I would not cite the performance claims as they stand.\n\nWho is this for? People working at the intersection of quantum ML and neuromorphic control might want to see it as a proof of concept and a cautionary example of comparison design. I'd send it to peer review if an editor asked me, because the question is legitimate and the flaws are fixable, but I would expect major revision. It is not a paper to accept as is.","headline":"A cleanly written exploratory study that combines a variational quantum circuit with spiking RL for grid navigation, but its central claim that the quantum layer helps is confounded by differing hidden sizes and independently tuned hyperparameters.","tokens_in":16615,"tokens_out":2816,"would_cite":false,"duration_ms":31810,"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 hybrid quantum-spiking reinforcement learning agent, QSNN, is claimed to achieve the strongest balance of success rate, path efficiency, and motion smoothness in obstacle-aware grid navigation, with 99% success in a 40x40 environment.","keywords":["quantum reinforcement learning","spiking neural networks","robot navigation","variational quantum circuits","Q-table conversion","dynamic obstacle avoidance","hybrid quantum-classical policy","grid-world navigation"],"falsifier":"Retrain all five agents with the state extended to include the agent's absolute position and distance to the target, keeping the rest of the protocol identical. If QSNN no longer outperforms the classical SNN, or if a brute-force check finds two states with equal four-component features but different optimal actions, the central claim about quantum-enhanced spiking navigation is undermined.","tokens_in":15632,"feed_emoji":"⚛️","tokens_out":10162,"duration_ms":107688,"temperature":0.7,"pith_summary":"This paper tries to establish that inserting a variational quantum circuit into a spiking neural-network policy improves reinforcement learning for obstacle-aware robot navigation. It presents Q-SpiRL, a framework that trains five agent families — tabular Q-learning, classical MLP, classical spiking (SNN), quantum-enhanced MLP (QMLP), and quantum-enhanced spiking (QSNN) — under a shared environment and evaluation protocol. The central claim is that QSNN achieves the best overall trade-off among task completion, path efficiency, and motion smoothness across 20x20, 30x30, and 40x40 grids, reaching 99% success in the hardest setting while keeping high success-weighted path length and low turn rate. A sympathetic reader would care because the paper offers a concrete, tested mechanism for hybrid quantum policies — spike-derived firing rates feeding a quantum feature transformation — and shows a single successful deployment on real quantum hardware.","feed_headline":"Quantum-spiking agent tops robot navigation tests at 99% success","feed_subtitle":"Spiking plus a variational quantum circuit gives the best success-efficiency-smoothness balance from 20x20 to 40x40 grids.","key_machinery":"The central mechanism is a hybrid QSNN pipeline: a frequency-based Poisson encoder converts the one-hot state into spike trains; leaky integrate-and-fire (LIF) layers process the trains over time; mean-pooling produces an 8-dimensional firing-rate vector; a variational quantum circuit (Hadamard gates, trainable controlled rotations, entangling layers, and Pauli-Z measurements) transforms that vector into quantum features; a final classical layer maps them to five Q-values. The secondary mechanism is the conversion of every trained policy into an explicit Q-table over the 2,560 discrete states, so all five agents are compared using the same deterministic greedy lookup.","core_discovery":"The central discovery claimed by the paper is that the QSNN — a spiking Q-network whose temporally averaged firing rates are passed through an eight-qubit, three-layer variational quantum circuit before action-value estimation — gives the strongest joint performance among the five evaluated families. In all three grid sizes, QSNN matches or exceeds the classical SNN on every metric; at 40x40 it reaches 99% success with SPL 0.902 and turn rate 0.260, while the classical MLP falls to 77% success. The paper interprets this as evidence that spike-based temporal representations are an effective pre-quantum interface, and that quantum enhancement is especially beneficial in the spiking pipeline ra","pith_inferences":["Because the state vector leaves out the agent's position and its distance to the goal, the policy cannot distinguish scenes that look identical in relative angles but differ in absolute location. A natural test is to add position and target distance to the state and see whether QSNN still beats the classical SNN; if the margin disappears, the reported quantum advantage depends on the specific disc","The Q-table construction for spiking agents uses one sampled Poisson-encoding realization per state. Repeating the conversion with different seeds would show how sensitive the reported QSNN numbers are to that single draw.","The hardware section labels its experiment as an indicative feasibility check and reports only one episode; averaging many hardware episodes would convert that check into an estimate of the noise-induced drop in SPL and reveal whether near-term devices can sustain the advantage.","The angular-bin state representation suggests a transfer path to real robots: quantize continuous range measurements into the same bins; whether the quantum layer still helps after that quantization would test the mechanism beyond grid worlds."],"forward_implications":["Quantum enhancement helps the spiking agent uniformly across all four metrics, while dense MLPs gain less and degrade sharply at 40x40.","Spiking policies remain more robust than dense policies as environment size and obstacle complexity grow.","The explicit Q-table conversion makes deployment inference a lookup operation and lets stochastic or neural policies be evaluated deterministically and comparably.","The successful single episode on real quantum hardware is a feasibility proof, though the paper explicitly says it is not a statistically matched benchmark.","Success rate alone can mislead: the tabular Q-learning baseline often matches success rate but trails on SPL and turn rate, so quality metrics are needed."],"fun_headline_variants":["Quantum spiking agent bests classical rivals at 99% success","QSNN tops navigation metrics across grid sizes up to 40x40","Spiking + quantum circuit yields 99% success in robot nav","Quantum-enhanced spiking nets lead robot navigation tests","Q-SpiRL: quantum spiking nets win on success and efficiency"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the four-component state (Ro, Do, RT, AT→o) fully captures what the agent needs to know, so a 2,560-entry lookup table can represent an optimal policy; if position and target-distance are necessary, the Q-table conversion cannot express the true optimal policy and the headline numbers rest on a partial state.","fun_headline_variants_meta":{"raw":{"variants":["Quantum spiking agent bests classical rivals at 99% success","QSNN tops navigation metrics across grid sizes up to 40x40","Spiking + quantum circuit yields 99% success in robot nav","Quantum-enhanced spiking nets lead robot navigation tests","Q-SpiRL: quantum spiking nets win on success and efficiency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000584,"raw_usage":{"total_tokens":2588,"prompt_tokens":757,"completion_tokens":1831,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":1755}},"tokens_in":501,"tokens_out":1831,"duration_ms":13150,"temperature":1.0,"reasoning_tokens":1755,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T13:29:56.928131+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain all five agents with the state extended to include the agent's absolute position and distance to the target, keeping the rest of the protocol identical. If QSNN no longer outperforms the classical SNN, or if a brute-force check finds two states with equal four-component features but different optimal actions, the central claim about quantum-enhanced spiking navigation is undermined.","supporting_citations":[],"review_version":2}