{"id":"ce784a0b-5954-4a24-bed0-2e5a8017b474","arxiv_id":"2509.00095","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A TD3 agent with Dirichlet priors and quantum-inspired genetic mutation matches Apple's historical R&D/SG&A splits on held-out quarters, but the evaluation metric nearly reproduces the training objective.","lead":"Apple's financials are used to train a reinforcement learning agent that splits spending between research and administration, with a genetic algorithm and a quantum-inspired mutation step added on top. Its near-perfect match to Apple's own past allocations mostly reflects that the training signal and the test metric are the same imitation score, not proof that the model is a better budget planner.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported cosine similarity is evaluated against the same future actual allocations used as the training reward (Eq. 5), so 0.9990 measures fit to the label, not predictive or prescriptive validity.","rationale":"The central claim is that the model 'demonstrates the promise' of adaptive budgeting because it aligns with actual allocations on unseen data. For that claim to hold, the evaluation must be independent of the training signal. It is not: Eq. (4) defines the target a_hat_t using next-quarter actuals, and Eq. (5) rewards the agent for matching that target. The Table I metrics compare against the same target, so high cosine similarity is an expected consequence of optimizing the reward. The chronological-split leakage, where the last training reward uses the first test quarter's label, makes the evaluation even less clean. A persistence baseline is the natural control: if last quarter's actual allocation already achieves ~0.999 cosine similarity on slowly varying ratios, the reported number carries no information about the RL, Dirichlet, or quantum components. I agree with the reader's weakest assumption; the concern is about internal circularity of the evaluation, not disagreement with field consensus. The paper does describe a concrete pipeline and reports some GA hyperparameters, but without code or data the central claim cannot be independently checked; the proposed baseline experiment is the minimal check that would resolve it.","tokens_in":10238,"tokens_out":4585,"duration_ms":46949,"concrete_test":"Evaluate a naive persistence policy on the same held-out 20% chronological split: at each test quarter, output the previous quarter's actual R&D/SG&A allocation ratio. Compute cosine similarity and KL divergence against the actual next-quarter ratios used in Table I. If this zero-parameter baseline achieves values close to 0.9990 and 0.0023, the reported headline metrics are not evidence that the RL, Dirichlet, or quantum-genetic components add predictive or prescriptive value.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (4) defines the target a_hat_t from R&D_{t+1} and SG&A_{t+1}, i.e., the next quarter's actual allocation. The first reward term in Eq. (5) is -||a_t - a_hat_t||_1, so the policy is explicitly trained to output the next quarter's observed allocation. Table I then evaluates the policy by comparing its outputs to exactly this same a_hat_t sequence. The 0.9990 cosine similarity and 0.0023 KL divergence therefore quantify how well the optimization fit the training label, not whether the model has discovered profitable or generalizable budgeting rules. The problem is compounded by the chronological split: because a_hat_t uses t+1 data, the last training transition's reward is computed from the first quarter of the held-out test period, leaking test information into training. A persistence baseline that repeats the previous quarter's actual allocation would likely also score near 0.9990 cosine similarity, since Apple's R&D/SG&A ratios are slowly varying. The profitability objective described in the text does not appear in Eq. (5), leaving the reported alignment as the entire basis for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid framework for quarterly budget allocation between R&D and SG&A, combining TD3 reinforcement learning with Dirichlet-based Bayesian belief updates and a genetic algorithm whose mutation operator is inspired by quantum rotation gates. The agent is trained on Apple Inc. quarterly data (2009–2025) with an 80/20 chronological split, and the authors report high alignment with actual allocations on the held-out test set (cosine similarity 0.9990, KL divergence 0.0023). The abstract claims that the model maximizes profitability while adhering to historical spending patterns, and the paper concludes that the framework can deliver 'adaptive, forward-looking policies' for corporate budgeting.","tokens_in":10611,"tokens_out":3888,"duration_ms":37895,"significance":"If the reported results were valid, the paper would offer a novel integration of Bayesian RL and quantum-inspired evolutionary search for a practical financial allocation problem, and the use of real Apple data is a useful feature. The paper also documents a complete, modular pipeline with explicit hyperparameters, which aids reproducibility. However, the central empirical claim is not currently supported: the reward function in Eq. (5) trains the policy to reproduce next-quarter actual allocations, and the evaluation in Table I measures agreement with the same target, making the headline numbers a measure of training fit rather than predictive or prescriptive validity. The absence of a profitability term in the reward, the lack of error bars, and post hoc seed selection further undermine the significance claims.","major_comments":[{"comment":"The reward function in Eq. (5) does not contain the profitability term that the text repeatedly describes. Section III-B states that the reward integrates profitability and allocation accuracy, and the introduction says the agent learns to 'maximize net income while staying close to historically realistic ratios.' However, Eq. (5) is rt = -||a_t - a_hat_t||_1 - lambda1 * ||a_t - a_{t-1}||_2 - lambda2 * KL(Dir(alpha_t)||Dir(alpha_prior)), with no net-income or profit term. Consequently, the model is not trained to maximize profitability at all; it is trained solely to match the next-quarter empirical allocation a_hat_t and to maintain smoothness and belief coherence. The abstract's claim that the framework 'maximizes profitability' is therefore unsupported by the stated objective.","section":"III-B, Eq. (5)"},{"comment":"The evaluation metric is the same target used in the training reward, making the reported results circular. Eq. (4) defines a_hat_t as the actual next-quarter allocation computed from R&D_{t+1} and SG&A_{t+1}, and the first reward term in Eq. (5) penalizes ||a_t - a_hat_t||_1. Table I then reports cosine similarity and KL divergence between the model's outputs and this same a_hat_t sequence on the held-out set. A cosine similarity of 0.9990 and KL divergence of 0.0023 therefore quantify how well the optimization reproduced the labels it was trained to imitate, not whether the model discovered profitable or generalizable budgeting rules. A persistence baseline that repeats the previous quarter's observed allocation would likely score similarly because Apple's R&D/SG&A ratios are slowly varying; no such baseline is reported.","section":"III-A Eq. (4) and IV Table I"},{"comment":"The chronological split does not fully prevent data leakage because the reward for the last training transition uses test data. Eq. (4) defines a_hat_t using R&D_{t+1} and SG&A_{t+1}, and Section IV states an 80/20 chronological split. Thus, the reward for the final training timestep at t = 0.8T is computed from the first quarter of the held-out test period (t+1 = 0.8T+1). This leaks test information into training and invalidates the claim that the evaluation is on 'unseen fiscal data.' The leakage also affects the Dirichlet update in Eq. (6), which uses the same a_hat_t.","section":"III-A, IV"},{"comment":"Table I reports only single-run values and explicitly states that the best scores were obtained by setting seed=60, which indicates post hoc seed selection. No error bars, standard deviations, or multiple-seed averages are provided, so it is impossible to assess the statistical significance of the improvements from the genetic and quantum-inspired components. The comparison in Table I is also incomplete: it does not include a persistence baseline, a random policy, or a standard GA without quantum mutation, so the marginal contributions of the Dirichlet prior and the quantum-inspired operator cannot be separated from variance or from the GA's general exploration.","section":"IV, Table I"}],"minor_comments":[{"comment":"The quantum-inspired mutation is described qualitatively as 'non-local' and 'probabilistic,' but no experiment isolates its effect. The paper should compare the quantum mutation against a Gaussian noise mutation with the same budget and also report the distribution of fitness across GA generations with error bars, as Figure 3 currently shows a single trajectory.","section":"III-D, IV"},{"comment":"The caption refers to 'RL + SPC + Quantum-Genetic Algorithm,' but the configuration in Table I and text is 'RL+DP+Genetic+Quantum' or 'RL+DP+Genetic+Quantum.' The acronym SPC is never defined.","section":"Figure 2 caption"},{"comment":"The final paragraph states that the framework enables firms to 'generate sustained and guaranteed returns over time.' Given the evaluation limitations and the absence of any counterfactual profitability analysis, 'guaranteed returns' is an overclaim and should be removed.","section":"V, Conclusion"},{"comment":"The financial disclosure says only that 'the authors declare the financial support received for this work' without identifying the funding source; if there is a sponsor, it should be named.","section":"Financial Disclosure"},{"comment":"The TD3 update equations are standard, but the notation for the target value y in Eq. (9b) omits the target-network parameter dependence; it should be written as y = r + gamma * min(Q1'(s', pi'(s')), Q2'(s', pi'(s'))) to be precise.","section":"III-C, Eq. (9)"}],"recommendation":"reject","confidential_remarks":"This manuscript reports an evaluation that is circular with respect to its training objective and omits the profitability term that motivates the paper. These are load-bearing issues that cannot be fixed by local revisions: the headline result (cosine similarity 0.9990) is essentially a training fit statistic. I would recommend rejection rather than major revision, though a fundamentally rewritten paper with a proper reward, baselines, and statistical evaluation could be resubmitted as a new study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the central result does not survive contact with the reward function. Eq. (5) penalizes -||a_t - a_hat_t||_1, where a_hat_t is the actual next-quarter allocation, and Table I evaluates the policy against exactly that same a_hat_t sequence. So the 0.9990 cosine similarity is a fit-to-label number, not evidence of predictive or prescriptive validity. The stress-test note holds up: the last training transition uses the first test quarter's actuals, so the chronological split leaks; and a persistence baseline that repeats the previous quarter's actual ratio would likely score near 0.9990 because Apple's R&D/SG&A split is smooth. The profit objective described in the text is absent from Eq. (5) — the equation only contains the L1 accuracy term, an L2 smoothness term, and the Dirichlet KL penalty. That is a load-bearing omission, not a typo.\n\nWhat is new and worth credit: the specific combination of TD3 with a Dirichlet belief update plus quantum-inspired genetic mutation applied to a corporate budget split appears not to have been published before. The methodology is described clearly enough to follow, the libraries are standard, and the GA hyperparameters are reported (10 generations, population 5, elite 0.4, mutation 0.1). The Dirichlet KL penalty is a reasonable regularization idea, even if the prior [5.0, 3.0] is set arbitrarily.\n\nSoft spots, in proportion: no code or data are provided; no seed variance is reported; the best seed is chosen post hoc (seed=60); TD3 hyperparameters are unreported. Minor but telling: Figure 2's caption says \"RL + SPC + Quantum-Genetic Algorithm\" while the text says RL+DP+Genetic+Quantum, likely a leftover from an earlier draft. Also reference [20] appears to be a convex optimization survey, not the convergence analysis of Bayesian RL that the sentence implies. The conclusion overclaims \"guaranteed returns.\"\n\nWho this is for: someone curious about how NOT to evaluate a financial RL agent, or a reader wanting a checklist of evaluation pitfalls. As a contribution, it does not deserve a serious referee slot yet. The right next step is to rewrite the reward to actually include profit, hold out a proper test set with no t+1 leakage, compare against persistence and simple baselines, and report distribution over seeds. If the authors do that, the quantum-genetic part might become interesting; right now it is decorative. My recommendation: desk reject.","headline":"The headline cosine similarity is an artifact of the reward function, which trains the policy to match next-quarter actual allocations and then evaluates against the same values; the paper does not demonstrate predictive or prescriptive budgeting value.","tokens_in":695,"tokens_out":909,"would_cite":false,"duration_ms":31688,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum-inspired RL matches Apple's budget splits to 0.999 cosine","keywords":["budget allocation","reinforcement learning","Dirichlet priors","TD3","genetic algorithm","quantum-inspired optimization","financial decision making","Bayesian updating"],"falsifier":"Run the same training and evaluation while replacing the reward target with a forecast of next-quarter allocations made from data available at time $t$, or compare the trained policy against a trivial baseline that outputs the previous quarter's split; if the trivial baseline also scores cosine similarity near 0.9990, or if removing the lookahead target destroys the score, then the result is an artifact of the reward design rather than evidence of adaptive budgeting.","tokens_in":9998,"feed_emoji":"📊","tokens_out":7756,"duration_ms":64716,"temperature":0.7,"pith_summary":"The paper tries to establish that a hybrid budget-allocation model—deep reinforcement learning with a Dirichlet prior over spending categories and a genetic optimizer that mutates policy weights with quantum-style rotations—can reproduce a large company's actual R&D versus SG&A splits on data it has not seen. Working from Apple's quarterly financials (2009–2025), the agent learns a two-way split that balances net income against three penalties: closeness to the observed next-quarter split, smoothness from one quarter to the next, and disagreement with a Dirichlet prior. The best configuration reaches cosine similarity 0.9990 and KL divergence 0.0023 against the real allocations, improving from 0.8832 without the genetic layer and 0.9813 with genetics but no quantum-style mutation. The value of the claim, if true, is that budget decisions can be cast as learnable, interpretable, uncertainty-aware policies rather than static rules.","feed_headline":"Quantum-inspired RL matches Apple's budget splits to 0.999 cosine","feed_subtitle":"Dirichlet priors plus genetic mutation help a TD3 agent reproduce R&D and SG&A ratios on held-out Apple quarters.","key_machinery":"The load-bearing object is the reward function in Eq. (5): $r_t = -\\|a_t - \\hat{a}_t\\|_1 - \\lambda_1 \\|a_t - a_{t-1}\\|_2 - \\lambda_2 D_{\\mathrm{KL}}(\\mathrm{Dir}(\\alpha_t) \\| \\mathrm{Dir}(\\alpha_{\\mathrm{prior}}))$. The first term pushes the agent to imitate the actual next-quarter allocation split, the second discourages abrupt changes, and the third keeps the agent's Dirichlet belief, updated as $\\alpha_t = \\alpha_{t-1} + \\hat{a}_t \\cdot c$, close to its prior. The trained actor's weights are then treated as genes, and a rotation gate $R(\\Delta\\theta)$ applied to each qubit-encoded gene supplies the mutation that lets the search escape local optima.","core_discovery":"The paper's central claim is that combining three components in one pipeline—TD3 deep RL, a Dirichlet belief state updated from observed allocation ratios, and a genetic algorithm whose mutation step rotates qubit-encoded gene amplitudes by a parameterized angle—produces budget allocations that track Apple's historical R&D/SG&A split almost exactly on held-out quarters. The reported test-set numbers for the full model are MAE 0.0229, RMSE 0.0283, cosine similarity 0.9990, and KL divergence 0.0023. Each ablation step degrades the fit, with RL plus Dirichlet priors alone at 0.8832 and the non-quantum genetic version at 0.9813, so the evolutionary fine-tuning stages are credited with most of the final accuracy.","pith_inferences":["The headline cosine similarity is computed against the same historical allocation ratios that the reward's first term pushes the agent to imitate, so the metric should be read as a fidelity-to-past-behavior score, not as evidence that the policy improves future profitability.","Because the target $\\hat{a}_t$ is the actual next-quarter split, the reward contains lookahead information; a deployable version would have to substitute a forecast, and the reported 0.9990 fit would almost certainly degrade under that change.","A natural stress test, not run in the paper, is to compare against a no-learning baseline that simply repeats the most recent observed split; if that baseline scores near 0.9990 cosine similarity, the learned policy's apparent success is largely the persistence of Apple's allocation ratios."],"forward_implications":["Adding the genetic and quantum-style refinement layers raises test-set cosine similarity from 0.8832 for RL plus Dirichlet priors to 0.9990, so the paper's central finding is that the evolutionary fine-tuning stage carries much of the performance.","Because the Dirichlet posterior is updated from observed splits and penalized against a prior, the policy retains an interpretable belief about plausible allocations rather than a single point estimate.","The ablation ordering implies the non-quantum genetic stage alone accounts for a large jump to 0.9813, with the quantum-style mutation supplying the remaining refinement.","The authors' stated extensions—adding macroeconomic indicators, constraints, real-hardware quantum tests, and multi-agent versions—outline a route from this two-category allocation problem to fuller budgeting systems."],"supporting_citations":[{"why":"Supplies the source of Apple's quarterly R&D, SG&A, and net income figures used to build the environment and the evaluation target.","marker":"[9]"},{"why":"Provides the environment interface that the custom budget simulator follows.","marker":"[25]"},{"why":"Introduces the TD3 algorithm used for policy learning.","marker":"[26]"},{"why":"Supplies the quantum-circuit machinery behind the mutation operator.","marker":"[28]"},{"why":"Provides the automatic-differentiation backend for the neural networks.","marker":"[29]"},{"why":"Provides the TD3 implementation used in the experiments.","marker":"[30]"},{"why":"Defines the KL divergence used both in the reward penalty and as an evaluation metric.","marker":"[31]"}],"fun_headline_variants":["Quantum-genetic RL nails Apple's R&D/SG&A split to 0.999","RL with Dirichlet priors and quantum mutation hits 0.999 cosine","Hybrid RL matches Apple's budget ratios with 0.999 similarity","TD3 + Dirichlet + quantum GA reproduces Apple's allocations","Quantum-inspired evolution tunes RL to Apple's exact split"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole evaluation rests on treating Apple's historical split between R&D and SG&A as the ground truth for good budgeting, and on making that split visible to the agent through the reward term; if matching past allocations is not the right objective, the reported 0.9990 similarity is a measure of imitation, not of decision quality.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-genetic RL nails Apple's R&D/SG&A split to 0.999","RL with Dirichlet priors and quantum mutation hits 0.999 cosine","Hybrid RL matches Apple's budget ratios with 0.999 similarity","TD3 + Dirichlet + quantum GA reproduces Apple's allocations","Quantum-inspired evolution tunes RL to Apple's exact split"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000755,"raw_usage":{"total_tokens":3330,"prompt_tokens":891,"completion_tokens":2439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":2343}},"tokens_in":507,"tokens_out":2439,"duration_ms":15385,"temperature":1.0,"reasoning_tokens":2343,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:49:13.383058+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training and evaluation while replacing the reward target with a forecast of next-quarter allocations made from data available at time $t$, or compare the trained policy against a trivial baseline that outputs the previous quarter's split; if the trivial baseline also scores cosine similarity near 0.9990, or if removing the lookahead target destroys the score, then the result is an artifact of the reward design rather than evidence of adaptive budgeting.","supporting_citations":[{"cited_title":"Apple income statement 2010–2024 — aapl,","cited_arxiv_id":null,"evidence_quote":"Supplies the source of Apple's quarterly R&D, SG&A, and net income figures used to build the environment and the evaluation target."}],"review_version":1}