{"id":"d84ef1cb-509a-496e-978b-26567438ce2b","arxiv_id":"2501.12823","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A cost-sensitive reinforcement learning environment shows an agent can learn when to pay for crop measurements to guide nitrogen fertilization in winter wheat.","lead":"This paper builds a crop-management simulator where an AI agent pays a cost to measure crop traits and decides when to measure and when to fertilize. Experiments show the agent learns to time measurements around key growth stages, and that realistic measurement costs still keep yields close to a no-cost system.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The outperformance claim over standard practice is not statistically supported: Realistic vs Standard-practice confidence intervals overlap and no paired test is reported.","rationale":"The reader's stated weakest assumption is the calibration of the hand-assigned feature costs and reward parameters, while my primary concern is the absence of a paired statistical comparison between the Realistic agent and the Standard-practice baseline. Both concerns point to the same conclusion: the headline performance claim is conditional and should not be accepted without further evidence. A paired test is the more immediate blocker because it addresses the reported numbers directly: the observed 0.16 t/ha advantage is smaller than the overlap of the confidence intervals would require for confidence. Cost sensitivity is a close second, since the Table 1 values determine the learned measuring policy and could plausibly change the ranking. I therefore agree with the CONDITIONAL verdict and recommend no change; the paper's environment contribution is real and reproducible, but the performance claim needs the additional analysis before it can be stated as established.","tokens_in":13413,"tokens_out":4175,"duration_ms":46778,"concrete_test":"Using the released code, evaluate the trained Realistic and fixed-schedule policies on identical seed-year pairs, compute per-pair yield differences, and run a paired Wilcoxon signed-rank test plus a 10,000-sample paired bootstrap; report the median difference and its confidence interval. Separately, retrain the Realistic scenario with TAGP and NAVAIL costs halved and LAI and SM costs doubled; if the Realistic minus Standard-practice difference is no longer positive, or is not significant, the headline claim is not robust to plausible cost uncertainty.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 3 is the only quantitative support for the central claim that, under realistic costs, the RL agent outperforms standard practice. The Realistic scenario median is 7.46 t/ha versus 7.30 t/ha for Standard-practice, but the bootstrapped 95% CIs are (6.45, 9.13) and (6.55, 8.65), respectively, so the intervals overlap substantially. The paper reports no paired comparison across the 10 seeds and 16 evaluation years, even though all scenarios are evaluated on the same seeded years and would permit a paired Wilcoxon or paired bootstrap test. The 0.16 t/ha margin is therefore consistent with seed or year noise. A secondary but related weakness is that the Table 1 costs are justified qualitatively rather than calibrated; a sensitivity analysis is needed to show that the Realistic scenario's policy, and its position relative to Standard-practice, is not an artifact of the specific hand-set numbers. Without these checks, the performance claim is conditional on both the evaluation statistics and the cost assumptions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an active feature acquisition formulation for crop management, extending AFA-POMDPs to include feature-specific measurement costs. An LSTM-based PPO agent acts in a WOFOST winter wheat environment with a weekly cadence, jointly deciding nitrogen fertilization amounts (0-60 kg/ha in 10 kg/ha increments) and whether to measure each of six crop features (LAI, SM, NuptakeTotal, NAVAIL, TAGP, and a random distraction feature). The environment is evaluated in four cost scenarios (No-cost, Flat-cost, Realistic, Exp-cost) plus two non-measuring RL baselines (All-observed, None-observed) and two fixed policy baselines (Standard-practice, Random-spread). The authors report that higher measurement costs hurt yield, that the Realistic-cost agent measures most often near flowering and preferentially measures cheaper features, and that the Realistic-cost agent achieves higher median yield than the standard fixed-date fertilization baseline. The paper also documents that the agent learns to ignore the random feature when measurements are costly.","tokens_in":13635,"tokens_out":4742,"duration_ms":47763,"significance":"If the performance claims are substantiated, the paper would provide a practically relevant demonstration that RL can simultaneously optimize fertilization and data collection, reducing measurement burden while maintaining yield. The manuscript's strengths include a carefully specified and released configurable environment (CropGym-ToMeasureOrNot), a clear training/evaluation split with 16 training years and 16 held-out years, multiple seeds, and a sensible set of baselines including a standard-practice fertilization schedule. The finding that the agent measures less often for costly, low-information features and aligns measurement times with critical crop development stages is a useful qualitative result. However, the central quantitative claim of outperforming standard practice rests on a small median difference with overlapping confidence intervals, and the cost parameters that define the 'Realistic' scenario are hand-assigned. These issues make the current evidence conditional and require targeted additional analysis before the main claims can be accepted.","major_comments":[{"comment":"The central claim that the Realistic-cost agent 'manages to achieve better performance compared to a baseline of standard practice' is not statistically supported by the reported results. The median yields are 7.46 t/ha for Realistic and 7.30 t/ha for Standard-practice, but the bootstrapped 95% confidence intervals are (6.45, 9.13) and (6.55, 8.65), respectively, showing substantial overlap. Since all scenarios are evaluated on the same seeded years and locations, a paired test (e.g., Wilcoxon signed-rank test or paired bootstrap over the 10 seeds and 16 evaluation years) is required to determine whether the observed margin is systematic or due to seed/year noise. Please report such a paired comparison for Realistic versus Standard-practice, and also for the other pairwise claims that depend on Table 3 (e.g., Flat-cost vs Realistic). If the paired test is not significant, the statement of outperformance should be softened accordingly.","section":"Experiments and Results, Table 3"},{"comment":"The feature measurement costs in Table 1 (LAI=5, SM=5, NuptakeTotal=20, NAVAIL=20, TAGP=25, Random=10) and the reward parameters beta=2 and D=10 are treated as fixed ground truth in all experiments. The qualitative justification is plausible, but the entire cost-scenario comparison, including the learned measuring policy in Table 2 and the performance ordering in Table 3, is conditional on these values. Different but equally realistic cost estimates could change the learned policy and the ranking of Realistic versus Standard-practice. I ask for a sensitivity analysis that varies the cost vector (e.g., scaling the expensive features by 0.5x, 2x, and 3x) and the reward parameters (beta and D) over a credible range, and reports whether the qualitative conclusions—especially the cost scenario ordering and the comparison with standard practice—remain stable. Without this, the 'realistic' scenario is an uncalibrated assumption rather than a validated setting.","section":"Design Rationale and Assumptions / Table 1"}],"minor_comments":[{"comment":"The text says Table 2 reports 'the average number of measurement actions performed by the agent in a one-year period', but the table shows values that are averaged across years and seeds. Please clarify the aggregation explicitly (e.g., mean over 16 evaluation years and 10 seeds) and state the standard deviation or MAD consistently in the caption.","section":"Experiments and Results, Table 2"},{"comment":"The sentence 'We set the Realistic cost to 10, same as Flat-cost' is ambiguous because it could be read as the entire Realistic cost vector being 10; in fact, Table 1 sets only the Random feature's Realistic cost to 10. Please rephrase to specify that the Random feature's cost is 10 in the Realistic scenario.","section":"Random feature measuring policy"},{"comment":"The term 'yearly cumulative minimum temperatures' (with values 1980.11 and 1488.36) is not standard and the units are unclear. Please define the quantity precisely, e.g., sum of daily minimum temperatures over the growing season in °C, and state what 'colder' means in this metric.","section":"Adaptive policy"},{"comment":"The cost term is written as sum from i=0 to Nm, while the text defines c as a vector of size Nm. The index should start at 1 (or the upper limit should be Nm-1) to avoid an off-by-one inconsistency.","section":"RL Environment, Eq. (1)"},{"comment":"The figure shows measurement actions as transparent vertical lines, but the legend does not distinguish which features were measured. Please indicate whether all measured features are pooled or only a subset; the caption should state this explicitly.","section":"Figure 3"},{"comment":"The hyperparameter description says 'each with 2 hidden layers with size of 256' and 'The rest of the hyperparameters we kept same as the default.' Please specify which defaults (e.g., Stable Baselines 3 PPO defaults) and list the exact architecture and any altered hyperparameters for reproducibility.","section":"RL Environment, RL agent"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a well-structured in silico study with released code, which is a positive. The main issue is that the headline claim of outperforming standard practice is not yet statistically supported, and the cost-sensitivity question is central to the paper's 'realistic' framing. Both issues are fixable within the manuscript's scope by adding paired statistical tests and a sensitivity analysis, so I recommend major revision rather than rejection. I would also ask the editor to ensure that the revised version clarifies the relationship to the earlier CropGym work, since the environment is an extension of that platform and the novelty claim should be positioned carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid, well-scoped environment paper with a genuine first—feature-specific measurement costs in RL for crop management—and it ships code. But it oversells one headline result: the outperformance over standard practice under realistic costs is a 0.16 t/ha median gap with heavily overlapping bootstrapped CIs and no paired test. That particular claim is not established, even though the rest of the paper mostly holds up.\n\nWhat's actually new: the AFA-POMDP extension is incremental (a weighted cost vector on top of Yin et al.), but applying it to WOFOST winter wheat management is new, and the environment is configurable and public. The agent learns to measure cheap features (LAI, SM) more often, to largely ignore the random distractor feature, and to shift measurement timing in cold years. Those qualitative results are convincing and align with agronomic intuition. Public code is real evidence and helps a lot.\n\nSoft spots, in order of severity. First, the same 10 seeds and 16 evaluation years are used across all scenarios, so paired tests are available; the paper reports only overlapping marginal CIs. A paired Wilcoxon could show the Realistic-vs-Standard gap is real, but as written the evidence is weak. Second, the cost vector and reward parameters (beta=2, D=10) are hand-set with qualitative justification; the scenario ranking could shift under plausible cost estimates. A cost-sensitivity analysis is needed before the performance claims are robust. Third, the Flat-cost vs Realistic comparison is similarly under-supported, though it is secondary.\n\nThe measuring-policy story is the stronger evidence; the yield-difference story is the weaker one. I would not push back on the environment or the problem framing—those are useful.\n\nThis paper deserves a serious referee. It is a good contribution for the crop-management RL community, and the environment will likely be reused. I would send it to review with a request for paired statistics and a cost-sensitivity check, not desk-reject. I would cite it as the first cost-sensitive measuring environment for crop management.","headline":"Useful environment paper with a genuine first, but the headline yield advantage over standard practice is not statistically supported.","tokens_in":14128,"tokens_out":2078,"would_cite":true,"duration_ms":21439,"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":"A recurrent PPO agent can simultaneously learn nitrogen fertilization and selective crop measurement in a WOFOST wheat environment, and with realistic per-feature costs it beats a fixed three-date schedule while measuring about a quarter…","keywords":["reinforcement learning","crop management","active feature acquisition","measurement costs","nitrogen fertilization","WOFOST","winter wheat","partially observable MDP"],"falsifier":"Re-run the Realistic-cost experiment with the cost vector scaled by 0.5, 1, 2, and 5, keeping everything else fixed, and compute the agent's net reward (yield value minus fertilizer, deployment, and measurement costs). If at any plausible scale the Realistic-cost policy's net reward falls below the fixed three-date schedule's net reward, the paper's claim that realistic costs still beat standard practice is falsified.","tokens_in":13234,"feed_emoji":"🌾","tokens_out":7942,"duration_ms":69457,"temperature":0.7,"pith_summary":"Farmers rarely have cheap, complete crop-state data before every fertilizer decision, yet most RL crop-management studies assume those observations are free. This paper tries to remove that assumption by making measurement itself a decision: in a weekly WOFOST winter-wheat simulation, a recurrent PPO agent chooses both a nitrogen application rate and which of six crop features to measure, paying an explicit cost for each measurement. The paper claims that with realistic feature costs the agent measures only about a quarter of the weeks, concentrates measurements near critical development stages, and still achieves a higher median yield (7.46 t/ha) than a fixed three-date fertilizer schedule (7.30 t/ha). It also finds that higher measurement costs degrade performance, and that in cost-free settings the agent measures roughly half the steps and matches an agent given complete observations. A sympathetic reader would care because the result suggests data-driven fertilizer optimization can be made practical under realistic data-collection budgets.","feed_headline":"RL agent times field measurements to beat fixed fertilizer schedule","feed_subtitle":"Cost-aware AI on a simulated wheat farm matches rich-observation yields while measuring only a quarter of the weeks.","key_machinery":"The load-bearing object is an active-feature-acquisition POMDP (a partially observable Markov decision process where the action space is a control action plus a measurement vector), extended so each feature has its own cost. Each week the control action is one of seven nitrogen rates, and the measurement action is a vector of binary choices for six crop features; the reward subtracts the chosen feature costs, a fixed field-deployment cost (10), and a fertilizer cost (β=2) from the weekly gain in simulated wheat yield (TWSO). Unmeasured features are presented to the policy as masked zeros, and the agent's recurrent LSTM memory must carry information across weeks. A single recurrent PPO (proximal policy optimization with LSTM actor and critic) jointly optimizes both action heads over a 47-week season in the WOFOST simulator. The mechanism is cost-driven: because DVS (development stage) and weather are always free, the agent can time its paid measurements to moments, such as flowering, where they most improve fertilizer decisions.","core_discovery":"The paper's central claim is that a single RL agent can learn a useful joint policy for applying nitrogen and for deciding when to pay for crop-feature measurements, and that doing so is necessary for realistic crop management. The authors adapt the AFA-POMDP formulation by assigning each measurable feature its own cost and appending observation masks to the input: unmeasured features appear as masked zeros, so the agent must remember past observations through its recurrent network. In the Realistic-cost scenario, the LSTM-PPO agent pays for cheap features (LAI, soil moisture) about five times a season, rarely buys the expensive lab-type measurements, and almost never measures a random distractor feature. The resulting median yield of 7.46 t/ha beats the standard-practice baseline of 7.30 t/ha and comes within about 5% of the cost-free upper bound of 7.86 t/ha. This is taken as evidence that selective, cost-aware measuring can approach complete-observation performance while cutting data collection effort.","pith_inferences":["The hand-set cost vector (5 for LAI and SM, 10 for the distractor, 20 for NuptakeTotal and NAVAIL, 25 for TAGP) is the switch that decides everything; a natural extension is to run the same training loop with cost multipliers taken from actual soil-lab and sensor-service price lists and check whether the Realistic policy still beats standard practice.","The same measure-and-control formulation could be moved from nitrogen to other costly sensing tasks, such as irrigation scheduling with soil-moisture probes or pest scouting, where the information value of a measurement has to justify its acquisition cost.","The paper's noiseless non-destructive measurement assumption probably understates the value of redundant measurement: with sensor noise and lab turnaround delays, measuring the same feature twice or measuring earlier may become rational, so field trials could reveal more measuring than the simulated policy.","Because the agent must be retrained for new sites, a cost-aware measurement policy learned in one region may not transfer; training on multiple sites with randomized initial soil conditions would test whether the flowering-stage measuring heuristic generalizes."],"forward_implications":["Under the paper's cost assumptions, a cost-aware RL policy needs far fewer observations than prior crop-management RL: about 12 of 47 weekly steps involved any paid measurement in the Realistic scenario, versus roughly half the steps when measurements were free.","Expensive features get used only when they are likely to matter: NuptakeTotal is measured about twice as often as NAVAIL despite equal cost, and the random distractor feature is measured least in every costed scenario.","The learned measuring policy is weather-adaptive: in the cold year (2010) the agent delays both measurements and fertilization relative to the normal year (1990), following the delayed crop development visible through free DVS observations.","Measurement costs change attainable yield: median yields fall monotonically from 7.86 t/ha (No-cost) to 6.63 t/ha (Exp-cost), so realistic cost assumptions are not neutral for policy evaluation.","Because the Realistic scenario beats the fixed three-date standard practice and the None-observed agent, the paper concludes that targeted measurement rather than either no observation or complete observation is the right operating point."],"supporting_citations":[{"why":"Formalizes AFA-POMDPs, the active feature acquisition setup that the paper extends with per-feature costs.","marker":"[Yin et al. 2020]"},{"why":"Extends WOFOST for nitrogen management and provides the winter wheat model used in the simulation environment.","marker":"[Berghuijs et al. 2024]"},{"why":"Introduces the original WOFOST crop growth model that simulates the weekly crop state.","marker":"[Van Diepen et al. 1989]"},{"why":"Documents WOFOST's development and validation, supporting the simulator's use as a faithful environment.","marker":"[De Wit et al. 2019]"},{"why":"Supplies the weekly time aggregation and nitrogen-management RL setup that the environment adapts.","marker":"[Kallenberg et al. 2023]"},{"why":"Introduces CropGym, the Gym-style wrapper around a crop model that this environment builds on.","marker":"[Overweg, Berghuijs, and Athanasiadis 2021]"},{"why":"Provides the recurrent PPO implementation used to train the policy.","marker":"[Raffin et al. 2021]"},{"why":"Motivates active-measure reinforcement learning for minimizing observation costs.","marker":"[Bellinger et al. 2021a]"},{"why":"Shows that mask-based observations work well with actor-critic agents for costly measuring.","marker":"[Bellinger et al. 2021b]"}],"fun_headline_variants":["AI learns when to measure crops for cheaper, better yields","Reinforcement learning cuts field measurements without yield loss","Cost-aware RL times soil checks to fertilize smarter","Selective measuring AI matches full data with fewer field visits","RL agent decides when to observe crops, cutting costs and boosting yield"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on the hand-assigned feature costs (5, 5, 20, 20, 25, 10), the fertilizer price ratio β=2, and the deployment cost D=10 being realistic; if actual measurement prices differ, the learned policy and the yield ranking could change.","fun_headline_variants_meta":{"raw":{"variants":["AI learns when to measure crops for cheaper, better yields","Reinforcement learning cuts field measurements without yield loss","Cost-aware RL times soil checks to fertilize smarter","Selective measuring AI matches full data with fewer field visits","RL agent decides when to observe crops, cutting costs and boosting yield"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000485,"raw_usage":{"total_tokens":2392,"prompt_tokens":940,"completion_tokens":1452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1371}},"tokens_in":556,"tokens_out":1452,"duration_ms":12653,"temperature":1.0,"reasoning_tokens":1371,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:45:08.350717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Realistic-cost experiment with the cost vector scaled by 0.5, 1, 2, and 5, keeping everything else fixed, and compute the agent's net reward (yield value minus fertilizer, deployment, and measurement costs). If at any plausible scale the Realistic-cost policy's net reward falls below the fixed three-date schedule's net reward, the paper's claim that realistic costs still beat standard practice is falsified.","supporting_citations":[{"cited_title":"N.; Silva, J","cited_arxiv_id":null,"evidence_quote":"Extends WOFOST for nitrogen management and provides the winter wheat model used in the simulation environment."},{"cited_title":"v.; Wolf, J","cited_arxiv_id":null,"evidence_quote":"Introduces the original WOFOST crop growth model that simulates the weekly crop state."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents WOFOST's development and validation, supporting the simulator's use as a faithful environment."},{"cited_title":"G.; Overweg, H.; van Bree, R.; and Athanasiadis, I","cited_arxiv_id":null,"evidence_quote":"Supplies the weekly time aggregation and nitrogen-management RL setup that the environment adapts."},{"cited_title":"CropGym: a Reinforcement Learning Environment for Crop Management","cited_arxiv_id":"2104.04326","evidence_quote":"Introduces CropGym, the Gym-style wrapper around a crop model that this environment builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the recurrent PPO implementation used to train the policy."}],"review_version":1}