{"id":"549f1b34-f3cb-4c43-9902-0d22ae44d6d4","arxiv_id":"2502.04408","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In a prostate radiotherapy simulator, iterative prompting of GPT-4V with Monte Carlo reward feedback produced better treatment plan scores than a DQN baseline and random gantry angles.","lead":"The authors turned GPT-4V into a radiotherapy planner by having it suggest beam angles, run a Monte Carlo dose simulator, and use the score to refine its next suggestion. In a small prostate cancer simulation, this iterative prompting beat a deep reinforcement learning baseline and random beam choices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"8-angle text-to-plan outputs vs. 5-beam environment/DQN (Secs. 2.1-2.2, Appendix) make the Sec. 3 reward gap an uncontrolled comparison until the action-space mapping is resolved.","rationale":"The reader's weakest assumption was that the DQN is a fair, adequately trained RL baseline. I agree that baseline quality matters, but the more immediate threat is internal: the action models are not solving the same problem unless the 8-angle outputs are mapped into MatRAD exactly as the 5-beam DQN actions are. The appendix shows repeated 8-angle JSON outputs, and the methods section never reconciles this with 'up to 5' input. This is checkable from the paper alone and can explain the entire mean-reward difference. I also note the sample-size mismatch (100 plans claimed vs degrees of freedom implying 30 per group), the unmeasured efficiency claim, and the absent code, but these are secondary. Because the issue is fixable by re-running on a matched action space and by releasing the parsing rule, the conditional verdict stands; it should be explicitly conditioned on this test, not only on replacing the DQN baseline.","tokens_in":10588,"tokens_out":7155,"duration_ms":68022,"concrete_test":"Re-run the Section 3 comparison with a single, documented action space. Specifically: (1) parse the text-to-plan outputs to at most 5 angles, reporting whether first-5, best-5, or some other rule is used, and recompute the 30 rewards per group; or, alternatively, train or evaluate the DQN with the same 8-angle outputs the LLM produces; (2) re-run the ANOVA and pairwise t-tests. If the text-to-plan advantage over the DQN disappears, shrinks, or reverses, the reported outperformance is an artifact of unequal beam counts. Also release the parsing and truncation code path so the mapping from the Appendix's 8-angle JSON to MatRAD is explicit.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that text-to-plan 'outperforms conventional RL-based approaches' (Abstract; Sec. 3) requires that the two methods solve the same planning problem. The paper does not establish this. Section 2.1 states the environment 'takes gantry angles (up to 5) as input,' and Section 2.2 says the DQN 'involves selecting up to 5 beam angles.' However, every complete text-to-plan interaction in the Appendix outputs eight-angle JSON lists (e.g., [10, 50, 90, 130, 170, 210, 250, 290]), with no statement of how these are reduced or parsed for a five-beam simulator. If the environment truncates or ignores angles, the evaluated plan is not the one the model reasoned about; if all eight are accepted, the DQN was constrained to a strictly smaller action space. Since the reward function sums PTV terms over all voxels and applies OAR penalties, an extra three beams can change the score for reasons unrelated to medical knowledge. Either way, the observed mean-reward gap (-211.88 vs -259.26 and -294.24) is not a controlled test of the proposed method. This is an internal inconsistency, not a difference of opinion with an external baseline choice, and it must be resolved before the comparison can be interpreted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces 'text-to-plan', a framework that turns GPT-4V into an action model for radiotherapy beam-angle planning. The model is prompted with CT images and a planning task, receives a scalar reward from the MatRAD Monte Carlo simulator, and iteratively proposes new gantry angles in JSON. The authors compare this method with a random baseline and a DQN baseline on prostate cases, reporting mean rewards of -211.88, -259.26, and -294.24 respectively, with ANOVA and pairwise t-tests. They conclude that the approach outperforms conventional RL in quality and efficiency.","tokens_in":10859,"tokens_out":4197,"duration_ms":38012,"significance":"The proof-of-concept is timely and could be of interest to the ML-for-health community if the comparison is valid. Strengths include the use of an open-source simulator, a transparent interactive prompt log in the Appendix, explicit statistical testing, and the fact that no fitted parameters are introduced in the reward. However, the manuscript currently does not establish that the text-to-plan and DQN methods solve the same planning problem, the reported sample size is inconsistent with the reported degrees of freedom, and the efficiency claim is not backed by timing data. These issues are load-bearing for the headline claims.","major_comments":[{"comment":"The environment and DQN are described as taking 'up to 5' gantry angles, but every text-to-plan interaction in the Appendix outputs eight angles (e.g., [10, 50, 90, 130, 170, 210, 250, 290]). The paper never states how these outputs are mapped to the five-beam environment. If the environment ignores or truncates angles, the scored plan is not the plan the model reasoned about; if all eight are accepted, the DQN was restricted to a strictly smaller action space. This makes the reward gap in Section 3 an uncontrolled comparison. The authors must specify the exact parsing and, ideally, rerun the comparison with matched action spaces.","section":"Sections 2.1-2.2 and Appendix"},{"comment":"The text says the models produced 100 treatment plans, but the reported ANOVA F(2, 87) implies 90 total rewards (30 per group) and the pairwise t-tests with df = 58 imply 30 per group. This discrepancy changes the sample size and the statistical support for the headline result. Please report the exact per-group sample sizes and rerun the tests accordingly.","section":"Section 2.4 and Section 3"},{"comment":"The claim that the method outperforms RL 'in both quality and efficiency' (Abstract) is not supported by any timing measurement in the paper. Section 2.2 reports that DQN training took about 7 days, but no wall-clock time, API call counts, or reward-versus-iteration curves are given for the text-to-plan method. Without such data, the efficiency claim should be removed or substantiated.","section":"Abstract, Introduction, and Discussion"},{"comment":"The DQN baseline is under-specified. The paper gives no hyperparameters, no architecture definition beyond '3D convolutions and batch normalization,' no reward curve, and no convergence check, and it does not compare against published beam-angle RL methods such as the 3D-Unet DRL in reference [1]. The conclusion that the method 'outperforms conventional RL-based approaches' therefore rests on a single, unvalidated baseline. Please provide full training details, learning curves, and ideally a comparison to a published RL baseline.","section":"Section 2.2"}],"minor_comments":[{"comment":"The text 'There is now training' appears to be a typo for 'There is no training'; similarly, 'trought' should be 'through'.","section":"Appendix"},{"comment":"The symbol P is used both as the OAR penalty and as an exponent in the same equation; please clarify whether P is a multiplicative penalty scalar or an exponent, and use distinct notation if needed.","section":"Section 2.1, Eq. (1)"},{"comment":"The DVH panel lacks axis labels and a legend identifying which curve corresponds to which structure, and the box plot panel does not state the sample size or the treatment of outliers.","section":"Figure 2"},{"comment":"The description 'optimizing the internal state of the large pretrained model rather than its weights' is vague; clarify what is actually updated between iterations, such as the conversational context or prompt history.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The action-space mismatch is the most serious issue; if the authors can show that the eight-angle outputs are faithfully reduced to five beams or rerun with matched action spaces, the comparison may become publishable. The paper would also benefit from framing the method as reward-guided prompting rather than 'few-shot RL,' since no gradient updates are performed. Given the current lack of timing data and baseline details, I cannot recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely interesting proof-of-concept, but the current version has an internal inconsistency in the action space that undermines the headline comparison, plus a few easily fixed reporting problems.\n\nWhat's new: combining a pretrained multimodal model (GPT-4V) with reward-guided few-shot prompting to pick gantry angles for prostate radiotherapy, using the MatRAD simulator. The related work is cited (Action-GPT, DRL for beam angle optimization), and the combination is new. The appendix gives the full prompting protocol, honestly showing the iterative loop. The authors also candidly acknowledge limitations: LLMs aren't reliable enough for clinical use, their vision is 2D, and they pushed GPT-4 beyond intended scope.\n\nSoft spots: The stress-test is right. Section 2.1 says the environment takes up to 5 gantry angles; Section 2.2 says the DQN selects up to 5; but every appendix interaction outputs 8 angles (e.g., [10,50,90,130,170,210,250,290]). There's no explanation of how 8 angles map onto a 5-beam environment. If the environment truncates or ignores angles, the evaluated plan isn't what the model reasoned about; if it accepts all 8, the DQN was handicapped. Either way, the reward gap in Section 3 is not a controlled comparison. This is a load-bearing flaw, not a cosmetic one.\n\nAlso: the paper says '100 treatment plans' but ANOVA F(2,87) implies 30 per group. That's a discrepancy worth fixing. The DQN baseline has no hyperparameters, no training curves, and no comparison to the published 3D-Unet DRL for beam angles, so 'outperforms conventional RL-based approaches' is overreach. The abstract claims efficiency gains, but no timing data appear anywhere. No code is released, which makes the whole thing harder to verify.\n\nWhat's good: The core idea is novel and the protocol is transparent enough that the issues are fixable. The honest limitation section is a plus. The statistical analysis, if the degrees of freedom are trusted, shows strong separation between groups.\n\nWho it's for: Anyone working on LLM/action models for medical planning, or on sample-efficient alternatives to RL-from-scratch in planning tasks. It's a proof-of-concept, not a clinical result.\n\nRecommendation: I'd accept it for peer review with heavy revision, not desk reject. The authors need to clarify the angle mapping, correct the sample size, add timing data, strengthen the baseline, and ideally release code. If they do that, the result could be credible. As is, I wouldn't cite it or rely on the numbers.","headline":"A novel prompt-based action model for radiotherapy that has a solid pilot design but an unresolved action-space mismatch makes the headline comparison uncontrolled.","tokens_in":11386,"tokens_out":4127,"would_cite":false,"duration_ms":36859,"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":"Reward-guided prompting of a multimodal model beats a DQN at simulated radiotherapy planning.","keywords":["radiotherapy treatment planning","multimodal foundation models","few-shot reinforcement learning","Monte Carlo dose simulation","gantry angle optimization","prostate cancer","text-to-plan prompting","action models"],"falsifier":"Run the same experiment on the MatRAD prostate data with a properly tuned deep Q-network (including hyperparameter search and a training budget comparable to the reported seven days) or with the published 3D-Unet DRL beam-angle method from the paper's reference [1], using the same reward function; if the text-to-plan mean reward no longer exceeds the RL baseline at the reported significance level, the central claim would be refuted. A cheaper additional falsifier is to check whether the advantage persists when the environment's maximum of five gantry angles is strictly enforced, since the paper's own appendix shows output with eight angles.","tokens_in":10371,"feed_emoji":"🎯","tokens_out":5113,"duration_ms":46899,"temperature":0.7,"pith_summary":"This paper claims that a large pretrained multimodal model can be turned into an action model for radiotherapy treatment planning by giving it a Monte Carlo simulator, a scalar reward, and a few iterative prompts, without updating its weights. The authors report that this 'text-to-plan' prompting method achieved a mean reward of -211.88 in a simulated prostate-cancer environment, significantly better than a deep Q-network baseline (-259.26) and random gantry selection (-294.24), with ANOVA F(2,87)=67.66, p<0.001 and pairwise t-tests all p<0.001. If true, the result matters because it suggests that the anatomical and dosimetric knowledge already stored in large models can be harnessed to speed up and standardize a traditionally manual, time-consuming step in radiotherapy, without the long training runs that standard reinforcement learning requires.","feed_headline":"Prompt-and-score beats RL at simulated radiotherapy planning","feed_subtitle":"Iterative reward-guided prompting of a multimodal model scored -211 vs -259 for a deep Q-network in a prostate Monte Carlo simulator.","key_machinery":"The mechanism is a reward-guided few-shot prompting loop that turns the model's hidden knowledge into actions. At each iteration, the model receives the patient CT, a textual goal, and the numeric reward from the previous plan, then outputs a JSON list of gantry angles; the Monte Carlo simulator (the MatRAD environment) computes a dose distribution and a reward that penalizes dose to organs at risk and rewards target homogeneity. The paper calls this 'optimizing the internal state of the pretrained model rather than its weights': the prompts and the model's internal activations are refined, not its parameters.","core_discovery":"On the paper's own terms, the central claim is that a pretrained vision-language model, prompted with a CT image and a text instruction to choose gantry angles, can iteratively improve its plans when given scalar reward feedback from a Monte Carlo dose simulator, and that in the MatRAD prostate benchmark this procedure outperforms a conventional deep reinforcement-learning baseline and random selection in plan quality. The reported statistics, mean rewards of -211.88, -259.26, and -294.24 for text-to-plan, RL, and random, with pairwise t-tests all below p<0.001, are offered as evidence that the gap is not noise. The authors interpret the result as showing that large multimodal models carry enough implicit knowledge of anatomy and dosimetry to bootstrap plan optimization that otherwise would have to be learned from scratch.","pith_inferences":["Inference: The headline comparison rests on a single, apparently lightly tuned DQN baseline; a fair test of the paper's claim would pit text-to-plan against stronger published beam-angle reinforcement-learning methods, such as the 3D-Unet DRL approach the authors cite, on the same data and with comparable compute.","Inference: If the method scales beyond prostate and beyond this simulator, a natural extension is to let the model propose dose constraints or beam energies, not just angles, and to use open-weights multimodal models so the whole loop is reproducible.","Paper-stated limitation: The Discussion notes that the vision-language backbone sees only 2D slices and that using GPT-4 beyond its intended scope limits generalizability; these are acknowledged by the authors as barriers to clinical adoption.","Inference: The appendix sample outputs list eight gantry angles although the paper's environment description allows up to five; enforcing that constraint is a natural robustness check that could change the measured rewards."],"forward_implications":["If the text-to-plan method is genuinely better, treatment planning can be improved in far fewer trials than RL needs, because the model starts with prior knowledge of anatomy and beam geometries.","The approach requires no gradient-based training of the policy, so it could be applied to a new patient or treatment site by simply changing the prompt and running the simulator, lowering the engineering barrier for automated planning.","The reward design used here, target homogeneity minus organ-at-risk penalties, can be reused as the objective for any future model or optimizer in the same environment.","The authors' observed 'dose bath' behavior, spreading beams to avoid critical structures, suggests the model can describe and follow a clinically sensible strategy, which is a step toward more explainable automated plans."],"supporting_citations":[{"why":"Supplies the MatRAD Monte Carlo simulator and the public prostate data that define the planning environment and the reward computation.","marker":"[7]"},{"why":"Provides the GPT-4V vision-language backbone that the text-to-plan method prompts and iteratively refines.","marker":"[19]"},{"why":"Introduces the action-model concept of a language model interacting with the world through functions, which the paper adapts to radiotherapy planning.","marker":"[11]"},{"why":"Defines the Gym interface template that the MatRAD environment follows, enabling the standard RL baseline setup.","marker":"[5]"},{"why":"Presents a deep reinforcement learning method for beam angle optimization that motivates the problem and represents the state of the art the authors claim to outperform.","marker":"[1]"}],"fun_headline_variants":["Prompt-and-score transforms multimodal models into radiotherapy planners","Reward-guided prompting beats RL in simulated radiotherapy planning","Few-shot RL turns vision-language models into dose planners","Multimodal model learns radiotherapy planning from reward feedback","Prompt-and-score outperforms deep Q-network in prostate plan simulation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the deep Q-network baseline is a fair, representative instance of conventional reinforcement learning for this problem; if the baseline is weak or undertrained, the reported advantage of text-to-plan over 'RL' would not generalize to stronger RL methods.","fun_headline_variants_meta":{"raw":{"variants":["Prompt-and-score transforms multimodal models into radiotherapy planners","Reward-guided prompting beats RL in simulated radiotherapy planning","Few-shot RL turns vision-language models into dose planners","Multimodal model learns radiotherapy planning from reward feedback","Prompt-and-score outperforms deep Q-network in prostate plan simulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000479,"raw_usage":{"total_tokens":2333,"prompt_tokens":870,"completion_tokens":1463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1368}},"tokens_in":486,"tokens_out":1463,"duration_ms":10273,"temperature":1.0,"reasoning_tokens":1368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:14:25.202650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same experiment on the MatRAD prostate data with a properly tuned deep Q-network (including hyperparameter search and a training budget comparable to the reported seven days) or with the published 3D-Unet DRL beam-angle method from the paper's reference [1], using the same reward function; if the text-to-plan mean reward no longer exceeds the RL baseline at the reported significance level, the central claim would be refuted. A cheaper additional falsifier is to check whether the advantage persists when the environment's maximum of five gantry angles is strictly enforced, since the paper's own appendix shows output with eight angles.","supporting_citations":[{"cited_title":"In: Jaffray, D.A","cited_arxiv_id":null,"evidence_quote":"Supplies the MatRAD Monte Carlo simulator and the public prostate data that define the planning environment and the reward computation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GPT-4V vision-language backbone that the text-to-plan method prompts and iteratively refines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the action-model concept of a language model interacting with the world through functions, which the paper adapts to radiotherapy planning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Gym interface template that the MatRAD environment follows, enabling the standard RL baseline setup."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents a deep reinforcement learning method for beam angle optimization that motivates the problem and represents the state of the art the authors claim to outperform."}],"review_version":1}