{"id":"31a7b97c-5bf6-4026-b301-28f9b8f58a0f","arxiv_id":"2608.09377","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Socratic analysis of research values produces a network of reasons for using or abstaining from foundation models in medical imaging.","lead":"This paper maps the values that lead medical imaging researchers to use or avoid foundation models, such as data efficiency, explainability, and fairness. It provides a structured way to see why these choices are made, which could help researchers and funders reflect on their own assumptions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The value network in Figure 1 is underdetermined by the Socratic method: the stopping rule and edge directions are subjective, so the central claim of a 'clear way' is not empirically established and an independent elicitation could yield a different network.","rationale":"The reader's weakest_assumption identifies exactly the same concern: the Socratic methodology yields a meaningful and complete map, but an independent elicitation could produce a different network. My stress-test agrees with this and sharpens it: the method's lack of a substantive stopping rule and the arbitrary edge directions make the graph underdetermined in a way that directly affects the central claim. The reader's CONDITIONAL verdict already captures this limitation, so my concern does not change the verdict. I did not identify a more specific internal inconsistency or a fatal flaw; the paper is honest about its scope, and the reasoning is generally sound conditional on the value network being accepted. The concrete test I propose would settle whether the network is robust or an artifact of the authors' perspective, which is precisely the condition needed to raise or lower confidence in the central claim. I agree with the reader's verdict and do not recommend a change.","tokens_in":8507,"tokens_out":6153,"duration_ms":64738,"concrete_test":"Recruit two independent teams experienced in medical imaging ML and in the Baxter & Eagleson Socratic method. Have each team independently construct a value network for the same technical decision (use foundation models vs. train from scratch) using a pre-registered protocol with a defined saturation criterion and the same source literature. Compare the two graphs against Figure 1 on: (a) value set overlap, (b) edge direction agreement, and (c) side assignments. If agreement is below 70% on any of these, the network is not uniquely determined and the central claim requires explicit qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 6—that 'specific research values do have a clear way of motivating the use of foundation models or abstaining from their use'—depends on the value network in Figure 1 being a faithful, complete, and uniquely determined representation. The Socratic method described in Section 3 has no operational criterion for when a value is 'fundamental' or 'already explored' (step 5), and the authors stop when they have 'exhausted all the values that could be cleanly associated' (Section 4). This stopping rule is subjective, and the directionality of the edges (e.g., why 'model recognition' is instrumental for 'publishability' and not vice versa) is asserted without a systematic justification or empirical evidence. The paper itself acknowledges exceptions: 'environmentalism' motivates both sides, and 'fairness' is explicitly called ambiguous in Section 5. These exceptions weaken the strong claim even as a theoretical statement. More importantly, the claim is about how researchers actually reason, but the graph is built from the authors' introspective reasoning, supplemented by selected citations. If an independent elicitation of the same technical decision produces a substantially different set of values, edge directions, or side assignments, then the 'clear way' is a product of the method rather than a stable property of research values. This underdetermination is the most load-bearing weakness because it directly threatens the paper's central conclusion, not merely its applicability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper explores how \"research values\"—normatively loaded properties such as data efficiency, explainability, reproducibility, environmentalism, and fairness—can motivate either the use of foundation models or the decision to train/adapt a model from scratch in medical imaging. The authors apply a Socratic methodology (previously introduced in their own work) to construct a graph of fifteen research values and their instrumental/conflict relationships (Figure 1), then discuss each value in turn. Section 5 treats fairness as genuinely ambiguous, and Section 6 concludes that specific research values have \"a clear way\" of motivating foundation-model use or abstention, contributing to philosophy of machine learning in medicine. The paper is explicitly exploratory and relies on introspective reasoning supplemented by selected citations rather than on empirical elicitation or a systematic literature review.","tokens_in":8805,"tokens_out":2425,"duration_ms":26812,"significance":"If the central conclusion is accepted in a suitably qualified form, the paper makes a worthwhile conceptual contribution: it gives a structured, discussable map of why researchers in medical image analysis might legitimately justify either using foundation models or abstaining from them. The individual value discussions are generally explicit and many are supported by relevant citations, and the paper is honest about its exploratory nature and about exceptions such as environmentalism, publishability, and fairness. The main value of the contribution is therefore as a hypothesis-generating philosophical taxonomy, not as an empirical measurement of how researchers actually reason. Its broader significance depends on whether the graph is presented as a plausible possible map rather than as a uniquely determined or complete one; that calibration is the main point that needs revision.","major_comments":[{"comment":"The central claim that \"specific research values do have a clear way of motivating the use of foundation models or abstaining from their use\" is stronger than the paper's own evidence supports. Sections 4 and 5 explicitly state that environmentalism \"is torn\" between creation and use, that publishability \"should not directly motivate a specific technical decision in a consistent way,\" and that fairness \"could be argued to motivate either side.\" With at least three of the fifteen values not having a clear side assignment, the concluding sentence should be qualified, e.g., to \"most of the values examined here\" or \"values often have a clear way, with notable exceptions such as fairness.\"","section":"Section 6"},{"comment":"The construction of the graph is underdetermined by the stated methodology. Step 5 of the Socratic procedure says to repeat until all values are \"fundamental or have already been explored,\" and Section 4 says the process stopped when \"we had exhausted all the values that could be cleanly associated\" with one side. This stopping rule and the directionality of edges (e.g., why model recognition is instrumental for publishability rather than vice versa) rest on the authors' introspection, with no operational criterion or inter-rater check. Because the central claim relies on Figure 1 as a faithful representation, the paper should either provide a more explicit coding protocol and reliability evidence, present the graph as one possible reconstruction, or weaken the conclusion accordingly.","section":"Section 3 and Section 4"},{"comment":"The meaning of the conflict edge is unclear for the model-novelty versus speed/memory relationship. The text states that \"despite this contradiction, there does not appear to be a possibility of an internal tension on the part of the researcher\" and that both values \"seem to both support the right side of the spectrum.\" Yet the legend defines a conflict as \"A and B are in conflict,\" and the figure draws a conflict edge between these two values. The paper should clarify whether conflict edges represent opposition in instrumental reasoning, opposition in side assignment, or something else; as written, the graph and text are hard to reconcile.","section":"Figure 1 and Section 4, Model novelty"}],"minor_comments":[{"comment":"The legend entry \"A is potentially an instrumental value for A, i.e. A could motivate B\" appears to contain a typographical error and a possibly reversed relation; it should read \"B is potentially an instrumental value for A\" or the sentence should be rephrased so the arrow direction is unambiguous.","section":"Figure 1 legend"},{"comment":"The sentence contains \"explainations,\" which appears to be a typo for \"explanations.\"","section":"Section 4, Clinical trust"},{"comment":"The phrase \"reproducibility is complicated highly from the presence\" should read \"is highly complicated by the presence.\"","section":"Section 4, Reproducibility"},{"comment":"The text says \"modal recognition\" in the first sentence; this should be \"model recognition\" to match the section title and the graph.","section":"Section 4, Model recognition"},{"comment":"The phrase \"researchers need fewer datasets to train a particular model\" mixes count and mass nouns; it should be \"less data\" or \"fewer training examples\" for consistency.","section":"Section 4, Data efficiency"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's exploratory/philosophical scope, and I do not see a fatal methodological flaw; the main issue is that the conclusion is phrased too categorically relative to the admitted exceptions and the subjective construction of the graph. A revision that explicitly frames the network as a candidate map and adds a transparent discussion of the stopping rule and edge-direction criteria would be sufficient for me to support publication. I would also suggest having the authors look once more at the figure legend and the model-novelty conflict edge, since those were the points where the graph and text most clearly diverged."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is not a technical paper, and it doesn't pretend to be. It takes the authors' own Socratic methodology from Baxter & Eagleson, applies it to the decision to use or avoid foundation models in medical imaging, and produces the value network in Figure 1. That specific graph, plus the treatment of fairness as an ambiguous value that pulls in both directions, is the new content. The paper is clearly written and honest about its exploratory nature.\n\nThe stress-test concern lands directly on the central claim. The method's stopping rule is subjective—Section 4 ends when the authors have exhausted values that can be 'cleanly associated'—and the edge directions are asserted rather than derived. Why is model recognition instrumental to publishability and not the other way around? Could another researcher not add a value like 'professional recognition' or invert the arrow between environmentalism and speed? The paper's own exceptions (environmentalism, fairness) show the boundaries are fuzzy. So the sentence in Section 6 that 'specific research values do have a clear way of motivating' foundation model use overstates what the method establishes. A fairer formulation would be that these values can motivate use or abstention, contingent on which values a researcher prioritizes and how the technology evolves.\n\nBut this is a soft spot in the strength of the conclusion, not a fatal one. The map is still useful. It gathers scattered justifications—data efficiency, reproducibility, clinical trust, research sovereignty, environmentalism—into one place, and it flags the pretraining/inference asymmetry and the fairness ambiguity well. The citation pattern is honest; the self-citation to [3] is to their own method, and the new graph is genuinely new. The paper doesn't claim empirical validation, and it explicitly limits its scope.\n\nWho gets value from this? Medical imaging researchers who want a vocabulary for reflecting on their own choices; philosophers of science who study values in applied AI; and regulatory or funding bodies thinking about what motivates foundation model adoption. It's not revolutionary, but it's a solid conceptual contribution.\n\nMy recommendation: send it to peer review. A good referee should push on the stopping rule and ask for either an independent elicitation or a more explicit justification of edge directions. But the paper deserves that push rather than a desk reject.","headline":"A useful conceptual map of research values for foundation models in medical imaging, but the map's underdetermination makes the 'clear way' claim stronger than the method supports.","tokens_in":9268,"tokens_out":3007,"would_cite":true,"duration_ms":30589,"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 network of research values explains why medical-imaging researchers adopt foundation models or refuse them.","keywords":["research values","foundation models","medical imaging","machine learning","Socratic methodology","value graph","fairness","reproducibility"],"falsifier":"A structured survey or interview study of medical-imaging researchers asking why they adopted or avoided foundation models: if the reasons they give do not match the values and connections in the paper's network, or if a substantial set of researchers cites values the network does not contain, the central claim would be falsified.","tokens_in":8294,"feed_emoji":"🩻","tokens_out":4755,"duration_ms":44499,"temperature":0.7,"pith_summary":"This paper tries to establish that the choice between using a foundation model, adapting it extensively, or training a model from scratch in medical imaging is systematically driven by a network of research values, not just by performance. Working from a Socratic method, it assembles a value graph in which values such as data efficiency and model reuse support minimal adaptation, while explainability, clinical trust, and knowledge extraction support developing one's own models. The paper's point is that specific research values give researchers coherent, identifiable reasons to adopt or abstain from foundation models. A sympathetic reader would care because this makes a familiar technical disagreement visible as a philosophical one about what makes research good.","feed_headline":"Why medical-imaging labs reach for foundation models or refuse them","feed_subtitle":"A Socratic value map ties data efficiency and reproducibility to adoption, and explainability and trust to abstention.","key_machinery":"The carrying object is the value graph produced by the Socratic methodology of the authors' prior work [3]. Starting from the technical decision of how much adaptation of a foundation model to perform, the method posits a motivating reason, tests whether the connection is direct or needs an intermediate value, adds that intermediate value, checks for direct conflicts with existing values, and repeats until only fundamental values remain. The resulting directed graph with conflict edges is what lets the paper attribute each technical choice to a chain of instrumental values rather than to a single preference. This machinery works because it converts an abstract question about research culture into concrete, inspectable links between values and the two ends of the adaptation spectrum.","core_discovery":"The central claim is that research values form an instrumental network, presented in the paper's Figure 1, whose structure explains both sides of the foundation-model decision. On the adoption side, data efficiency, model reuse, reproducibility, research speed and facility, model recognition, and research sovereignty each provide a chain of motivation for using a pretrained model with minimal adaptation. On the abstention side, speed and memory consumption, model novelty, explainability, clinical trust, and knowledge extraction motivate extensive adaptation or training from scratch. Some values sit on both sides: environmentalism is split between the cost of creating AI and the cost of using it, publishability inherits its direction from whatever instrumental values it recruits, and fairness currently supports neither side cleanly because foundation models can both broaden representation and perpetuate pretraining biases. The paper concludes that the value network gives the scientific community a more self-reflective understanding of why foundation models spread or fail to spread.","pith_inferences":["One could test the network empirically by content-coding the justification sections of medical-imaging papers that use or reject foundation models and checking whether the stated reasons match the graph's edges.","The same Socratic value-graph method could be applied to other contested technical choices, such as adopting federated learning or synthetic data, producing comparable networks with their own ambiguous values.","If different research communities espouse different fundamental values, foundation-model adoption should vary systematically across those communities in ways the paper does not try to measure.","The graph's claim of completeness is the part most exposed to new evidence, since independent elicitation could add values or redraw edges the authors did not consider."],"forward_implications":["If the network is right, a researcher's choice to adopt or avoid foundation models can be predicted from which values they espouse, and justifications in papers can be read as expressions of these values.","The conflict between model novelty and speed and memory consumption shows that values can push in the same direction while opposing each other, so the same technical choice can be overdetermined by different value sets.","Fairness does not currently favor either side, so the paper implies that external regulation and auditing, rather than researcher values alone, may be needed to make foundation models fair.","Publishability is not a single driver: depending on which instrumental values it recruits, it can justify either using or ignoring foundation models.","Future developments such as federated learning could change the value network by adding new benefits and costs, so the map is time-dependent rather than fixed."],"supporting_citations":[{"why":"Supplies the Socratic methodology and the notion of instrumental value graphs on which the whole network construction rests.","marker":"[3]"},{"why":"Defines foundation models and documents data efficiency, model reuse, and zero-shot and few-shot adaptation that anchor the pro-adoption values.","marker":"[13]"},{"why":"Provides the primary stated motivation of data efficiency for using foundation models in healthcare.","marker":"[37]"},{"why":"Grounds the epistemic and non-epistemic value distinction used to frame reproducibility and other values.","marker":"[24]"},{"why":"Distinguishes dead neurons from under-utilised units, which the paper uses to argue foundation models incur speed and memory costs.","marker":"[12]"},{"why":"Exemplifies model distillation as the kind of extensive adaptation that speed- and memory-oriented researchers pursue.","marker":"[16]"},{"why":"Supports the claim that clinical trust relies on explainability, linking that value to the from-scratch side.","marker":"[15]"}],"fun_headline_variants":["Why medical imaging labs adopt or reject foundation models","The hidden values steering medical AI toward or away from pretrained models","Research values map why pretrained models win or lose","Socratic value map: adoption and abstention in medical imaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire network is the product of the authors' own Socratic reasoning, so if an independent elicitation of medical-imaging researchers' values produced different values or different instrumental links, the claim that these specific values systematically motivate foundation-model use would not be established.","fun_headline_variants_meta":{"raw":{"variants":["Why medical imaging labs adopt or reject foundation models","The hidden values steering medical AI toward or away from pretrained models","Research values map why pretrained models win or lose","Socratic value map: adoption and abstention in medical imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000517,"raw_usage":{"total_tokens":2449,"prompt_tokens":829,"completion_tokens":1620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":1553}},"tokens_in":445,"tokens_out":1620,"duration_ms":12562,"temperature":1.0,"reasoning_tokens":1553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:23:35.696851+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A structured survey or interview study of medical-imaging researchers asking why they adopted or avoided foundation models: if the reasons they give do not match the values and connections in the paper's network, or if a substantial set of researchers cites values the network does not contain, the central claim would be falsified.","supporting_citations":[{"cited_title":"Medical Image Analysis102, 103494 (2025)","cited_arxiv_id":null,"evidence_quote":"Supplies the Socratic methodology and the notion of instrumental value graphs on which the whole network construction rests."},{"cited_title":"IEEE Reviews in Biomedical Engineering (2025)","cited_arxiv_id":null,"evidence_quote":"Defines foundation models and documents data efficiency, model reuse, and zero-shot and few-shot adaptation that anchor the pro-adoption values."},{"cited_title":"ACM Computing Surveys58(11), 1–35 (2026)","cited_arxiv_id":null,"evidence_quote":"Provides the primary stated motivation of data efficiency for using foundation models in healthcare."},{"cited_title":"In: Cur- rent controversies in values and science, pp","cited_arxiv_id":null,"evidence_quote":"Grounds the epistemic and non-epistemic value distinction used to frame reproducibility and other values."},{"cited_title":"IEEE Transactions on Artificial Intelligence 4(4), 959–971 (2022)","cited_arxiv_id":null,"evidence_quote":"Distinguishes dead neurons from under-utilised units, which the paper uses to argue foundation models incur speed and memory costs."},{"cited_title":"CAAI Transactions on Intelligence Technology9(2), 286–302 (2024)","cited_arxiv_id":null,"evidence_quote":"Supports the claim that clinical trust relies on explainability, linking that value to the from-scratch side."}],"review_version":1}