{"id":"5627dad6-ebb8-4920-82fe-098b2c3895e3","arxiv_id":"2603.00777","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Tool-using chest X-ray agents show demographic disparities in tool usage, routing, and reasoning that are not visible from end-to-end accuracy alone.","lead":"This paper audits demographic fairness in chest X-ray AI agents that use a large language model to call specialized tools, and proposes a method to break observed unfairness into tool-exposure, tool-routing, and reasoning stages. It reports that gaps appear in intermediate agent behavior that end-to-end accuracy alone would miss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tool-exposure bias (Eq. 2) conditions on agent-chosen tool use, so the observed accuracy gap may reflect routing selection rather than the tool's own utility; this undermines the causal decomposition that is the paper's central contribution.","rationale":"The reader's weakest assumption identified the same issue, and I agree it is the load-bearing point. I considered two other candidates: (1) the 'not predictable from end-to-end' claim is not statistically tested, and (2) the transition-matrix differences may reflect marginal tool-use rates rather than routing patterns. However, the tool-exposure metric is the cornerstone of the proposed decomposition—the term 'tool-exposure bias' is part of the title and abstract. If it is confounded, the decomposition's causal narrative collapses, even though the raw descriptive findings (e.g., intermediate disparities exist) may remain valid. The forced-exposure test is a direct intervention that would distinguish tool-specific utility from selection; it is feasible on the MedRAX codebase and would provide the evidence needed to upgrade the claim from associative to causal. Without such a check, the CONDITIONAL verdict is appropriate: the paper should either perform this validation or soften the decomposition language to 'exposure-conditioned associations'.","tokens_in":10321,"tokens_out":4313,"duration_ms":49211,"concrete_test":"Run a forced-exposure experiment: for a random subset of instances from each demographic group, bypass the agent's planner and invoke each tool A with a fixed, balanced policy, then recompute subgroup accuracy gaps on the resulting trajectories. If the gap under forced exposure is close to the observed ΔTEB(A), selection is not the driver; if it shrinks or changes sign, the original metric is confounded by routing. A complementary analytical check: re-estimate ΔTEB(A) with inverse probability weighting over a difficulty proxy (e.g., whether the QA tool was also used, question type) and compare adjusted vs. unadjusted gaps. Either check would settle whether the tool-exposure 'bias' is a tool property or an artifact of agent routing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The tool-exposure bias metric ΔTEB(A) in Eq. (2) is defined as the subgroup accuracy gap on instances for which the tool was invoked (E_A(τ)=1). Tool invocation, however, is selected by the same LLM agent during execution, based on the query, image, and prior observations. If the distribution of cases routed to a tool differs across demographic groups—e.g., female patients may be routed to segmentation only for harder cases—then the accuracy gap among exposed instances reflects selection (or collider bias) rather than the tool's independent utility. The paper acknowledges this risk by reporting exposure rates, but reporting the rate does not adjust for case-mix differences; it only makes the confounding visible. Moreover, the metric uses final-answer accuracy as the 'utility,' so any downstream LLM reasoning disparity is also baked into ΔTEB, meaning it is not a clean measure of tool-level bias. Because the central claim is that unfairness can be decomposed into tool-exposure, tool-transition, and LLM-reasoning sources, the validity of this decomposition hinges on ΔTEB having a causal, or at least confound-adjusted, interpretation. Without that, the paper demonstrates intermediate disparities but not their causal localization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DUCX, a stage-wise fairness audit framework for tool-using chest X-ray agents built on MedRAX. It defines three process-level bias metrics—tool-exposure bias (subgroup accuracy gap conditioned on tool invocation), tool-transition bias (difference in Markov transition matrices over tool calls), and LLM reasoning bias (subgroup gaps in judge-scored reasoning quality, hedging frequency, and demographic term usage)—and reports end-to-end fairness metrics (ACC, ΔACC, DP, EoD, FUT) across five driver LLMs and two benchmarks (CheXAgentBench and a newly curated MIMIC-FairnessVQA). The central claim is that agentic unfairness arises at intermediate stages and cannot be predicted from end-to-end evaluation alone, motivating process-level audits and targeted mitigation.","tokens_in":10608,"tokens_out":3070,"duration_ms":34719,"significance":"The topic is timely: as medical AI moves from standalone models to agentic pipelines, existing fairness audits that only examine final predictions may miss disparities introduced by tool routing and reasoning. The paper's main contribution is a concrete evaluation recipe with released code and a new demographic-aware VQA benchmark, plus a systematic comparison across five backbones. If the decomposition claims held, this would be a useful step toward process-level fairness auditing in clinical agents. The authors also deserve credit for reporting bootstrapped confidence intervals and for openly acknowledging the confounding role of exposure rates, even though, as argued below, the acknowledgment is insufficient for the paper's central attribution claim.","major_comments":[{"comment":"The tool-exposure bias ΔTEB(A) conditions on E_A(τ)=1, i.e., on the agent having chosen to invoke tool A. Tool invocation is a decision made by the same agent under evaluation and depends on the query, image, and prior context. If cases routed to a tool differ in difficulty across groups—for instance, if female patients are routed to segmentation only on harder cases—the accuracy gap among exposed instances measures a mixture of tool utility and routing selection, not a clean tool-level disparity. The paper reports exposure rates, but reporting the rate does not adjust for case-mix; it only makes the confounding visible. Because the abstract and Sec. 2.3 claim that the framework 'decomposes' end-to-end bias into tool-exposure, transition, and reasoning sources, this metric's causal or confound-adjusted status is load-bearing. The authors should either (a) define a causal estimand and adj","section":"Sec. 2.3, Eq. (2)"},{"comment":"The word 'decomposition' is not supported by any formal relationship. End-to-end bias is reported as ΔACC/DP/EoD; tool-exposure bias is a conditional accuracy gap; tool-transition bias is a matrix difference P^(g1)−P^(g2); reasoning bias is a set of response-feature gaps. These quantities are not connected by any identity, inequality, or accounting equation that would allow a reader to attribute a share of end-to-end bias to each stage. The central claim in the abstract—that the framework 'separates end-to-end bias from three agent-specific sources'—therefore overstates what the metrics establish. At minimum, the paper should either derive a plug-in relation (e.g., how ΔACC relates to exposure-conditioned gaps and exposure-rate differences) or soften the claim to 'three complementary process-level audits,' rather than a decomposition.","section":"Sec. 2.3, Eqs. (2)–(5)"},{"comment":"Several reported reasoning-bias gaps have confidence intervals that include zero or are extremely wide. For example, under Gender on CheXAgentBench, Qwen3VL has ΔHedge = 534.95 with CI [3.58, 1438.18]; on MIMIC-FairnessVQA, Qwen3VL Age ΔHedge is 1165.50 with CI [89.77, 3893.21]. A gap whose 95% CI spans more than two orders of magnitude is not evidence of a robust 'distinct subgroup disparity.' Likewise, in Table 1, EoD values of 20.79% carry CIs [8.91, 33.72] and 16.89% CIs [7.96, 26.50], indicating substantial uncertainty. The qualitative conclusion that 'intermediate behaviors exhibit distinct subgroup disparities' is not uniformly supported by the inferential statistics. The authors should apply multiple-comparison corrections, report effect sizes with tighter bounds (e.g., by increasing resamples or using analytic standard errors), and restrict strong claims to effects whose CIs exc","section":"Table 2"},{"comment":"Tool-transition bias is defined as a full matrix difference ΔTTB = P^(g1) − P^(g2), but no scalar summary, significance test, or overall magnitude is provided. The narrative in Sec. 3 relies on visual inspection of the matrices (e.g., 'female patients are more likely to proceed directly from the Classifier or Report Generator'), without any test of whether the observed transition differences exceed sampling variability. Given that the agent trajectories are finite and tool-call counts are modest, a permutation test or a norm-based summary (e.g., total variation distance with bootstrap CI) is needed before claiming 'distinct tool transition patterns.' This is especially important because the transition analysis is one of the three pillars of the proposed decomposition.","section":"Sec. 2.3, Eq. (4) and Fig. 3"}],"minor_comments":[{"comment":"The abstract calls the method 'DUCK' while the title and the rest of the text call it 'DUCX.' This inconsistency should be fixed.","section":"Abstract"},{"comment":"The footnote 'Dataset is provided here' contains no URL; a working link or repository identifier should be added.","section":"Sec. 2.4"},{"comment":"The Markov transition matrix is defined over the tool set T, but Fig. 3 includes a 'START' state. The equation and the figure should be aligned: either add a START row/column in Eq. (3) or explain how initial transitions are encoded.","section":"Sec. 2.3, Eq. (3)"},{"comment":"The violin plots are described as distributions across driver LLMs, but the number of LLMs is only five; the resulting distributions are quite coarse. Consider plotting individual points or using a strip chart to avoid over-interpreting density shapes.","section":"Sec. 3, Fig. 2"},{"comment":"The 'JudgeGap' metric adapts an MT-Bench prompt [31] to medical chest X-ray reasoning, but the validity of an LLM judge for this domain is not established. The paper should include a small human-validation or inter-judge agreement analysis, or at least acknowledge this limitation prominently.","section":"Sec. 2.3, LLM reasoning bias"},{"comment":"Table 2 states 'All values are multiplied by 100,' but the formatting of CIs is inconsistent (e.g., some entries read '0.10[0.30,21.70]', with a misplaced decimal). This is hard to parse and should be cleaned.","section":"Sec. 3, Table 2"},{"comment":"The phrase 'tool-used based agentic frameworks' (also used in the abstract) is grammatically awkward; 'tool-using agentic frameworks' is clearer.","section":"Introduction, Sec. 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a medical-imaging or AI-fairness venue and makes a plausible empirical contribution as a descriptive audit. The main obstacle is that the title, abstract, and Sec. 2.3 promise a 'decomposition' that the metrics do not deliver, and one of the three central metrics (ΔTEB) has a selection-bias problem that is acknowledged but not remedied. These issues are fixable by reframing the claims and adding sensitivity analyses, so I do not recommend rejection, but the revision needs to be substantive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere's the short version: DUCX is a genuinely useful empirical audit of where demographic disparities show up inside a MedRAX-style chest X-ray agent, but the 'decomposition' is more of a descriptive breakdown than a causal attribution. If you read it as 'intermediate behaviors differ across groups in ways end-to-end numbers miss,' it holds up. If you read it as 'we localized the origin of unfairness to specific tools or stages,' it doesn't.\n\nWhat's new: the stage-wise framing—tool exposure, tool transition, reasoning synthesis—is a sensible way to organize process-level fairness checks for agentic medical AI. The empirical sweep is real: five driver LLMs, two datasets, confidence intervals on the end-to-end and reasoning metrics, and code released. The finding that segmentation and visualization tools show bigger exposure-conditioned gaps than classifiers or grounding is the kind of concrete observation that could guide future mitigation.\n\nWhere it gets soft: Eq. (2), the tool-exposure bias, conditions on the agent's own decision to invoke a tool. If the planner routes harder cases from one demographic to that tool, the accuracy gap among exposed instances is a selection effect, not a tool-utility gap. Reporting exposure rates makes the confound visible but doesn't adjust for it. The paper also uses final-answer accuracy as the utility, so downstream reasoning bias is baked in. And the three components aren't additive—they're three different conditional comparisons—so 'decomposition' is a stretch. The transition-bias matrices are averaged over LLMs without confidence intervals, and some of the reasoning-bias gaps (e.g., Qwen3VL hedging) have extremely wide CIs, so treat those as suggestive. The MIMIC-FairnessVQA dataset is generated by LLM with manual verification; useful, but the curation pipeline isn't described in enough detail to fully assess.\n\nNet: this paper is for readers who care about fairness auditing for medical agents. It deserves a serious referee, and with revision (softer attribution language, confound-adjusted exposure metric, error bars on transitions) it could be a solid contribution. My advice: engage.","headline":"A useful empirical audit of process-level fairness in medical agents, but the causal 'decomposition' is not established by the metrics.","tokens_in":11073,"tokens_out":2813,"would_cite":true,"duration_ms":28553,"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":"Tool-using chest X-ray agents exhibit subgroup disparities in tool usage, transition patterns, and reasoning—disparities that end-to-end fairness metrics cannot predict.","keywords":["fairness","medical AI agents","chest X-ray","tool use","LLM agents","fairness decomposition","tool-exposure bias","tool-transition bias"],"falsifier":"For a fixed agent, compare ΔTEB computed from its natural trajectories with ΔTEB computed after forcing both subgroups to use the same tools on matched questions (or after randomizing tool exposure). If the forced-exposure gap is near zero while the natural one is large, the measured 'tool-exposure bias' is really a routing artifact.","tokens_in":10217,"feed_emoji":"🩻","tokens_out":3019,"duration_ms":28497,"temperature":0.7,"pith_summary":"The paper argues that fairness in medical AI cannot be judged solely by comparing final outputs across demographic groups. Tool-using chest X-ray agents—systems where a language-model planner calls specialized vision tools like classifiers, segmenters, and report generators—route different patient groups through different internal decision paths, and disparities appear in those paths even when end-to-end error rates look similar. The authors introduce DUCX, a stage-wise decomposition that splits an agent's end-to-end bias into three components: tool-exposure bias (utility gap when a tool is used), tool-transition bias (differences in which tool sequences are followed), and LLM-reasoning bias (differences in how final answers are phrased). Across five language-model backbones and two benchmarks, they find persistent subgroup gaps in all three stages, with exposure-conditioned utility gaps up to 50%. The point of the paper is that process-level auditing of agent trajectories is necessary to find and eventually correct these hidden disparities.","feed_headline":"Bias hides in X-ray agents' tool choices and reasoning","feed_subtitle":"End-to-end fairness checks miss subgroup gaps in how medical AI agents pick and chain tools; process audits would find them.","key_machinery":"The central object is the stage-wise fairness decomposition. Tool-exposure bias (ΔTEB) is defined as the accuracy gap between subgroups conditioned on the event that a particular tool appears in the trajectory; tool-transition bias (ΔTTB) is the difference between estimated Markov transition matrices of tool-to-tool routing per subgroup; LLM-reasoning bias is the max-min gap across subgroups in response-level features (judge-assigned reasoning quality, hedging-word counts, and demographic-term counts). These three metrics partition the agent trajectory into exposure, routing, and synthesis stages, so that a measured disparity can be mapped to the stage most responsible.","core_discovery":"On the paper's own terms: tool-using chest X-ray agents exhibit distinct subgroup disparities in tool usage, transition patterns, and reasoning traces that are not predictable from end-to-end evaluation alone, and these disparities can be separated into tool-exposure, tool-transition, and LLM-reasoning bias. The key claim is that end-to-end fairness metrics (accuracy gap, demographic parity, equalized odds) average over heterogeneous trajectories and therefore conceal where unfairness originates. The decomposition shows, for example, that conditioning on segmentation-tool availability yields subgroup accuracy gaps as high as 50% even when overall accuracy gaps are small.","pith_inferences":["If the decomposition holds, fairness audits of medical agents should log and analyze trajectories (tool calls, transitions, reasoning text), not just final predictions.","The confounding of tool exposure with agent routing suggests a sharper test: randomizing or counterfactually re-assigning tool access across groups could verify whether the exposure gap is a property of the tool or of the selection policy.","The findings likely generalize beyond chest X-rays: any LLM-driven pipeline that chooses among specialist modules could harbor stage-level demographic bias.","A testable extension: applying stage-targeted debiasing (e.g., equalizing transition matrices or exposure rates) should reduce end-to-end equalized odds more than outcome-level postprocessing alone."],"forward_implications":["End-to-end fairness metrics alone understate or miss subgroup disparities in tool-using medical agents.","Disparity can originate at any of three distinct stages: which tool is used, which tool chain is followed, and how the final answer is phrased.","The stage in which disparities concentrate shifts with dataset and sensitive attribute (e.g., segmentation tools on one benchmark, visualizer on another).","Different reasoning-bias measures are independent: a model can show no demographic framing gap yet have large hedging gaps.","The decomposition provides modular intervention targets for debiasing agentic medical systems."],"fun_headline_variants":["X-ray AI bias hides in tool choices, not just outputs","Medical AI fairness misses tool-use disparities","Tool-using X-ray agents conceal 50% subgroup gaps","End-to-end fairness fails to catch agent tool bias","Process audits reveal hidden bias in X-ray agents"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The tool-exposure metric assumes that comparing accuracy only on cases where the agent already chose to use a tool isolates the tool's own bias, but the agent's choice itself depends on query difficulty and subgroup, so the gap can reflect routing selection rather than the tool's utility.","fun_headline_variants_meta":{"raw":{"variants":["X-ray AI bias hides in tool choices, not just outputs","Medical AI fairness misses tool-use disparities","Tool-using X-ray agents conceal 50% subgroup gaps","End-to-end fairness fails to catch agent tool bias","Process audits reveal hidden bias in X-ray agents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":1049,"prompt_tokens":787,"completion_tokens":262,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":186}},"tokens_in":531,"tokens_out":262,"duration_ms":3420,"temperature":1.0,"reasoning_tokens":186,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:48:00.876191+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed agent, compare ΔTEB computed from its natural trajectories with ΔTEB computed after forcing both subgroups to use the same tools on matched questions (or after randomizing tool exposure). If the forced-exposure gap is near zero while the natural one is large, the measured 'tool-exposure bias' is really a routing artifact.","supporting_citations":[],"review_version":1}