{"id":"03fce284-7f3b-4a82-8671-8fda52c94107","arxiv_id":"2506.06574","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A component-optimized multi-agent system with strong process metrics underperformed simpler configurations in diagnostic accuracy by about 10 points on the same test data.","lead":"A Stanford-led team tested 2,400 real emergency cases of abdominal pain with AI systems that either use one model for everything or split the job among specialized agents. A team built from the best individual components scored better on process checklists but worse on the final diagnosis, so component performance alone is not enough to judge a clinical AI system.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper assumes the Retriever LLM's 100% accuracy on a 20-case pilot transfers to all 1,190 test cases; retrieval failures would be misattributed as Best-of-Breed agent failures and could explain the paradox.","rationale":"The paper has real strengths: a held-out test set, a wide matrix of 34 evaluated systems, released code, and a design that separates process metrics from clinical outcomes. The reader's CONDITIONAL verdict is appropriate. I agree with the reader's weakest assumption: the Retriever is the hidden bridge between component behavior and the failure taxonomy used to explain the paradox. The concern is load-bearing because Section 3.2.5 uses the failure counts to conclude that BoB fails due to agent compatibility, and Section 4 generalizes this into a call for end-to-end validation. If retrieval is imperfect, the same failure counts would arise from shared infrastructure rather than from the interaction of the selected agents. The proposed audit is inexpensive and decisive: it directly checks whether retrieval errors contaminate the comparison. I also noted a secondary concern, that BoB's test-set coverage falls to 0.46, below the 0.5 threshold used to select its information-gathering agent, which complicates the 'excellent process metrics' narrative; but the retriever audit is more fundamental because it determines whether the entire failure attribution is valid. This is a data-integrity and assumption-support issue, not an ad hominem one, and it is testable from existing logs and code. Until that check is run, CONDITIONAL remains the right verdict, so I do not change the reader's recommendation.","tokens_in":15530,"tokens_out":17310,"duration_ms":180972,"concrete_test":"Re-run Phase 2 with retrieval auditing enabled: for every RetrieveResults call, log the requested test or imaging string, the exact patient-record entries matched, and a rule-based ground-truth check of whether the requested entity exists under an alias. Compute per-request and per-system retrieval accuracy on all 1,190 test cases. Then re-run BoB with an oracle retriever that always returns the exact alias-matched data when present. If BoB's 67.7% diagnostic accuracy moves toward the top system's 77.4%, or its insufficient-info, unauthorized-ordering, and hallucination counts drop substantially, the Optimization Paradox is an artifact of retrieval failures rather than agent incompatibility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the shared Retriever LLM (GPT-4o) retrieves requested data perfectly on all 1,190 held-out test cases. Section 2.3 states only that it achieved 100% retrieval accuracy on the 20-case preliminary set used for prompt development, while Appendix 5.3 shows the Retriever is the sole mechanism mapping natural-language test requests to patient records, with 'if a test is not available, state that' as the fallback. Retrieval misses or synonym failures would therefore produce exactly the symptoms attributed to BoB in Section 3.2.5: insufficient information gathering (84 cases), unauthorized test ordering (165 cases), and hallucinated results (165 cases). The conclusion that these failures stem from 'fundamental agent compatibility issues rather than individual component deficiencies' is valid only if retrieval is perfect. That assumption is not secure: 20/20 successes give only a weak lower bound on per-request reliability, the test set spans 480 unique lab tests and 5,959 imaging reports, and BoB requests far fewer tests per case (3.62 total, 1.70 labs) than the top system (6.47 total, 4.25 labs). A single missed retrieval is therefore less likely to be compensated by redundant data in BoB, making a retriever failure a plausible alternative explanation for BoB's low-coverage, downstream-starved error profile.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates single-agent and multi-agent LLM systems for clinical diagnosis on 2,400 MIMIC-CDM abdominal pain cases, with a 20-case pilot, a 1,190-case development set for component selection, and a 1,190-case held-out test set. The authors construct a Best-of-Breed (BoB) system from the development-set winners for information gathering, interpretation, and differential diagnosis, and compare it against 8 single-agent and 26 multi-agent systems on the test set. They report that BoB achieves strong process metrics (e.g., 85.5% lab interpretation accuracy) but significantly lower diagnostic accuracy (67.65%) than a top-performing multi-agent system (77.39%), which they attribute to an 'Optimization Paradox' arising from agent incompatibility and information-flow breakdowns rather than individual component deficiencies.","tokens_in":15802,"tokens_out":4525,"duration_ms":44335,"significance":"If the core finding holds, the paper makes an important contribution to clinical AI evaluation: it would provide a concrete, real-patient-data demonstration that component-level metrics can diverge sharply from system-level diagnostic performance, supporting the need for end-to-end validation. The study's strengths include a sizable pathology-stratified dataset, a clean development/test split for component selection, evaluation across multiple model families, reported process and cost metrics beyond accuracy, and publicly released code. The central claim, however, rests on a test-set-selected comparator and an unvalidated retriever component, both of which currently weaken the causal interpretation that the paradox stems from fundamental agent incompatibility.","major_comments":[{"comment":"The headline comparison of 67.65% vs. 77.39% selects the 77.39% system as the maximum-accuracy multi-agent system on the same held-out test set used for the comparison. Choosing the best of 25 systems on the test set and then applying McNemar's test to that selected pair does not account for the selection process, so the reported p<0.0001 and the 9.75% gap are inflated. The authors should either pre-specify the comparator (e.g., a development-set-selected system), compare BoB against the full distribution of multi-agent accuracies, or adjust for multiple comparisons; this is load-bearing for the central paradox claim.","section":"§3.2.2 and Table 6"},{"comment":"The Retriever LLM is credited with 100% retrieval accuracy based only on the 20-case pilot set, and no retrieval audit is reported on the 1,190 test cases. Because BoB requests substantially fewer tests per case (3.62 total vs. 6.47 for the top system), a single retriever miss is less likely to be masked by redundant data, so the observed failure categories—insufficient information gathering, unauthorized test ordering, and hallucinated test results—could be caused by retriever failures rather than by BoB's agent-incompatibility mechanism. Reporting test-set retrieval accuracy or a sensitivity analysis that treats retrieval misses as a separate error source is necessary to support the causal interpretation in §3.2.5.","section":"§2.3 and Appendix 5.3"},{"comment":"The claim that BoB's failures stem from 'fundamental agent compatibility issues rather than individual component deficiencies' is not uniquely supported by the data. The information-gathering agent was selected for the highest coverage-to-test ratio while maintaining coverage >0.5, a criterion that explicitly rewards low test counts. BoB's low request volume (3.62 tests vs. 6.47 for the top system) could therefore be a direct consequence of the selection metric, not of inter-agent incompatibility. A cross-over experiment—e.g., pairing BoB's diagnosis agent with the top system's information-gathering agent, and vice versa—would help separate component behavior from interaction effects.","section":"§2.5.1 and §3.2.5"},{"comment":"The backbone-composition analysis compares small groups (7 homogeneous, 12 mixed, 6 heterogeneous systems) after excluding one outlier, and the reported p-values do not survive the stated Bonferroni correction. The conclusion that heterogeneous systems face 'inherent diagnostic challenges' is therefore over-interpreted; this analysis should be presented as exploratory, with confidence intervals and an explicit caveat about the small sample sizes.","section":"§3.2.4"}],"minor_comments":[{"comment":"The text says the Best of Breed system was constructed by 'selecting the next best performing LLMs for each component,' which contradicts §2.5.1 where the top-performing agents are selected; the intended construction rule should be clarified.","section":"§2.5.2"},{"comment":"The abstract calls 85.5% 'excellent process metrics' and implies BoB had superior components, but Table 7 shows another multi-agent system (multi_gemini-flash_gpt_gpt) achieving 85.78% lab interpretation accuracy; the wording should be qualified.","section":"Abstract and §3.2.1"},{"comment":"The table note states 'Bold indicates large or medium effect size,' but the table as rendered in the manuscript does not show bold markers; the formatting should be fixed so the note is meaningful.","section":"Table 2"},{"comment":"The row labeled 'gemini_claude' appears to be an inconsistent or incomplete agent-system identifier compared with the naming convention used in Table 6; please correct or explain.","section":"Table 7"},{"comment":"The word 'baseline' is used for the top-performing multi-agent system, which is not a baseline in the usual sense; consider using 'comparator' to avoid confusion.","section":"§3.2.2"}],"recommendation":"major_revision","confidential_remarks":"The central empirical finding is potentially valuable, but the current manuscript's headline gap is inflated by test-set-based selection of the best comparator, and the mechanism claim depends on the unmeasured reliability of the retriever component. If the authors can provide a test-set retrieval audit or a sensitivity analysis, and reposition the comparison against a pre-specified or distribution-based reference, the paper could become publishable at a strong venue. The inconsistent description of BoB construction and the over-interpreted backbone-composition analysis should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this paper has a real empirical result worth taking seriously—a component-optimized multi-agent system lands near the bottom of 26 systems on held-out diagnostic accuracy—but the headline gap is inflated by picking the best test-set system, and a key assumption about the retriever is shaky.\n\nWhat's new and good: the setup is clean. They split 2,380 cases into dev/test, select components on dev, evaluate all systems on test, and report outcome, process, and cost metrics. The error analysis (hallucinated results, unauthorized test ordering, insufficient gathering) is a useful way to look at failure modes. Code and data are available. The call for end-to-end validation rather than component metrics is practical and timely.\n\nSoft spots: the 67.7 vs 77.4 comparison uses the single best multi-agent system from the same test set. Against the average multi-agent system (about 72.3) the gap is roughly 4-5 points, still a real gap but not the 10-point story the abstract sells. The Retriever LLM's 100% accuracy is claimed from 20 pilot cases; if it misses tests on the 1,190 test cases, then BoB's low test counts (3.6 vs 6.5 per case) mean less redundancy, and retrieval failures would show up exactly as the 'insufficient info' and 'hallucinated results' they attribute to agents. The paper doesn't report retrieval accuracy on the test set, so the compatibility conclusion is not secure. The subgroup analysis in §3.2.4 drops the DeepSeek outlier post hoc and doesn't reach corrected significance—minor, but worth noting. The mechanism story—agent incompatibility—is a hypothesis, not proven. The BoB info-gathering agent's selection metric (coverage-to-test ratio) may simply be a poor proxy for information adequacy; that would be a component-level explanation, not a systems-level paradox.\n\nBottom line: the empirical pattern is credible enough to engage with, and the paper is well-organized. This is for people in clinical AI evaluation, procurement, and regulation. The interpretation needs work: pre-specify the baseline, verify retriever performance on the test set, and treat the compatibility mechanism as a hypothesis. It deserves a serious referee, probably with major revision requests.","headline":"Worth reading: the component-optimized 'Best of Breed' system underperforms most multi-agent systems on held-out diagnosis, but the headline gap leans on a cherry-picked baseline and an untested retriever assumption.","tokens_in":16339,"tokens_out":7229,"would_cite":true,"duration_ms":60017,"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":"Building a clinical AI from its best parts made it diagnose worse.","keywords":["multi-agent systems","clinical decision support","optimization paradox","best of breed","component-level evaluation","diagnostic accuracy","large language models","end-to-end validation"],"falsifier":"Re-run the Best of Breed system on the same 1,190 test cases but replace the natural-language retriever with a perfect oracle that always returns exactly the requested labs and imaging reports; if the diagnostic accuracy gap against the top multi-agent system narrows or disappears, then the Optimization Paradox is really a retrieval artifact, not evidence of agent incompatibility.","tokens_in":15335,"feed_emoji":"🏥","tokens_out":3319,"duration_ms":33086,"temperature":0.7,"pith_summary":"This paper sets out to show that optimizing the individual components of a clinical AI system can hurt the system as a whole. It compares single-agent systems, multi-agent systems, and a Best of Breed system assembled from the individually top-scoring agents on 2,400 real emergency cases for four abdominal conditions. The Best of Breed system scored highest on process metrics like lab interpretation accuracy (85.5%) yet delivered significantly lower diagnostic accuracy (67.7%) than a top multi-agent system (77.4%). The paper attributes this gap to incompatibility between agents and breakdowns in information flow, not to weak components, and argues that clinical AI must be validated end-to-end rather than by component metrics alone.","feed_headline":"Best parts, worse diagnosis: AI component optimization backfires","feed_subtitle":"A clinical AI assembled from top-scoring agents lagged 10 points on diagnoses, showing component metrics cannot predict system outcomes.","key_machinery":"The central mechanism is the multi-agent decomposition of clinical diagnosis into three stages: information gathering, information interpretation, and differential diagnosis, coordinated by a sequential orchestrator with explicit handoffs. The Best of Breed construction selects the top single agent for each stage on a development set and then observes its behavior when combined, revealing that process metrics such as coverage-to-test ratio and lab interpretation accuracy do not track diagnostic outcomes. The paper introduces an error taxonomy distinguishing insufficient information gathering, unauthorized test ordering, and hallucinated test results, and uses it to attribute the performance gap to agent compatibility rather than component deficiencies.","core_discovery":"The paper's central claim is the Optimization Paradox: a multi-agent clinical diagnostic system constructed from the best-performing component for each task can underperform systems made of less individually capable agents. In the authors' experiments, the Best of Breed system, built from GPT-4o for information gathering, GPT-4.1 for lab interpretation, and Gemini-2.0-Flash for differential diagnosis, achieved 85.5% interpretation accuracy but only 67.65% diagnostic accuracy, versus 77.39% for the top multi-agent system, a drop of 9.75 percentage points (McNemar's test, p < 0.0001). Error analysis attributes the failure to information-flow breakdowns during agent handoffs: the diagnosis agent received incomplete data (7.06% of cases), attempted unauthorized test ordering (13.87%), and hallucinated test results (13.87%, a 33-fold higher rate than the top system). The authors conclude that component-level metrics cannot predict or validate system-level clinical performance, and that end-to-end evaluation against clinical outcomes is necessary.","pith_inferences":["A testable design principle emerging from these results is that handoff interfaces between agents should be standardized, lossless, and backbone-independent, since the paper's error analysis suggests formatting and communication mismatches drive the failures.","The same paradox likely applies outside clinical settings: any modular system where upstream components are optimized independently may fail if the interface between modules is not jointly optimized with the downstream consumer.","One could extend the paper's approach by measuring, for each agent pair, the mutual information between what the upstream agent produces and what the downstream agent actually uses, to quantify compatibility directly.","A simpler retrospective test of the paper's attribution would be to rerun the Best of Breed system while keeping the same backbone for the interpreting and diagnosis agents, isolating whether the deficit comes from cross-backbone communication rather than from any single component."],"forward_implications":["If the Optimization Paradox holds, then component-level validation of clinical AI systems is insufficient and can be dangerously misleading.","A heterogeneous mix of model backbones in a multi-agent system may degrade diagnostic accuracy even when the same backbones perform well individually.","Process metrics like guideline coverage and lab interpretation accuracy should not be used as proxies for diagnostic correctness in deployment decisions.","Regulatory approval and hospital procurement of modular AI systems should require system-level performance evidence against clinical outcomes."],"supporting_citations":[{"why":"Supplies the MIMIC-CDM dataset of 2,400 real patient cases that all evaluations rely on.","marker":"[11]"},{"why":"Shows agent-based systems matching or exceeding primary care physicians, the backdrop for expecting multi-agent success.","marker":"[5]"},{"why":"Provides a conversational evaluation framework for clinical LLMs that informs the multi-agent system design.","marker":"[7]"},{"why":"Establishes the clinical complexity and risk of acute abdominal pain, grounding why the reported accuracy drop is clinically significant.","marker":"[12]"},{"why":"Appendicitis guideline whose recommendations define the coverage metric for one of the four pathologies.","marker":"[16]"},{"why":"Cholecystitis guideline contributing recommended exam and lab categories to the process metrics.","marker":"[17]"},{"why":"Diverticulitis guideline contributing recommended tests to the coverage scoring.","marker":"[18]"},{"why":"Pancreatitis guideline contributing lab tests that shape the information gathering coverage metric.","marker":"[23]"}],"fun_headline_variants":["Top components, flawed teamwork: clinical AI integration fails","Best parts don't make best system: AI diagnostic paradox","Optimal agents, poor diagnosis: the AI integration trap","Superior pieces, inferior whole: clinical AI paradox","Best-of-breed AI lags behind: integration beats components"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the retriever that fetches patient data from records is 100% accurate on the held-out test cases, based only on its perfect performance on a preliminary set; if retrieval ever fails on the test set, those failures would be misclassified as the information-gathering and hallucination errors that drive the central conclusion.","fun_headline_variants_meta":{"raw":{"variants":["Top components, flawed teamwork: clinical AI integration fails","Best parts don't make best system: AI diagnostic paradox","Optimal agents, poor diagnosis: the AI integration trap","Superior pieces, inferior whole: clinical AI paradox","Best-of-breed AI lags behind: integration beats components"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3699,"prompt_tokens":963,"completion_tokens":2736,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":2656}},"tokens_in":579,"tokens_out":2736,"duration_ms":20701,"temperature":1.0,"reasoning_tokens":2656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:54:17.328336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Best of Breed system on the same 1,190 test cases but replace the natural-language retriever with a perfect oracle that always returns exactly the requested labs and imaging reports; if the diagnostic accuracy gap against the top multi-agent system narrows or disappears, then the Optimization Paradox is really a retrieval artifact, not evidence of agent incompatibility.","supporting_citations":[{"cited_title":"Evaluation and mitigation of the limitations of large language models in clinical decision-making","cited_arxiv_id":null,"evidence_quote":"Supplies the MIMIC-CDM dataset of 2,400 real patient cases that all evaluations rely on."},{"cited_title":"Sara Mahdavi, Christopher Semturs, Juraj Gottweis, Joelle Barral, Katherine Chou, Greg S","cited_arxiv_id":null,"evidence_quote":"Shows agent-based systems matching or exceeding primary care physicians, the backdrop for expecting multi-agent success."},{"cited_title":"Tran, Daniel I Schlessinger, Shannon Wongvibulsin, Zhuo Ran Cai, Roxana Daneshjou, and Pranav Rajpurkar","cited_arxiv_id":null,"evidence_quote":"Provides a conversational evaluation framework for clinical LLMs that informs the multi-agent system design."},{"cited_title":"Macaluso and Robert M","cited_arxiv_id":null,"evidence_quote":"Establishes the clinical complexity and risk of acute abdominal pain, grounding why the reported accuracy drop is clinically significant."},{"cited_title":"Diagnosis and treatment of acute appendicitis: 2020 update of the wses jerusalem guidelines.World Journal of Emergency Surgery, 15:27, 2020","cited_arxiv_id":null,"evidence_quote":"Appendicitis guideline whose recommendations define the coverage metric for one of the four pathologies."},{"cited_title":"Gomes, Andrew W","cited_arxiv_id":null,"evidence_quote":"Cholecystitis guideline contributing recommended exam and lab categories to the process metrics."},{"cited_title":"Hall, Patricia L","cited_arxiv_id":null,"evidence_quote":"Diverticulitis guideline contributing recommended tests to the coverage scoring."},{"cited_title":"American college of gastroenterology guideline: management of acute pancreatitis","cited_arxiv_id":null,"evidence_quote":"Pancreatitis guideline contributing lab tests that shape the information gathering coverage metric."}],"review_version":1}