AI Consilium: Benchmarking LLM Medical Second Opinions

Background and Context

Large language models can already list a differential diagnosis for a hard case. A useful second opinion asks for more. It must be grounded in facts, coherent in its clinical logic, actionable for the next step of care, and safe. Most existing evaluations check only whether the final diagnosis is correct. That metric says little about whether a report would help a clinician decide what to do on Monday morning.

The AI Consilium paper, posted on arXiv, proposes a benchmark built around this gap. It models a multidisciplinary consultation, the kind of case conference where several specialists review a patient and agree on a plan. The authors use 300 NEJM-MedQA cases with confirmed final diagnoses. They split the set in two. 110 cases serve for physician reference scoring and for calibrating the automated judge. The other 190 cases are held out for the formal evaluation.

The paper asks several linked questions on the same shared cases. How large are the capability differences between models? Does multidisciplinary synthesis beat a single specialty opinion? Does a structured clinical reasoning requirement help? How do the discussion dynamics behave during the consultation? How much reasoning effort does each model need? Putting all of these into one public case set is the main contribution of the work.

Deep Analysis

The design uses eight single-model pipelines. In each pipeline, one model plays every specialist role. Each role first writes its own opinion. The roles then enter an interactive consilium stage. They review each other's work over several rounds until they reach consensus, and the pipeline outputs a final report. Because one model fills every role, differences between pipelines can be traced to the model itself and not to a mix of models. Each pipeline runs under two conditions. In the free-form condition, the model writes without a fixed format. In the structured condition, the model must organize its output along a fixed clinical reasoning framework. A physician-calibrated LLM judge then scores the output. The team first aligned the judge with physician reference scores on the 110 calibration cases. They then used it on the 190 held-out cases. The criteria include factual grounding, reasoning coherence, problem prioritization and management. The judge scored both the final reports and, under the structured condition, the primary-role output produced before the consultation. That second set lets the authors isolate the gain from multidisciplinary synthesis. In total the judge scored 3,040 final reports.

Which matters more for second-opinion quality: the model or the answer format? In the AI Consilium results, model identity separates quality more than answer format does. Swapping the underlying model moves the scores further than switching from free-form to structured output. Format is not irrelevant, though. Structured final reports earn a higher overall score in seven of the eight pipelines. The clearest gain appears in management, where structured reports average 2.740 against 2.436 for free-form reports. The structure therefore helps most on the actionable question of what to do next. It does not lift every criterion evenly. Does the consultation step itself add value? The authors compare the full structured synthesis with the structured primary-role opinion written before the consultation. The synthesis scores higher on all five shared criteria. The largest improvements come in problem prioritization, factual grounding and reasoning coherence. This supports the claim that role-to-role review corrects ranking bias and factual slips that a single viewpoint leaves in place. At the model level, GPT-5.6-Sol ranks first on every evaluated component. It also needs the fewest review rounds to reach consensus. The abstract gives no finer ablation figures, so readers who need them should consult the full paper.

Industry Impact

For teams building medical AI, the first lesson concerns evaluation. Splitting a second opinion into grounding, coherence, actionability and safety is closer to clinical use than a diagnosis accuracy score. Scoring with a judge calibrated on physician ratings also makes this kind of review repeatable at scale. The benchmark gives developers a common yardstick. They can compare models, output formats and collaboration schemes on the same 190 held-out cases.

The second lesson is practical. If a team wants better output from a multi-specialty system, the choice of base model is likely to pay off more than a change of output format. The structured format still earns a place where the system must give management advice, given the 2.740 versus 2.436 gap. A sensible deployment would therefore pick the strongest base model first. It would then add structured reasoning to the steps that produce treatment and follow-up recommendations.

The third lesson concerns multi-agent design. The comparison before and after the consultation shows that letting roles review each other improves prioritization and factual grounding. The extra compute is not wasted. The number of review rounds needed for consensus also matters. GPT-5.6-Sol reached agreement in the fewest rounds, so discussion length and reasoning cost should be tracked as engineering variables and not treated as fixed overhead.

Outlook

The authors' own design sets the limits of the claims. The cases come from NEJM-MedQA and all carry confirmed diagnoses. Real clinical work often lacks that clarity. The scoring relies on an LLM judge. Physician calibration on 110 cases reduces the risk of drift, but it does not equal prospective validation in a live clinical setting. The results show how models behave on a curated benchmark. They do not show patient outcomes.

The single-model design is a strength for clean comparison, and it also leaves a question open. The paper evaluates pipelines in which one model plays all roles. It does not report on mixed-model panels, in which different models take different specialties. A natural next step is to test whether such panels beat the best single-model pipeline. The same shared case set would make that comparison straightforward.

For now, AI Consilium offers a reference point. It shows that model choice dominates format, that structure helps most on management, and that multidisciplinary synthesis beats the pre-consultation primary opinion on all five shared criteria. Future work should add prospective testing with clinicians in the loop and report the cost of each review round. Those two additions would show whether the measured quality gains hold up at the bedside and at an acceptable price.