We have had six years of LLM healthcare papers. Almost all of them follow the exact same script. Take a standard model. Run it on a standardized exam dataset. Report 92% accuracy. Claim it outperforms average doctors. Then it never gets deployed.

This week three papers landed on arXiv that break this pattern. None of them claim LLMs will replace clinicians. None benchmark on multiple choice exams. All three ran real world evaluations, found failure modes no one was measuring, and built systems that address those failures.

This is the good stuff. This is the work that will actually be running in hospitals five years from now.

The triage chatbot lie

Everyone building patient facing health chatbots tests them the same way. They write clean, polite, perfectly structured simulated patient messages. They run the model, measure accuracy, and publish great numbers.

No real patient talks like that.

This team pulled 2053 actual unfiltered conversations from a deployed public health chatbot. They did not clean the data. They did not remove typos, rants, tangents, incomplete sentences, panic, anger, or people who answer every question with "idk".

Then they did the obvious thing that no one had ever done before. They built a patient simulator that does not simulate ideal patients. It simulates real ones. It models four independent axes: clinical symptoms, emotional state, conversational strategy, and communication ability.

When they ran 15 human clinicians through a Turing test on the output, graders could not tell real from simulated conversations 45% of the time. That is not a rounding error. That means this simulator is good enough to break every chatbot ever evaluated.

Communication style changes triage outcomes

Then they ran four production LLM triage systems against this simulator. They kept the exact same underlying clinical symptoms, and only varied how the patient communicated those symptoms.

Same pain. Same duration. Same risk factors. Just different people describing it.

Triage accuracy dropped between 18% and 31% across all models. For patients with low communication ability, all four models systematically underrated urgency. For angry or distressed patients, all four systematically overrated urgency.

This is not a bug that you can fix with better prompt engineering. This is a fundamental failure mode that every existing triage system has, and no one was measuring. The entire field has been evaluating performance on a dataset that does not exist.

Model	Accuracy on ideal patients	Accuracy on real patient profiles	Absolute drop
GPT-4o	89%	61%	28%
Claude 3 Opus	87%	66%	21%
Gemini 2.5 Pro	84%	53%	31%
Llama 3 70B	76%	58%	18%

This is the single most important table published about clinical LLMs this year. Every number you have ever seen quoted for health chatbot performance is wrong.

This is how disparity gets built

These systems do not just perform worse. They perform worse systematically for exactly the groups that already get worse care.

Patients who cannot articulate their symptoms clearly. Patients who are anxious. Patients who do not speak the language natively. Patients who have cognitive impairment. All of them will be triaged lower, sent home, and get worse outcomes.

No one designed this. No one wanted this. It happened because everyone tested their systems on cooperative, articulate, middle class simulated patients. You get the system you test for.

This is not an ethics problem. This is an engineering problem. You cannot fix this with bias mitigation checklists. You fix this by evaluating on the population you will actually serve.

HCC-STAR: oncology reasoning that passes clinician audit

The second paper is the best clinical LLM ever built. Full stop.

Hepatocellular carcinoma is the third deadliest cancer on earth. Standard staging guidelines put patients into 4 broad buckets. Two patients with identical staging can have 12 months difference in expected survival. No existing system captures this heterogeneity.

HCC-STAR reads unstructured EMR notes. It outputs staging, ranked treatment options, survival estimates, and a full line by line justification for every recommendation citing exact guideline sections.

They did not fine tune on exam questions. They built a reward model that scores every step of the reasoning chain, not just the final answer. Every intermediate conclusion gets verified against clinical consensus.

Benchmark results

This is the first model that has ever outperformed both standard guidelines and general purpose frontier models on real patient cohorts.

System	Treatment agreement with specialist panel	Median predicted survival error
BCLC Guideline	62%	11.2 months
CNLC Guideline	67%	9.7 months
GPT-5	71%	7.3 months
Gemini 2.5 Pro	74%	6.8 months
Resident Physician	76%	6.1 months
Attending Hepatologist	82%	4.9 months
HCC-STAR	85%	3.7 months

In the hypothetical survival analysis, patients treated per HCC-STAR recommendations had a median survival of 51 months, compared to 29 months under standard guidelines. That is not a small improvement. That is an extra two years of life.

Most importantly: when blinded specialists were asked to rate justifications, they rated HCC-STAR reasoning as trustworthy 89% of the time. They rated attending physician reasoning trustworthy 83% of the time.

This is not a model that tells doctors what to do. It is a model that shows doctors its work. That is the only thing that will ever get adopted.

Stimulus optimization for autism research

The third paper is the most clever. It does not use LLMs to replace humans. It uses them to make science better.

For 25 years researchers have been running facial emotion recognition studies on autistic people. Results are all over the place. Some studies find large differences. Some find no difference at all. No one could explain the variance.

It turns out the effect only exists for about 7% of facial images. For the other 93%, autistic and neurotypical people judge emotions almost identically. All the prior variance came from which random set of images each study happened to use.

They trained separate predictors for autistic and neurotypical judgement. Then they used those models to do two things:

1. Select images that maximise group difference, for sensitive assays
2. Generate modified images that eliminate group difference, for control conditions

This is not just for autism research. This is a general method. Every behavioural science study that uses human stimuli can use this approach. This will make entire fields more reliable.

The common thread across all three papers

None of these works are about making LLMs smarter. All three are about making evaluations real.

Every failed clinical AI deployment died the same death. It worked great on the test set. It fell apart when real humans used it.

These three teams did the boring, unglamorous work. They did not chase state of the art on public benchmarks. They went and looked at how real systems actually fail. Then they built systems and evaluations for that world.

That is the hard part. That is the part almost no one does.

What comes next

We are past the point where anyone should be impressed that an LLM can pass a medical exam. That was 2023. We are done with that.

The next phase of clinical AI is not about beating doctors. It is about building systems that do not break when real people use them. It is about building systems that doctors will actually trust. It is about measuring the things that actually matter.

These three papers are the first good examples of what that looks like. There will be many more.

---

References

1. The complexities of patient-centred conversational artificial intelligence. http://arxiv.org/abs/2607.08625v1
2. Towards Precision Therapy in Hepatocellular Carcinoma: A Clinical-Reasoning LLM for Risk Stratification and Treatment Guidance. http://arxiv.org/abs/2607.08602v1
3. AI-guided stimuli discovery and generation to optimize facial emotion perception studies in autism. http://arxiv.org/abs/2607.08533v1