acceptodds
Under review as a conference paper at ICLR 2027

Bayesian Agentic Medical Diagnosis

Abstract

Efficient sequential clinical diagnosis typically requires maintaining calibrated beliefs over diseases and selecting tests that maximally resolve uncertainty. Current LLM-based diagnostic agents handle both tasks inside the model itself, through accumulated context, verbalized confidence, or prompt sampling, inheriting the LLM's miscalibration and discarding the principled structure of Bayesian decision theory. We revisit this architectural miscalibration and introduce Bayesian Agentic Medical Diagnosis, abbreviated as “BAMD", which confines the LLM to a predictive role while delegating belief maintenance and test selection to an external Bayesian engine. Beliefs are updated via Bayes' rule after each observation, and test selection is governed by information-theoretic operators on the evolving belief state. On four clinical benchmarks, BAMD integrated on an 8B Bayesianized Llama model running on a single local GPU matches or exceeds prompt sampling with a locally deployed 120B LLM across all four datasets, while producing the best-calibrated probability estimates of any LLM-based approach we compare against. These results highlight that the limiting factor in LLM-based sequential diagnosis is architectural, not scale-related, and that a small model embedded in the proper framework can outperform much larger proprietary systems accessed through commercial APIs.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Sequential medical diagnosis, Bayesian decision-making, Large language models, Uncertainty quantification, Active information acquisition

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_zWu3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper combines a Bayesian LoRA predictor with external belief updates, information-gain-based test selection, and value-of-information stopping. Experiments on four binary diagnosis datasets report improved discrimination and calibration over several LLM baselines.

Critical strengths

1.Separating prediction from explicit belief tracking makes the diagnostic process easier to inspect.

2.The MLE and MC-Dropout comparisons share the training data and external controller with BAMD, providing useful evidence about the contribution of Bayesian parameter estimation.

Critical weaknesses

1.The experiments do not establish the central claim that architecture, rather than model scale, limits sequential diagnosis. The evaluation covers four small tabular binary tasks, and the small-versus-large model comparison changes both training and inference procedures. A conventional probabilistic predictor using the same external controller is also missing. The controlled MLE comparison supports a benefit from Bayesian estimation, but not the broader architectural conclusion.

2.The controller does not consistently optimize the utility used to justify continued testing. Equation 13 checks whether any remaining test warrants continuation, whereas Equation 9 selects by information gain per cost. These criteria need not select the same test, and the selected test may have negative net VOI. The paper does not establish conditions that prevent this mismatch.

Questions for authors

1.Can the authors provide a conventional probabilistic predictor using the same training data, observed features, and external controller? This would help distinguish the contribution of the LLM from that of the controller.

2.Whenever Equation 13 permits continuation, is the test selected by Equation 9 guaranteed to have nonnegative net VOI? If not, how often does this mismatch occur in the reported experiments?

Reviewer_UCweOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors introduce a framework for Bayesian Agentic Medical diagnosis. Building on the ACTMED framework, the authors study sequential decision making and evidence accumulation in medicine. Their approach involves finetuning a 8B Llama model and sampling over the model weight posterior. Their method involves separation of model reasoning in the output space and the LLM as a generative model. The authors report performance on 4 different UCI tabular datasets and several ablations of their methods.

Critical strengths

Stepwise medical diagnosis with sequential evidence accumulation is a timely and important problem. The usage of open-source local models is increasingly important in data sensitive applications. The authors test their method using 2 open-source models and provide results on 4 datasets and several ablations of their method. Calibration gains over other methods Practical deployment as model can run on a single GPU

Critical weaknesses

The claim that "the limiting factor is architectural, not scale-related" is unconditional and general, but every demonstrated gain is tied to the small-data, structured-tabular regime, and "not scale-related" rests on a single model scale (8B) plus a cross-study GPT-4o citation. The datasets on which the method is evaluated are very small and only single disease. Could the authors please evaluate their approach on more novel step-wise LLM diagnosis benchmarks e.g. VivaBench, Agentclinic or Abdo-Mimic (Hager et al). Also why was the OSCE style dataset from the ACTMET paper not run?

Questions for authors

In your framework, could the LLM predictive layer be replaced by a natively Bayesian predictor (e.g., a Bayesian network or GP classifier) under the same controller? Given that LA-CDM is evaluated out of its native setting, would the authors either provide a stronger or in-setting RL comparison, or narrow the claim to state that this particular realization underperforms here, rather than that internalizing control is architecturally inferior? Section 4.1 states that an 80/20 train/test split was used. How was this performed with ACTMED that operates in zero shot settings? The performance gains appear to be simply due to fine-tuning a model on parts of the dataset using LoRA which is not very novel by itself. The performance of BAMD without fine-tuning seems to be worse than ACTMED as well. For fairness the model used for ACTMED would have to be finetunes on the test prediction for the task as well. Why are the results on ACTMED for the heart dataset not reported? Figures 2 and 3 are very cluttered and contain a lot of generic objects to fill space. Why use an 8B model? Larger models can be run on Macbooks or small scale GPUs with modern architectures that perform better than GPT-4o on many benchmarks Calling GPT-4o a frontier API is a bit of a stretch and more modern LLMs are missing. GPT 6 luna costs a fraction of the cost calculation in the Appendix and probably should be used as a comparison instead. Can the authors please explain how the model is diagnosing accurately for CKD with an average of 0.23 tests? This seems like an error.

Reviewer_tgAeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Efficient sequential clinical diagnosis typically requires maintaining calibrated beliefs over diseases and selecting tests that maximally resolve uncertainty. Current LLM-based diagnostic agents handle both tasks inside the model itself, through accumulated context, verbalized confidence, or prompt sampling, inheriting the LLM's miscalibration and discarding the principled structure of Bayesian decision theory. We revisit this architectural miscalibration and introduce Bayesian Agentic Medical Diagnosis, abbreviated as ``BAMD", which confines the LLM to a predictive role while delegating belief maintenance and test selection to an external Bayesian engine. Beliefs are updated via Bayes' rule after each observation, and test selection is governed by information-theoretic operators on the evolving belief state. On four clinical benchmarks, BAMD integrated on an 8B Bayesianized {Llama} model running on a single local GPU matches or exceeds prompt sampling with a locally deployed 120B LLM across all four datasets, while producing the best-calibrated probability estimates of any LLM-based approach we compare against. These results highlight that the limiting factor in LLM-based sequential diagnosis is architectural, not scale-related, and that a small model embedded in the proper framework can outperform much larger proprietary systems accessed through commercial APIs.

Critical strengths
  1. This is an Novel framework for the agentic workflow in Medical Diagnosis. 2.. Although the paper looks very rough, its innovative insights can still stimulate some thinking, namely, exploring the effects of different prompts on different models and find out the actual performance of this model under various promots.
Critical weaknesses
  1. Because it requires extracting K=10K=10 posterior weight samples for each LLM query and evaluating candidate tests in each decision round, its evaluation and inference time on local hardware is relatively long.
  2. Unlike some zero-shot methods that do not require task-specific training, this framework requires offline Bayesian fine-tuning based on fully observed patient data with known disease states and test results.
  3. The formula derivations and experimental evaluations in the paper mainly focus on binary diagnostic scenarios (such as healthy/ill, normal/abnormal), which limits its direct application to more complex multi-class or continuous clinical cases.
Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.