BALAR: The Art of Diagnosis—When AI Learns to Ask the Critical Question Like a Master Physician
Paper Information
- Title: BALAR: A Bayesian Agentic Loop for Active Reasoning
- Authors: Aymen Echarghaoui, Dongxia Wu, Emily B. Fox (Stanford University)
- arXiv: 2605.05386
- Released: 2026-05-06
- AR-Bench-DC (detective reasoning): +14.6% accuracy
- AR-Bench-SP (situation puzzles): +38.5% accuracy
- iCraft-MD (clinical diagnosis): +30.5% accuracy
- No fine-tuning required—purely a prompt-based outer loop
- Task-agnostic
- Dimension 1: vascular involvement? {vascular, non-vascular}
- Dimension 2: onset pattern? {episodic, chronic, acute}
- Joint state space: 2 × 3 = 6 possibilities
- "yes, definitely": 0.91
- "maybe": 0.07
- "no": 0.02
Key Results
Feynman-Style Deep Dive
🎭 Opening: The Question That Was Never Asked
Imagine walking into a clinic and telling the doctor: "Doctor, I've been having headaches lately."
A mediocre doctor would immediately write a prescription—maybe symptomatic treatment, maybe a generic painkiller. But a great physician would put down the pen, look you in the eye, and ask: "How long has it hurt? Is it throbbing or stabbing? Any nausea or vision changes?"
These questions aren't stalling. Each one narrows the space of possible diagnoses—from the vague entry point of "headache" toward one concrete answer among "migraine," "tension headache," or "hypertensive crisis."
What BALAR does is teach AI to become that great physician.
🔍 Act 1: Why Current AI Doesn't "Ask Good Questions"
Current LLMs are passive in interaction. You give a prompt, it gives an answer. It's like walking into a pharmacy, describing your symptoms, and having the pharmacist recommend drugs directly—without ever asking follow-ups.
This isn't because the models "don't want to ask," but because: 1. Tree-of-Thoughts (ToT): focuses on internal search over reasoning trees, not interactive questioning 2. Uncertainty-of-Thoughts (UoT): simulates interaction trees but at enormous cost, without an explicit probabilistic state model 3. CollabLLM: requires RL fine-tuning—expensive and training-dependent 4. MediQ-Expert: fixed pipeline, no formal model of user intent
The result: state-of-the-art LLMs score far below the oracle that holds all private information on AR-Bench. They lack a *principled mechanism for asking questions*.
🧠 Act 2: Bayesian Reasoning—Treating Uncertainty as a Map
BALAR's core is an old mathematical framework: Bayesian inference.
Imagine you're a detective facing an ambiguous clue. You don't guess the culprit directly—you first build a "possibility space."
BALAR models user intent as a latent discrete variable θ, distributed over a structured product space. Each dimension captures one facet of possible ambiguity.
For the "headache" example:
During "sleep time" (before interaction begins), BALAR initializes this belief state via parallel LLM calls: 1. Step 1 – Propose dimensions: the LLM proposes ambiguity dimensions 2. Step 2 – Prior elicitation: assign probability labels to each dimension value (e.g., "likely," "neutral," "unlikely") 3. Step 3 – Generate questions: generate clarifying questions, each linked to multiple dimensions 4. Step 4 – Build likelihood tables: parallel LLM calls fill in the table of P(answer | state)
💡 Act 3: Mutual Information—Measuring "Question Value" Mathematically
With a belief state and likelihood tables, how does BALAR pick the next question?
The answer: maximize Expected Mutual Information (EMI).
Mutual information I(θ; Y) measures how much our uncertainty about state θ decreases once we know the answer Y.
It's like a detective mentally calculating before interrogating a witness: "If this question yields answer A, how many suspects can I eliminate? If answer B, how many?" Then maximizing the expected "information gain."
BALAR provides a theoretical guarantee: under conditional independence assumptions, greedy MI maximization satisfies
Gk ≥ (1−1/e)G*
where Gk is the greedy strategy's expected information gain after k rounds and G* is that of the optimal adaptive strategy. (1−1/e) ≈ 63%—a classic approximation guarantee meaning the greedy policy captures at least 63% of the optimal solution's value.
🔄 Act 4: Soft Bayesian Updates—When Witnesses Don't Follow the Script
Users won't give you discrete multiple-choice answers. They say: "Hmm… it's kind of throbbing, I guess, but not too bad—mostly when I'm tired from work."
BALAR uses a "Soft-Map" LLM to map free-form answers to a probability distribution over discrete choices, then updates the Bayesian belief with this soft observation.
For example, if the user says "yes, it really hurts," the soft-map might produce:
🌱 Act 5: Dynamic State Expansion—When Existing Dimensions Aren't Enough
Sometimes the initial dimension space is insufficient. Midway through an interview, you realize you also need to consider "family history" or "medication side effects"—dimensions absent from the original belief state.
BALAR detects this via an Entropy Gap Criterion:
If the gap between the current belief's entropy and a target entropy cannot be closed by the best possible questions within the remaining interaction rounds, the system triggers an EXPAND operation: 1. Generate new dimensions 2. Elicit priors for the new dimensions 3. Expand the belief state 4. Recompute likelihood tables 5. Generate new questions
It's like a detective discovering a new suspect trait ("left-handed") mid-investigation and adding "handedness" to the suspect feature dimensions.
📊 Selected Experimental Results
| Benchmark | BALAR | Best Baseline | Improvement | |------|-------|---------|------| | AR-Bench-DC | ~72% | ~63% | +14.6% | | AR-Bench-SP | ~65% | ~47% | +38.5% | | iCraft-MD | ~58% | ~44% | +30.5% |
AR-Bench-DC consists of 100 detective cases, each with 5 suspects; the agent must identify the culprit through questioning. AR-Bench-SP contains situation puzzles requiring creative thinking. In iCraft-MD, the agent is a doctor and the user is a patient.
🎬 Closing: Back to the Clinic
BALAR's story tells us: good AI interaction isn't about "answering faster"—it's about "asking smarter."
It requires no fine-tuning and no domain-specific training data. It's simply an elegant Bayesian outer loop that uses the LLM as a component of a probabilistic reasoning engine.
Like an experienced physician facing vague symptoms, it knows what to ask, how to update its judgment, and when to introduce new dimensions of examination.
As the paper states: "The bottleneck is not the raw reasoning capacity of the LLM, but the absence of a principled mechanism..."
The bottleneck was never the model's intelligence—it's how that intelligence is used.
---
References
1. Echarghaoui, A., Wu, D., & Fox, E. B. (2026). BALAR: A Bayesian Agentic Loop for Active Reasoning. *arXiv preprint arXiv:2605.05386*. 2. Zhou, et al. (2025). AR-Bench: Active Reasoning Benchmark. *NeurIPS 2025*. 3. Li, et al. (2024). iCraft-MD: Interactive Clinical Diagnosis. *ICML 2024*. 4. Horvitz, et al. (1984). PATHFINDER: Decision-theoretic diagnosis. *Proceedings of the AAAI*. 5. Yao, et al. (2023). Tree of Thoughts. *NeurIPS 2023*. 6. Hu, et al. (2024). Uncertainty of Thoughts. *ICLR 2024*. 7. Wu, et al. (2025). CollabLLM: Fine-tuned proactive dialogue. *ACL 2025*.
---
*Feynman-style deep dive | Auto-collected 2026-05-09*