Teaching AI How Experts Investigate: A Plant Disease Action Tree Experiment

Better AI answers may need a procedure, not another manual

Language models can already chew through piles of documentation, weigh alternatives and give decent answers. So when a specialized domain needs better results, the reflex is to feed the model more of the same: more manuals, more reference material, more examples, more context.

There's a second kind of expert knowledge that gets far less attention. It isn't what an expert knows. It's what an expert looks at next when they don't know the answer. A small plant disease experiment makes the difference concrete.

Why plant diagnosis isn't image recognition

On the surface it looks like a classification job. One leaf has spots, another is yellow, a third has distorted growth, so name the symptom and name the disease.

Professionals don't work that way. A single symptom can come from a virus, a fungus, bacteria, insects, a nutrient problem, chemical injury, water stress or some other environmental damage. What you see is just an observation. The actual task is to work out what could explain it and which examination would separate those explanations from each other. Plant pathology already has solid reasoning methods for that.

Abduction and deduction

Diagnostic reasoning splits neatly into two kinds of inference.

Abduction starts from what you observe and asks which hypotheses could explain it. This is hypothesis generation:

Observation
    ↓
Possible explanations
    ↓
H1
H2
H3

Deduction then takes each hypothesis and asks what else you should see if it were true:

Hypothesis
    ↓
Expected observation
    ↓
Test / inspection

Whatever that check turns up becomes new information, and the loop starts again:

Observation
    ↓
Abduction
    ↓
Hypotheses
    ↓
Deduction
    ↓
Next observation or test
    ↓
New observation
    ↓
Update hypotheses

That looks a lot like how modern AI agents operate: reason, act, observe the result, continue. Which raises an obvious question. What if you handed an AI this investigation process explicitly?

Turning the process into an Action Tree

The experiment converted the diagnostic routine into a plain Action Tree. A fragment of it:

START
 │
 ▼
Is the host sufficiently identified?
 ├─ No → Get host information
 └─ Yes
      │
      ▼
Are symptoms and signs sufficiently observed?
 ├─ No → Observe them
 └─ Yes
      │
      ▼
Is the damage pattern known?
 ├─ No → Examine spatial distribution
 └─ Yes
      │
      ▼
Is the spread over time known?
 ├─ No → Examine time course
 └─ Yes
      │
      ▼
Generate competing hypotheses
      │
      ▼
What observation would distinguish them?
      │
      ▼
Choose the next action

The tree doesn't replace the model's reasoning, and the model stays free to propose hypotheses. What the tree adds is a frame for deciding what's already established, what remains unknown, which evidence is missing and what to examine next. Think of it as a reasoning policy.

Three setups for the same task

Three configurations were compared. The first is a raw model with only the case:

Case
 ↓
LLM
 ↓
Diagnosis + next action

The second adds expert knowledge, meaning diagnostic reference material, on top of the case:

Case
+
diagnostic reference material
 ↓
LLM

The third gives the model the same knowledge and also the Action Tree:

Case
+
same diagnostic knowledge
+
expert investigation procedure
 ↓
LLM

The comparison that matters is second against third. Both have essentially identical domain knowledge. The only difference is whether the model was merely handed the knowledge or also given a method for using it.

An underwhelming first result

The first case was fairly simple, and everything did well. The raw model, the one with expert knowledge and the Action Tree version all landed on broadly reasonable diagnostic paths. Judged only on final-answer accuracy, the scoreboard would read:

Raw AI ≈ Knowledge AI ≈ Action Tree AI

By that measure the experiment looked like a dud. Something else had changed, though.

What improved first was verifiability

The Action Tree run left a visible trail, an explicit sequence of steps:

Source observation
      ↓
Decision point
      ↓
YES / NO / UNKNOWN
      ↓
Interpretation
      ↓
Current hypotheses
      ↓
Missing evidence
      ↓
Next action

That meant you could inspect the structure of the work and not just the answer. A single step looks like this:

Question:
Is the field-level damage pattern known?

Observation:
Symptoms are present on several plants,
but the overall distribution has not been examined.

Decision:
INSUFFICIENT_EVIDENCE

Next action:
Map symptomatic and asymptomatic plants across the field.

The output wasn't one more block of prose. It was a chain of decisions that someone else can review. The name that stuck for this was a Verification Spine. Two systems can reach the same answer while one is far easier to inspect, reproduce and correct.

Raising the reasoning difficulty

Making the diseases rarer would have been the lazy way to get harder cases. The experiment raised reasoning difficulty instead, in six levels.

The question stopped being whether the AI can answer harder questions. It became how the AI behaves when the evidence gets harder to reason about.

How the three setups behaved across the levels

The results weren't a simple win for the Action Tree.

At the easy levels, all three reached a sensible answer, and the gain was structure and verifiability. Mid-range difficulty is where domain knowledge started to matter. The open question there was whether an explicit procedure could further improve which action the AI picks next and cut down on premature closure. At the top level, success no longer meant finding the answer. It meant seeing that the evidence couldn't support one.

That points to three distinct kinds of benefit:

L0–L1
Structure and verification
        ↓
L2–L4
Reasoning policy and action selection
        ↓
L5
Graceful degradation

One result was steady from the easiest level to the hardest: the Action Tree left behind the cleanest Verification Spine. Even where its final answer was identical to the others, it still laid this out:

Observation
    ↓
Decision point
    ↓
Evidence
    ↓
Hypothesis
    ↓
Missing information
    ↓
Next action

That visibility may end up mattering as much as raw accuracy.

Misleading evidence

One case was built to pass for a viral disease. The plants had virus-like symptoms and aphids were nearby. The quick, intuitive reaction goes like this:

Virus is likely
      ↓
Run a virus test

That isn't a ridiculous call. It's plausible, which is the whole reason the case is hard.

The tree's first question was whether anyone had pinned down how the damage was distributed. Nobody had, so the next step wasn't an expensive molecular test. It was a plain look at how the damage was laid out in the field. The new observations:

Almost all affected plants were inside one spray-width.

Different plant species showed similar damage.

The boundary did not expand over the following days.

Now the picture changes. A transmissible pathogen looks much less convincing, and an abiotic cause such as chemical injury moves up. The tree didn't hand the AI extra knowledge. It changed what the AI decided to look at next, and that may be the more valuable ability.

Knowing versus finding out

This suggests a useful split. Traditional knowledge augmentation asks what experts know. An Action Tree asks how experts find things out. They're different assets.

Picture a specialist facing a hard case. Their skill isn't just a big store of facts. It also includes rules of thumb like these:

• Don't conclude from appearance alone.
• Check the spatial pattern first.
• This observation separates two large groups of causes.
• Hold off on the expensive test.
• This result eliminates several hypotheses at once.
• The evidence isn't enough to decide.
• The case doesn't match known patterns, so escalate it.

That's procedural knowledge, and a lot of it is rarely written down.

When the answer can't be known

The top difficulty level held another surprise. One case was built so that the evidence couldn't separate the last two hypotheses. No further usable test existed in the scenario, and no amount of clever reasoning can recover information that isn't there. The right output wasn't a diagnosis at all:

INSUFFICIENT EVIDENCE

That shifts the question from how far the AI can keep giving correct answers to what it does at the edge of what the evidence supports. Systems can degrade in two quite different ways. The dangerous way:

Correct
  ↓
Correct
  ↓
Plausible guess
  ↓
Confidently wrong

And the graceful way:

Correct
  ↓
Uncertain
  ↓
Insufficient evidence
  ↓
Escalate

In real professional work, the graceful pattern can matter as much as higher accuracy.

A reasoning capability frontier

The experiment hints that the payoff of structured expert procedures shifts with difficulty.

When the problem is easy and the AI already knows the answer, the gains are consistency, structured interpretation, traceability and reproducibility.

When difficulty is moderate and several plausible paths compete, the procedure can help avoid premature closure, spot missing evidence, pick actions that reveal more, and test competing hypotheses against each other.

When the evidence itself falls short, the benefit becomes recognizing the limit, keeping unresolved hypotheses alive, and escalating instead of making up an answer.

So asking whether an Action Tree makes the AI smarter may be the wrong question. A better one is how it moves the model's reasoning capability frontier.

Is this just the old expert system idea?

The objection is easy to raise. If you write decision trees for an AI, aren't you rebuilding the rule-based expert systems of the 1980s?

Not quite. A rigid expert system says something like:

IF A AND B
THEN diagnosis = X

The approach here splits the work. The tree governs which questions get considered, and the language model still does the hard interpretation of messy observations:

Action Tree
→ controls the investigation process

LLM
→ interprets messy observations
→ generates hypotheses
→ recognizes unusual cases

Action Tree
→ checks what remains unresolved
→ guides the next information-gathering action

The tree constrains the process without fully dictating the answer. The combination works only because language models can reason over ambiguous, unstructured input in ways older rule engines never could.

More context has limits

The usual response to a weak AI system is to add context. That will stay important, but returns from piling on documents may diminish.

Another lever is a thing organizations hardly ever record as data, namely the way skilled people work through uncertainty. What gets checked first? What makes them drop an initial hypothesis? Which cheap observation cuts the search space in half? When do they declare the evidence insufficient, and when do they leave the standard procedure and escalate? None of that is plain fact. It's search strategy.

A better question to ask

The starting question was whether an expert Action Tree can make an AI produce better answers. The experiment left a more interesting one: can AI do better work if it's taught how experts reduce uncertainty?

Better work includes the right answer, but also reaching it consistently, exposing the key decisions, choosing stronger next actions, resisting tempting early conclusions, and knowing when the evidence simply doesn't suffice.

The next step in applied AI may not be only teaching machines what experts know. It may be capturing something subtler: how experts find out.