Greedy, but Not Blind: Teaching Optimization to Listen

Budget meetings have a familiar rhythm.

Someone brings the spreadsheet. Someone brings the map. Someone else brings the sentence that ruins the spreadsheet: “This district looks inefficient on paper, but the roads are worse than the data says.”

Classical optimization knows what to do with numbers. It does not naturally know what to do with that sentence. In public health planning, infrastructure rollout, retail site selection, and ESG investment, those sentences are often where the real institutional knowledge lives. Unfortunately, once the sentence enters the room, the algorithm usually leaves through the back door. Or worse, the organization pretends the sentence has been “encoded” into a weight, because apparently all human judgment becomes rigorous once it is multiplied by 0.37.

The paper behind today’s article, Health Facility Location in Ethiopia: Leveraging LLMs to Integrate Expert Knowledge into Algorithmic Planning, proposes a more disciplined handoff.¹ Its framework, called LEG — Large language model and Extended Greedy — does not ask an LLM to directly choose where health facilities should go. That would be the glamorous version. It would also be the dangerous version.

Instead, LEG gives the LLM a narrower job: adjust high-level district allocations in response to natural-language advice. The actual grid-cell choices are still made by a constrained greedy algorithm, GuidedGreedy, which preserves a tunable coverage guarantee. The interesting contribution is not “LLMs can help optimization.” That slogan is too cheap. The interesting contribution is that the LLM is allowed to listen, but not allowed to drive off the cliff.

The real problem is not facility location; it is controlled surrender

The application is concrete. Ethiopia’s Ministry of Health is upgrading selected health posts into more comprehensive facilities that can provide additional essential services, including childbirth and postnatal care. Budgets are limited, so the planner must decide which facilities to upgrade. The paper models this as a coverage problem: choose a limited number of locations so that as much population as possible falls within a service-accessibility region.

That much is familiar. Facility location has a long optimization history. The paper uses a monotone submodular coverage function, meaning each new facility adds population coverage, but the marginal benefit can shrink when nearby populations are already covered. This structure is friendly to greedy algorithms: repeatedly pick the cell with the highest marginal coverage gain, and one gets the classic $(1 - 1/e)$-style approximation behavior.

But the actual planning problem is not merely “maximize covered population.” Stakeholders may care about remote districts, underserved communities, terrain difficulty, political feasibility, existing health burdens, or fairness across zones. Some of that can be measured. Some cannot. Some may even conflict.

The paper’s key move is to separate the decision into two levels:

This is the mechanism-first insight. The paper is not saying, “replace optimization with language.” It is saying, “put language at the level where human advice is meaningful, then put optimization at the level where guarantees are enforceable.”

That distinction matters. If an LLM directly selects facilities, it may produce a plausible map that quietly undercovers the population. If a pure greedy algorithm selects facilities, it may maximize coverage while ignoring context that experts consider operationally decisive. LEG is built around a middle position: the LLM influences the allocation, but the solver remains the gatekeeper.

The LLM only moves district budgets, not individual facilities

A likely first misreading of this paper is that the LLM is doing the facility-location optimization. It is not.

The process starts with a standard greedy baseline. That baseline selects grid cells to maximize coverage, then summarizes the result into a district allocation vector. For example, if the greedy solution places three upgraded posts in District A, five in District B, and two in District C, that vector becomes the high-level object shown to the LLM.

The LLM then receives the current allocation, previous allocation, budget, district-level population statistics, initial coverage information, and a set of advice sentences. Its task is tightly constrained: produce a new district allocation, moving only one or two units between districts per iteration, while preserving the total budget and respecting minimum allocation constraints.

This is a useful design choice. Moving one or two units is not glamorous, but it is governance-friendly. It prevents the LLM from hallucinating a radically different plan after one eloquent prompt. It also makes each iteration interpretable: a planner can ask, “Why did one unit move from this district to that district?” That is far easier to audit than a full map of unexplained location changes.

Once the LLM produces the district allocation, GuidedGreedy translates it back into grid-cell choices. This is where the paper’s real engineering lives. The LLM says, in effect, “try giving this district more capacity.” GuidedGreedy then asks, “Can I do that without sacrificing too much coverage?”

This division of labor is the core business lesson. In high-stakes allocation systems, the point is not to make the AI “smarter” in the abstract. The point is to define where the AI is allowed to be subjective.

GuidedGreedy is the guardrail, not a decorative algorithm

GuidedGreedy works by comparing two possible choices at each step.

The first candidate is the unrestricted greedy choice: the grid cell with the highest marginal coverage gain anywhere. The second candidate is the best grid cell that respects the LLM-proposed district allocation. The algorithm follows the district-guided candidate when it is good enough; otherwise, it falls back to the unrestricted greedy candidate.

The “good enough” threshold is controlled by $\beta$. If the district-guided cell achieves at least a $\beta$ fraction of the unrestricted greedy marginal gain, the algorithm can obey the LLM’s district plan. If not, it protects the coverage objective.

The second control parameter, $\alpha$, governs how much of the selection process must retain this greedy-quality protection. The paper’s guarantee can be summarized as:

Here, $f(S)$ is the population coverage of the returned allocation, and $OPT_b$ is the best possible coverage allocation under budget $b$. The larger $\alpha$ and $\beta$ are, the stronger the theoretical coverage floor. The smaller they are, the more freedom the process gives to the LLM-guided allocation.

This is not just a mathematical footnote. It is the governance knob.

A health ministry, retailer, logistics operator, or development agency does not always want the same balance. In one region, maximizing population coverage may be politically and ethically compelling. In another, the priority may be to correct long-term neglect of remote districts, even if the immediate coverage count falls. LEG makes that trade-off explicit rather than burying it inside an opaque “AI recommendation.”

The paper’s mechanism therefore turns a vague governance question — “how much should we listen to experts?” — into a parameterized system design question: “how much coverage loss are we willing to permit in exchange for qualitative alignment?”

That is progress. Not because it eliminates judgment, but because it stops pretending judgment has been eliminated.

Verbal feedback is not decoration; it changes the allocation path

The experiments use three Ethiopian regions: Afar, Somali, and Benishangul Gumuz. The authors use projected 2026 population data, walking accessibility based on a global friction map, and a maximum two-hour walking-distance assumption. To emulate multi-stakeholder guidance, Gemini-2.5-Pro generates 20 advice sentences, organized as four experts with five sentences each and deliberate contradictions. The same stronger model is also used to synthesize an alignment scoring function from 0 to 1. The iterative allocation and feedback steps use Gemini-2.5-Flash.

That setup matters for interpretation. The spatial and population data are real. The expert-advice environment is simulated through LLM-generated advice, and the alignment score is also LLM-generated. So the experiments are best read as a controlled system test, not as proof that real Ethiopian stakeholders would endorse the resulting plans.

The first experiment asks whether verbal feedback is useful. The paper compares two variants:

The pattern is sensible. Quantitative feedback stays closer to the measurable objective, so it tends to preserve more coverage. Verbal feedback helps the system better follow the advice sentences, so alignment improves. The paper reports that this verbal-feedback advantage appears across all three regions, with the largest gap in Benishangul Gumuz.

That is an important result, but it should not be overread. The result does not prove that verbal feedback captures “true human values.” It proves something narrower: given advice sentences and an LLM-generated alignment evaluator, the verbal-feedback loop improves alignment with that advice proxy more effectively than numeric feedback alone.

For business use, that narrower claim is still valuable. Many enterprise allocation problems already rely on imperfect proxy goals. The question is not whether the proxy is metaphysically pure. The question is whether the system makes the proxy visible, adjustable, and auditable. LEG moves in that direction.

The coverage-alignment trade-off is the feature, not the bug

One of the paper’s most useful observations is also the least surprising: as advice alignment rises, coverage tends to fall.

A lazy reading would treat that as a weakness. A better reading treats it as the point.

If expert advice never changes the coverage-maximizing solution, then either the advice is redundant or the evaluation is too blunt. Real stakeholder input should sometimes move the allocation away from pure population coverage. A district with poor health outcomes, weak existing infrastructure, or difficult terrain may deserve priority even if its immediate coverage gain is lower. The optimizer sees fewer people covered. The planner may see a more defensible policy.

Experiment 2 explores this trade-off by varying $\alpha$ while keeping $\beta = 1$. The reported pattern is that larger $\alpha$ values generally produce higher coverage and lower advice alignment. Smaller $\alpha$ values allow more qualitative conformity. The paper notes an Afar-specific exception where $\alpha = 0$ and $\alpha = 0.25$ behave closely enough that a small enforced greedy component offers limited additional practical coverage benefit.

This experiment is best understood as a sensitivity test, not a second thesis. It does not merely say that one parameter changes the output. It shows that $\alpha$ behaves in an interpretable way. That matters because a control knob is only useful if users can predict what turning it will do.

For enterprise readers, $\alpha$ is the kind of parameter that should appear in governance documentation. In a public-sector deployment, it might be described as the minimum protected coverage discipline. In a logistics rollout, it might represent service-level discipline. In a retail expansion model, it might represent revenue or footfall discipline. In an ESG allocation model, it might represent the minimum measurable impact retained while incorporating community priorities.

The name changes. The design pattern does not.

The appendix-style tests are about operational discipline

The paper also examines whether a longer feedback history helps the verbal-refinement loop. It compares using only the previous iteration with using a three-step window. After ten iterations, the difference between the one-step and three-step variants is minimal in both coverage and advice alignment. The authors adopt the one-step setting as the default because it is comparable, slightly more stable, and more computationally efficient.

This is not the headline result. It is an implementation sanity check. But it is useful.

In business AI systems, teams often assume that more context is automatically better. More history, more documents, more conversation, more traces — surely the model will become wiser after being fed the entire attic. Often it becomes slower, more expensive, and only theatrically more informed.

Here, the paper suggests that the feedback loop does not need a long memory window to work. The immediate allocation difference and coverage difference are enough to guide the next iteration. That finding is specific to this experimental setup, but the operational lesson generalizes: before expanding context windows, test whether the extra context actually changes decisions.

The paper’s visual comparison in Afar is another interpretability check. With $\alpha = 0.5$, the allocation leans more toward dense population clusters. With $\alpha = 0$, it better reflects qualitative advice. The overlap between the two settings shows that the system still agrees on core high-priority areas, while the differences reveal what the parameter is doing spatially.

That is exactly the kind of diagnostic planners need. A parameter should not be a mystical slider. It should leave a visible footprint.

What this means for business allocation systems

The direct application is health facility upgrading in Ethiopia. The broader business relevance is resource allocation under competing objectives.

Many organizations face the same pattern:

The Cognaptus inference is not that every allocation problem should use LEG exactly. The deeper point is architectural: separate language interpretation from hard optimization, then define the interface between them.

That interface should answer four questions:

1. What level of decision can the LLM modify?
2. What measurable objective is protected?
3. What parameter controls the trade-off?
4. What feedback lets humans understand why the recommendation changed?

If those questions are unanswered, the system is probably not “AI-aligned.” It is just AI-flavored.

The boundaries are not small details

The paper is careful enough to give us useful boundaries.

First, the advice is generated by Gemini-2.5-Pro, not collected from actual Ethiopian planners for the experiment. The authors are collaborating with Ethiopian public-health institutions, but the experimental advice environment is synthetic. That makes the test cleaner, but it limits claims about real stakeholder acceptance.

Second, the alignment evaluator is also synthesized by an LLM. This is practical, because scoring natural-language advice is difficult. It is also a source of circularity. An LLM generates advice, another LLM turns advice into a scoring function, and an LLM helps improve allocation. That does not invalidate the framework. It means deployment would need human-audited alignment metrics, especially where advice encodes political, ethical, or clinical priorities.

Third, the paper focuses on coverage and advice alignment. Real facility upgrades involve land availability, construction constraints, staffing, supply chains, security, seasonal accessibility, maintenance, and budget execution risk. The authors note that the framework can operate over predefined candidate sets, but a real deployment would still need domain-specific feasibility filters before any allocation becomes a plan.

Fourth, the guarantee protects coverage relative to the formal coverage function. It does not guarantee fairness, stakeholder legitimacy, or health outcomes. A coverage guarantee is valuable. It is not a moral certificate. AI systems remain stubbornly bad at becoming moral certificates, despite a thriving industry in pretending otherwise.

These boundaries do not weaken the paper’s contribution. They clarify what kind of contribution it is: a controlled mechanism for integrating qualitative advice into optimization while retaining a measurable performance floor.

The better mental model: constrained listening

The usual debate around AI in decision-making is poorly framed. One side wants algorithms because humans are biased. The other side wants human judgment because algorithms are blind. Both are correct often enough to be annoying.

LEG suggests a more useful framing: constrained listening.

The optimizer should listen because the formal objective is incomplete. It should be constrained because natural-language advice is inconsistent, hard to verify, and easy to overfit. The LLM should translate and reflect, but not become the final authority. The greedy algorithm should protect coverage, but not pretend coverage is the only value in the room.

That is the central mechanism. Not automation replacing experts. Not experts overriding automation. A structured negotiation between measurable utility and qualitative judgment.

For public health planning, that structure may help ministries explore allocation plans that are both efficient and politically explainable. For businesses, it offers a template for AI-assisted decisions where local knowledge matters but cannot be allowed to casually destroy ROI, service levels, or compliance constraints.

The future of decision AI will not be won by systems that “understand everything.” Conveniently, no such system is currently available, though many demos have excellent lighting. It will be won by systems that know exactly where they are allowed to be flexible and exactly where they are not.

Greedy, in other words, can learn to listen. But someone still needs to keep a hand near the steering wheel.

Cognaptus: Automate the Present, Incubate the Future.