LLM classifiers seem decent but fail quietly

LLMs often perform well at classification tasks, but they lack calibration, threshold control, and interpretability. Their verdicts are hard labels with no principled way to trade off precision and recall.

“LLMs-as-classifiers, prompts applied to a context and returning a label, suck to work with.”

Calibration is broken

LLM confidence scores are not well-calibrated. You can extract token log probabilities or ask for confidence, but neither reliably reflects true probability.

Structured data gets ignored

LLMs excel with unstructured text but often ignore structured data pasted into prompts. You cannot verify if the model actually used the information you provided.

Prior knowledge creates distribution mismatch

LLMs have built-in priors that may not match your data distribution. They won't know if your positive class is rare or enriched, and adapting this context for each population is cumbersome.

Interpretability is an illusion

While prompts are written in prose, you cannot verify what the LLM actually does internally or which parts of your prompt it follows correctly.

Reframe LLMs as feature engineers

Instead of using LLM verdicts as final labels, treat them as input features for a downstream classifier like logistic regression. This unlocks all the properties you actually need.

Logistic regression recovers calibration

Wrapping the LLM verdict in logistic regression produces calibrated probabilities in expectation. You can now choose operating thresholds to trade off precision and recall.

Add structured data as covariates

Logistic regression lets you include structured features alongside the LLM verdict. The model automatically adapts to your training data distribution and covariate structure.

Interpretability improves with features

You still face the black box problem of the LLM itself, but now you can see exactly how much the LLM verdict contributes to your final decision relative to other features.

Improve performance systematically

With LLM-as-classifier, your only option is tweaking prompts. With feature engineering, you can collect more data, screen features, create new features, or swap model architectures.

You need training data anyway

The feature engineering approach requires labeled training data, which breaks the appeal of training-free LLM classification. But you need test data anyway, so the cost is modest.

Debug features empirically

Screen features by checking if they always lead to expected outcomes. Treat features as secondary classification targets and recursively improve them.

Extract richer LLM signals

Use log probabilities of verdict tokens, run multiple LLM passes, or ask for sub-verdicts and intermediate reasoning. Each becomes a feature for your downstream model.

Model architecture is plug-and-play

Start with logistic regression but swap in XGBoost, neural networks, or rule-based systems as needed. The LLM features work with any downstream model.

Test case: irony detection

Using SemEval 2018 Task 3 with 4618 tweets, a simple LLM prompt achieved 0.747 F1 score, beating the competition winner at 0.705. But calibration was poor.

Logistic regression improved calibration

Adding logistic regression on top of the LLM verdict reduced Brier score from 0.259 to 0.175 while maintaining F1 at 0.747. Probabilities now matched empirical proportions.

Nineteen LLM features boosted performance

Asking the LLM to evaluate 19 dimensions of irony (humor, contrast, rhetorical questions, etc.) and adding deterministic features improved F1 to 0.779 and Brier to 0.127.

Matched state-of-the-art results

The feature engineering approach achieved 0.779 F1, overlapping with post-competition LSTM results at 0.786 F1. All using logistic regression on LLM-extracted features.

Coefficients reveal what matters

Logistic regression coefficients show which features the model actually relies on. This transparency helps you understand and debug your classifier.

LLM classification research is growing

More papers use LLMs for classification in research. Treating them as feature engineers rather than classifiers produces higher quality, more reliable results.