Generative Evaluation via Internal Model Likelihood

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 136 papers44 conceptsfrozen 2026-10-04

Traditional text evaluation metrics like BLEU or ROUGE rely on surface-level overlaps—counting how many words a machine-generated sentence shares with a human-written one. However, these metrics often fail to capture semantic meaning, coherence, or factual accuracy. A growing body of research suggests a more intuitive alternative: if a highly capable language model "expects" a piece of text (assigns it a high probability), that text is likely to be high-quality.

This shift moves evaluation from a string-matching problem to a probabilistic one, leveraging the internal "world model" of pre-trained transformers to score fluency, consistency, and relevance.

The Intuition: Quality as Probability

The core idea behind generative evaluation is that a well-trained language model acts as a density estimator for "good" language. In a generative framework, the quality of a text segment is quantified by its conditional log-probability under a model.

As Jinlan Fu and Pengfei Liu demonstrate with GPTScore, we can treat evaluation as a generation task in reverse. Instead of asking a model to write a summary, we provide the model with a source and a candidate summary, then calculate the mathematical likelihood that the model would have generated that specific candidate. High-quality text is more "probable" because it aligns with the linguistic and logical patterns the model learned during pre-training.

Beyond Prompting: The "Vibe Check" vs. The Math

While the "LLM-as-a-judge" approach often involves prompting a model to output a numerical score (e.g., "Rate this 1-10"), researchers like Bhuvanashree Murugadoss and Christian Poelitz have found that models don't always follow complex rubrics. In Evaluating the Evaluator, they show that a model's raw perplexity—a measure of how "surprised" the model is by the text—often correlates better with human judgment than the scores the model provides when prompted.

This suggests that the model’s internal probability distribution is a more "honest" evaluator than its prompted outputs. By calculating the likelihood of a response given a specific instruction (e.g., "Is this summary consistent?"), metrics like T5Score and BARTScore can assess specific facets of quality—like style, factuality, or coherence—without needing a human-written reference.

Reference-Free and Context-Aware Scoring

One of the greatest strengths of likelihood-based evaluation is its ability to work in "reference-free" settings. Traditional metrics require a "gold standard" human reference, which is expensive to produce and often limits the evaluator to one "correct" way of expressing an idea.

Generative metrics bypass this by measuring the relationship between the input and output directly:

  • Coherence: Jiwei Li and Dan Jurafsky showed that generative models can measure "discourse coherence" by tracking how likely a sequence of sentences is to appear together in Neural Net Models of Open-domain Discourse Coherence.
  • Controlled Generation: Hao Zhou introduced CTRLEval, which uses masked language models to see how well a sentence fits a specific "control" attribute (like sentiment or topic) by treating evaluation as a fill-in-the-blank task.
  • Visual Alignment: Even in image captioning, Elisa Kreiss and Eric Zelikman found in ContextRef that likelihood-based scores can determine if a description fits both the image and the surrounding text context.

The Narcissism Challenge

Despite their power, these metrics introduce a unique bias. Yiqi Liu and Nafise Sadat Moosavi identified a "narcissistic" bias in LLMs as Narcissistic Evaluators. Because a model's scoring is based on its own internal distribution, it naturally assigns higher probabilities to text that looks like its own training data or its own previous outputs.

This means that if you use GPT-4 to evaluate a summary written by GPT-4, the score may be inflated not because the summary is objectively better, but because the evaluator "recognizes" its own statistical patterns. Understanding and mitigating this "model ego" is a primary focus for the next generation of probabilistic metrics.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_generative_evaluation_via_internal_model_58614545,
  title = {Generative Evaluation via Internal Model Likelihood},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-evaluation-via-internal-model-likelihood/rdr_58614545eca74abba256ab4c3a7ed9a4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_generative_evaluation_via_internal_model_58614545_v1,
  title = {Generative Evaluation via Internal Model Likelihood},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-evaluation-via-internal-model-likelihood/rdr_58614545eca74abba256ab4c3a7ed9a4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}