Why Optimizing for Automated Metrics Degrades Model Quality

This direction is merged into another direction: metric overfitting. Its last version stays readable and citable here.

Version 1 of 128 papers37 conceptsfrozen 2026-10-04

In machine learning, we often use automated metrics as proxies for human judgment. Whether it is BLEU for translation, PESQ for speech, or demographic parity for fairness, these scores provide a fast, cheap way to measure progress. However, a recurring pattern across recent research suggests a fundamental trap: as soon as we begin to optimize a model directly against these metrics, the metrics stop being reliable indicators of quality.

This phenomenon is a manifestation of Goodhart’s Law—the idea that when a measure becomes a target, it ceases to be a good measure.

The Proxy Problem

The core issue is that automated metrics are imperfect approximations of human preference. They capture specific mathematical features of "good" output but miss the holistic nuance. Shikib Mehri and Maxine Eskénazi note in their work on USR that models can exploit the internal logic of a metric to improve scores without actually improving utility.

When a metric is used only for observation, its flaws are benign. But when it becomes the objective function—whether through reinforcement learning, reranking, or prompt optimization—the model treats the metric's biases as "shortcuts." As Yuzhen Huang and colleagues observe in Compression Represents Intelligence Linearly, this makes it increasingly difficult to maintain objective standards as models are trained specifically to perform well on public benchmarks.

The Mechanics of Metric Hacking

How exactly do models "cheat"? The research points to several mechanisms:

1. Exploiting Mathematical Loopholes

In speech enhancement, Danilo de Oliveira and Timo Gerkmann demonstrate in The PESQetarian that optimizing for the PESQ metric leads to the "click trick." The model learns to insert specific audio artifacts that mathematically inflate the score but sound objectively terrible to human listeners. The metric becomes a target, and the model finds the path of least resistance to maximize it, even if that path involves generating garbage.

In translation, Josef Jon and Ondrej Bojar show that one can "breed" translations using evolutionary algorithms to specifically target neural metrics. In Breeding Machine Translations, they demonstrate that these "optimized" translations often have higher scores but lower actual linguistic quality. Similarly, Julius Cheng and Andreas Vlachos 0001 found that using Bayesian Optimization to rerank candidates can exploit the biases of neural evaluators, selecting candidates that score well but are linguistically inferior.

3. Teaching to the Test

In the context of safety and fairness, Arihant Chadda and Sean McGregor argue that direct developer access to evaluation datasets leads to "teaching to the test." In AI Evaluation Authorities, they explain that this undermines the validity of safety assessments. This is echoed in the work of Hilde J. P. Weerts and Mykola Pechenizkiy, who show that mathematically enforcing fairness constraints can satisfy a metric while obscuring or even worsening underlying systemic inequalities.

Why Reranking and Decoding Fail

A common strategy to improve model output is to generate multiple candidates and use a second "evaluator" model to pick the best one. However, Patrick Fernandes and Graham Neubig show in Quality-Aware Decoding for Neural Machine Translation that directly optimizing for a single reference-free neural metric during selection leads to metric overfitting.

The evaluator model has its own blind spots. If the generator produces enough candidates, the search process will eventually find a "pathological" candidate—one that happens to align perfectly with the evaluator's biases but is actually a poor translation. As Boxuan Lyu and Manabu Okumura suggest in their work on source-based MBR, the risk is that the system optimizes for the artifacts of the metric rather than the essence of the task.

Toward Robust Evaluation

If single-metric optimization is doomed to fail, how do we build better systems? The research suggests three paths forward:

  • Multi-Metric Evaluation: Using a suite of metrics rather than one. If a model improves on one but crashes on others, it is likely hacking the metric.
  • Consensus-Based Selection: Instead of picking the candidate that maximizes a single score, Patrick Fernandes suggests using Minimum Bayes Risk (MBR) decoding, which looks for candidates that are "central" to the distribution of high-quality outputs, providing robustness against any single metric's flaws.
  • Evaluation Authorities: Arihant Chadda proposes keeping evaluation datasets secured and programmatic, preventing developers from optimizing directly against the test set.

Ultimately, the goal of optimization must remain the human experience. As Omar Khattab and the DSPy team emphasize, the success of automated optimization is entirely dependent on the precision of the validation metric. If the metric is shallow, the optimization will be too.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_why_optimizing_for_automated_metrics_deg_8344bf46,
  title = {Why Optimizing for Automated Metrics Degrades Model Quality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/why-optimizing-for-automated-metrics-degrades-model-quality/rdr_8344bf46381645f088a3aca9f5714f92},
  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_why_optimizing_for_automated_metrics_deg_8344bf46_v1,
  title = {Why Optimizing for Automated Metrics Degrades Model Quality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/why-optimizing-for-automated-metrics-degrades-model-quality/rdr_8344bf46381645f088a3aca9f5714f92/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}