Fine-Tuning Changes What Your Model’s Errors Reveal

TL;DR for operators

A fine-tuned model can become only slightly more accurate while its remaining errors become substantially easier to distinguish from correct answers.

That matters when uncertainty scores feed operational controls. If a production workflow accepts an answer, abstains, calls another model, or sends a case to human review according to a detector threshold, fine-tuning changes more than the benchmark score. It can change the detector itself as an operating signal.

Across the settings evaluated by Hu and colleagues, methods based on whether repeated answers converge on the same meaning and methods based on sequence-level confidence improved more consistently after adaptation than token-level entropy measures.¹ Yet probes of the models’ internal representations did not improve consistently. The evidence therefore supports a narrower conclusion: adaptation can make errors more observable through particular behavioral signals without demonstrating that the model has become generally better calibrated or intrinsically more uncertainty-aware.

For teams deploying fine-tuned models, uncertainty controls should be regression-tested after adaptation. Reusing the old detector and threshold without retesting is an unsupported assumption.

Fine-tuning changed error observability more than error frequency

Consider a routine deployment decision. A team fine-tunes a model for an internal question-answering task, reruns the benchmark, and sees a two- or three-point improvement. The task score has moved, but not enough to justify rebuilding the surrounding workflow.

The less visible question is whether the errors that remain still look the same to the system responsible for catching them.

The paper tests that question across LLaMA-3.2-3B-Instruct, Qwen2.5-3B-Instruct, and Mistral-7B-Instruct-v0.3, using TriviaQA, NQ-Open, and SQuAD. The authors compare base models with LoRA, DoRA, PiSSA, and a full-parameter supervised fine-tuning reference.

Accuracy improvements under the parameter-efficient methods were modest: roughly 0.4–3.1 percentage points on NQ-Open, 0.1–1.6 on SQuAD, and 2.6–6.1 on TriviaQA. Detector performance could move much further.

That distinction changes what adaptation evaluation needs to measure. Error frequency tells an operator how often the system fails. Error detectability tells the operator whether those failures can be routed differently.

A model whose accuracy rises only slightly may still support a materially better abstention or review policy if incorrect answers become easier to separate from correct ones.

Detector choice becomes part of the adaptation test

The improvement was not uniform across uncertainty measures.

One class of detectors asks the model the question repeatedly and examines whether the resulting answers converge on the same meaning. If repeated generations concentrate around one semantic answer, uncertainty is lower; if they split across incompatible meanings, uncertainty is higher. Semantic Entropy, SelfCheckGPT, and Degree of Uncertainty belong to this broader behavioral family.

Another family evaluates confidence in the generated sequence itself, including maximum sequence probability and perplexity. These signals also improved relatively consistently after fine-tuning.

Token-level entropy was less dependable. Under the paper’s full-SFT specificity check, for example, gains in Semantic Entropy across the nine backbone-dataset pairs ranged from 5.38 to 14.24 percentage points, SelfCheckGPT from 2.49 to 16.36, and the sequence-confidence measure MSP from 4.53 to 12.81. Mean Token Entropy ranged from a 3.45-point decline to a 10.62-point increase and fell in three of the nine pairs.

The full-parameter SFT arm is useful here as a specificity test. Its qualitative detector-family pattern appeared across all nine backbone-dataset combinations, suggesting that the phenomenon is not confined to low-rank PEFT.

It does not establish that PEFT and full SFT are equivalent. The full-SFT arm uses one untuned configuration per pair, so the experiment cannot rank the two adaptation approaches or show that all SFT recipes produce the same behavior.

Better black-box detection is not evidence of better internal calibration

The strongest interpretive trap is to move from “detectors improved” to “the model understands its uncertainty better.”

The paper tests part of that interpretation by training logistic-regression probes on the models’ hidden states. These probes ask whether correct and incorrect answers become more linearly separable inside the model after PEFT.

The answer is inconsistent.

Probe performance improves on TriviaQA, degrades on NQ-Open, and varies by model on SQuAD. That is difficult to reconcile with a detector-independent claim that fine-tuning has uniformly strengthened an internal uncertainty representation.

The more defensible reading is narrower. Fine-tuning changes model behavior in ways that some external detectors can exploit. The observed black-box gains are real within these experiments, but their existence does not establish universal calibration improvement.

This distinction matters for governance. A team validating an abstention system needs evidence that the chosen signal predicts error under the new model configuration. It does not need to claim that the model has acquired a general faculty of self-knowledge.

Reduced output variability is plausible, but answer shortening is insufficient

The mismatch between strong black-box gains and mixed internal probes creates a mechanism question: what changed in the observable behavior?

Within the PEFT experiments, the authors find an association between stronger multi-sample detection and lower variability in generated-answer length. For LLaMA-3.2-3B on NQ-Open, for example, the base model produced answers averaging 11.92 tokens with a standard deviation of 5.02. The PEFT variants averaged about 5.3–5.4 tokens, with standard deviations around 1.77–1.83.

More regular output can make repeated generations easier to compare. If irrelevant variation declines while differences in answer meaning remain, semantic-consistency measures may separate correct and incorrect answers more cleanly.

But shorter answers alone do not explain the result.

The paper’s few-shot control changes prompting so that the base model produces shorter answers without changing its weights. That intervention alters length and style but does not reproduce the same reduction in variability or the corresponding detector improvements. The control therefore weakens a simple “shorter is better” explanation.

The supported mechanism is more qualified: weight adaptation appears to reduce at least one dimension of output-format diversity, and that behavioral regularization is associated with better multi-sample detection. Length variability is only a proxy, the mechanism analysis covers PEFT rather than full SFT, and the evidence does not establish a universal causal pathway.

Revalidate the routing signal, not just the adapted model

For an operator, the central decision follows directly from the evidence.

Suppose uncertainty determines whether a generated answer is automatically accepted, rejected, passed to another model, or escalated to a reviewer. After fine-tuning, three components need retesting: the model’s task accuracy, the detector’s discrimination performance, and the threshold that converts the detector score into an action.

The paper gives no basis for assuming the third component survives merely because the first improved.

Cognaptus therefore reads the result as an adaptation-governance requirement: treat uncertainty instrumentation as part of the model configuration. A detector validated against the base model is not automatically validated against its fine-tuned descendant.

The study also suggests where testing effort should begin. In comparable short-form factual QA systems, repeated-answer semantic agreement and sequence confidence deserve explicit post-tuning evaluation because they showed the most consistent gains. Token-level entropy should be tested independently rather than inherited as a default.

That is an inference for deployment practice, not a paper result about production ROI. Whether the additional sampling required by multi-generation detectors is economical depends on review cost, error severity, latency budgets, and the volume of uncertain cases in the actual workflow.

The evidence stops at short-form fact-seeking QA

The experiments are substantial enough to support the comparative finding: three open-weight backbones, three QA datasets, three PEFT methods, a full-SFT reference, seven black-box detectors, hidden-state probes, a prompting control, and robustness checks over sampling temperature, low-rank values, and target modules.

Those rank, module, and temperature experiments serve mainly as sensitivity checks. Their value is showing that the primary PEFT detector pattern is not tied to one narrow configuration; they do not extend the result to arbitrary model families or applications.

The remaining boundary is consequential. The study concerns answer-level detection in fact-seeking QA using relatively small open-weight models. It does not establish equivalent behavior for long-form generation, claim-level verification, reasoning, coding, multimodal systems, or interactive dialogue. Nor does it show PEFT superiority over full-parameter tuning.

So the operational conclusion is not that fine-tuning makes models reliably aware of their mistakes.

It is that fine-tuning can change the observability of those mistakes enough that the surrounding control system deserves a new evaluation. For teams already using confidence, abstention, or review routing, that is a separate post-tuning acceptance criterion—one that raw task accuracy cannot substitute for.

Cognaptus: Automate the Present, Incubate the Future.