PaperlayerView exact revision on arXiv

Source-linked AI summary

Do Prompt Optimizers Break Requirements They Are Not Scored On? A Preregistered Study

Sangdo Han, Moonsoo Kim

arXiv:2610.13287v1cs.CLcs.LG

TL;DR

Prompt optimizers target one metric even when prompts contain other requirements that go unscored. This preregistered study tests three optimizers with deterministic, per-item checks and finds that apparent regressions on HotpotQA are largely tied to pre-answer truncation under a tight output cap.

  • Problem

    Prompt optimizers can rewrite requirements such as output format, citations, or abstention while maximizing a metric that typically scores only answer correctness.

    Source: §1 · p. 1

  • Method

    The study preregisters hypotheses and measures, then compares initial and optimized prompts per test item using deterministic checks for unscored requirements.

    Source: §3 · p. 3

  • Results

    14% to 47% of HotpotQA responses end before an answer appears under the 200-token cap, narrowing the preregistered regressions to pre-answer truncation.

    Source: §1 · p. 1, §7 · p. 6

  • Takeaways & Limitations

    Report unparseable and truncated outputs alongside target metrics, and decompose requirement violations before attributing them to individual requirements.

    Source: §7 · p. 6

    The headline result holds for the preregistered 200-token cap; the decomposition and cap-sensitivity analyses are post hoc.

    Source: Limitations · p. 7

Abstract

from arXiv · show

Prompt optimizers rewrite a prompt to raise one target metric, but prompts also state requirements the metric does not score, such as an output format, citations, or abstention. In a preregistered study, we run GEPA, MIPROv2, and TextGrad on HotpotQA, SQuAD 2.0, and IFEval and count per-item violations of unscored requirements with deterministic checks. By the preregistered criterion, all three optimizers produce regressions, substantially on HotpotQA, where net regression in citation validity reaches 10.7 to 41.6 points while exact match rises by 21 to 34 points. A post-hoc decomposition changes this reading. The optimized prompts elicit reasoning before the required JSON object, and under the 200-token output cap used throughout, 13.9% to 47.0% of responses end before any answer appears. Among parseable outputs, net regression is within 3 points of zero. With a 1,024-token cap, extraction failures disappear, content regression stays below 1 point, and exact match rises further. Two simple gates fail the preregistered criterion. The side effect is narrower than hypothesized: optimized prompts are more accurate and break an explicit output contract, and a lenient target metric does not show the cost.

1 Introduction

The paper asks whether single-metric prompt optimization harms requirements the metric does not score, and tests this with preregistered, deterministic checks across three optimizers and tasks.

1 Introduction

  • All three optimizers significantly regress on HotpotQA citation validity and answer length; SQuAD 2.0 improves on unscored requirements, while IFEval shows essentially no target change.These are the preregistered findings; the post-hoc analysis narrows their interpretation.Source: §1 · p. 1
  • Under the 200-token cap, optimized prompts elicit reasoning before the JSON object, and responses ending before it appears are counted as violating every content requirement.Among parseable outputs, there is no content regression.Source: §1 · p. 1
  • With a 1,024-token cap, extraction failures disappear, content regression stays below 1 point, and exact match is higher.The remaining format violation shifts from code fences around the object to a reasoning preamble before it.Source: §1 · p. 1
  • The broad hypothesis that single-objective optimization erodes unscored requirements generally is unsupported; the narrower finding is greater accuracy alongside a broken output contract whose cost depends on output budget.A selection-time filter rejects every optimized HotpotQA candidate, while an in-loop constraint stalls one optimizer entirely.Source: §1 · p. 2

2 Related Work

Prior work evaluates prompt optimizers by target performance, prompt properties, robustness, or multiple objectives; this study instead tracks whether outputs retain requirements omitted from the metric.

2 Related Work

  • GEPA uses natural-language reflection and a Pareto front, MIPROv2 optimizes instructions and demonstrations using a learned objective surrogate, and TextGrad improves system components from textual feedback.These optimizer papers evaluate methods by target metrics on each task.Source: §2 · p. 2
  • PrefPO (Singhal et al., 2026) examines prompt hygiene and prompt hacking, whereas this study measures whether outputs satisfy initial-prompt requirements left unscored.PrefPO evaluates optimized prompt text and its relation to criteria rather than output compliance per test item.Source: §2 · p. 2
  • Multi-objective methods optimize objectives supplied to them, while this study measures effects on requirements standard single-objective optimizers are not given.Examples include ParetoPrompt, MOPO (Menchaca-Resendiz and Klinger, 2025), and MO-CAPO (Büssing et al., 2026).Source: §2 · p. 2
  • Robustness studies vary inputs or model size; this study holds both fixed and examines requirements stated in the prompt itself.Shi et al. (2024) study input perturbations, while Shim et al. (2026) use smaller models as optimizer and solver.Source: §2 · p. 2
  • Prior work finds format restrictions can reduce reasoning performance and tight output limits can cause truncation; this study links those issues to optimized prompts that elicit open reasoning.The reported interaction is between reasoning preambles and a tight output budget.Source: §2 · p. 2
  • Per-example negative flips in image classification motivate a similar view of prompt updates, while this study separates content violations from failures to produce parseable output.The authors report finding no prior study of reasoning preambles against retained format instructions and their interaction with output caps, but note that their search was not exhaustive.Source: §2 · pp. 2–3

3 Experimental Setup

The study evaluates three prompt optimizers on three tasks, compares initial and optimized prompts item by item, and preregisters its hypotheses, measures, tests, and stopping rules.

3 Experimental Setup

  • HotpotQA and SQuAD 2.0 use self-contained QA inputs; test sets add unanswerable items, including HotpotQA questions with supporting paragraphs removed.IFEval is also included, alongside the two QA tasks.Source: §3 · p. 3
  • The optimizer rewrites the whole system prompt and receives only the target metric; IFEval’s unscored quantities are compliance rates across nine instruction groups.The QA prompt states requirements in separate sentences, while IFEval begins with a single instruction to follow every instruction.Source: §3 · p. 3
  • If no JSON object can be extracted, exact match is zero and every applicable content requirement counts as violated.The strict format check and extraction rule were fixed before the runs.Source: §3 · p. 3
  • GEPA, MIPROv2, and TextGrad rewrite instruction text without few-shot examples; each run allows at most 500 task-model calls and explores 1 to 10 distinct prompts.The task model is Claude Haiku 4.5 at temperature 0, and QA output caps are 200 tokens.Source: §3 · p. 3
  • The conditions include an initial-prompt seed, ungated optimization, a selection gate, and an in-loop gate that scores requirement violations as zero and returns them as feedback.The selection gate chooses the best passing candidate or falls back to the initial prompt.Source: §3 · p. 3
  • Net regression is new violations minus repairs divided by all test items, and preregistered hypotheses specify regression, noise-floor, gate, and phrase-retention tests.The hypotheses, metric definitions, tests, and stopping rules were committed and tagged before the main runs.Source: §3 · p. 3

4 Preregistered Results

The preregistered results support regression criteria for all three optimizers, with substantial effects on HotpotQA, while findings on other tasks and requirements are mixed. The section also reports that neither gate meets the preregistered criterion and that regressions occur even when requirement phrases remain in prompts.

4 Preregistered Results

  • Ungated optimization raises exact match substantially on both QA tasks and leaves IFEval essentially unchanged, with mean change between −0.9 and +0.1 points.The initial IFEval prompt scores 0.908.Source: §4 · p. 4, Table 3 · p. 4
  • All three optimizers support H1, with substantial net regression on HotpotQA in citation validity and answer length.On SQuAD 2.0, abstention violations decrease and length is largely unchanged; format improves on both QA tasks because initial strict-format violations leave little room for further regressions.Source: Table 4 · p. 4, §4 · p. 4
  • New-violation rates exceed the noise floor in all 36 HotpotQA run-by-requirement comparisons, versus 4 of 27 on SQuAD 2.0 and 6 of 81 on IFEval.All IFEval cases are in the length-constraint group; the baseline new-violation rate is at most 0.7% on every requirement and task.Source: §4 · p. 4
  • No QA configuration of either gate meets both parts of the preregistered criterion.The criterion required reducing new violations by at least half while keeping target loss within one point.Source: §4 · p. 4
  • Requirement phrases remain in 61 of 63 QA requirement-by-run cases, but 27 of those cases still regress.The substring check cannot detect wording that has been weakened or contradicted.Source: §4 · p. 4
  • In 17 of 54 runs, the test-set target metric did not rise, including 13 IFEval runs and 4 MIPROv2 runs under the in-loop gate.The study completed 54 optimization runs, 27 selections, and 87 evaluations, producing 33,089 test outputs.Source: §4 · pp. 3–4

5 What the Regressions Are

Post-hoc analyses traced most apparent content regressions to optimized outputs that ran out of room before a JSON object appeared. A larger output cap removed extraction failures, but reasoning preambles still broke the JSON-only contract.

Reading

  • The reading of the regression analyses is qualified because they were not preregistered; the authors did not test reasoning in isolation.Whether the format change matters depends on the output consumer and budget: strict parsers fail, while generous budgets let tolerant parsers succeed.Source: §5 · pp. 4–5

Decomposition

  • 29.5% of ungated optimized HotpotQA outputs failed extraction, versus 0.7% for the initial prompt; rates were 27.5% for GEPA, 47.0% for MIPROv2, and 13.9% for TextGrad.Of 1,110 new citation violations, 1,092 (98.4%) were extraction failures.Source: §5 · p. 5
  • Among items where both prompts yielded extractable JSON, net regression was −0.018 to −0.027 for citation, +0.001 to +0.010 for abstention, and −0.002 to +0.004 for length.SQuAD 2.0 had no extraction failures in any condition.Source: §5 · p. 5

Truncation

  • Every unparseable QA output used exactly 200 tokens, the cap; the authors take this as evidence of truncation, though stop reasons were not stored.The lenient exact-match extraction penalizes a preamble only when the JSON object does not fit.Source: §5 · p. 5

Sensitivity to the output cap

  • At a 1,024-token cap, all 4,500 responses were parseable and net regression on citation, abstention, and length ranged from −3.6 to +0.5 points.Exact match rose for every optimizer; these post-hoc results were not significance-tested.Source: §5 · p. 5, Table 5 · p. 5

Format change

  • At 1,024 tokens, optimized outputs still broke the JSON-only instruction, chiefly with reasoning preambles: 38.9% for GEPA, 62.2% for MIPROv2, and 25.7% for TextGrad, versus 2.7% initially.The 90th-percentile output length was 85 tokens initially and 303, 355, and 235 tokens for the three optimizers, respectively.Source: §5 · p. 5

6 Gates

Neither simple gate met the preregistered criterion on the QA tasks. The selection gate often reverted to the initial prompt, while the in-loop gate traded accuracy for stricter format compliance, with results varying by optimizer.

Gate results

  • Neither the selection gate nor the in-loop gate met the preregistered H3 criterion on the QA tasks.The gates were included as reference conditions, not proposed methods.Source: §6 · p. 5

Selection gate

  • On HotpotQA, the selection gate chose the initial prompt in all nine runs because all 21 optimized candidates regressed on validation items under the 200-token cap.Violations returned to the noise floor, but the accuracy gain was lost (EM 0.187).Source: §6 · p. 5

In-loop gate

  • GEPA and TextGrad’s in-loop gates increased strictly valid output and reduced truncation, but cost 6 and 14 exact-match points; new violations fell 28% to 69%, short of the 50% target for most requirements.MIPROv2 returned the initial prompt in all six gated QA runs.Source: Table 6 · p. 6, §6 · p. 6

Selection gate

  • On SQuAD 2.0, the selection gate retained an optimized candidate in five of nine runs; mean EM was 0.497 to 0.612 with the gate versus 0.720 to 0.780 without it.The zero-regression rule was strict enough that one flipped item could reject a candidate on validation sets of 150 to 200 items.Source: §6 · p. 6

In-loop gate

  • At 1,024 tokens, gated prompts had higher format compliance but lower accuracy than ungated prompts: GEPA/TextGrad EM was 0.549/0.428 versus 0.622/0.549.Reasoning preambles occurred in 25.7% and 13.1% of gated outputs, versus 38.9% and 25.7% ungated.Source: §6 · p. 6

7 Conclusion

The conclusion narrows the apparent regressions to a broken output contract and, under a tight budget, answers crowded out by reasoning preambles. It recommends reporting unusable outputs alongside target scores and decomposing violations before interpreting them.

Conclusion

  • On HotpotQA, optimizers were more accurate but elicited open reasoning before the required JSON, crowding out answers in 14% to 47% of responses under a tight output budget.With a larger budget, the same prompts were more accurate and content-requirement regression was below 1 point.Source: §7 · p. 6

Practical implications

  • A tolerant answer extractor can show higher target scores while strict consumers receive unusable outputs, so unparseable and truncated response rates should be reported alongside the metric.Source: §7 · p. 6
  • Requirement-violation counts can be dominated by one upstream failure, so they should be decomposed before being read as evidence about each requirement.Source: §7 · p. 6

Limitations

The findings are bounded by a single-model evaluation, a small search budget and few seeds, and a 200-token cap whose headline effect was not designed to test truncation. The initial prompt’s format failures and coarse measures also constrain what the gates and requirement checks establish.

Limitations

  • The study uses one task model and one provider; reasoning-before-structured-answer behavior may differ across models.The planned open-weight task-model replication was not run.Source: Limitations · p. 6
  • Each run used 500 task-model calls and explored 1 to 10 prompts; intervals based on few runs may miss effects below about 15 points.At an intraclass correlation of 0.05, the minimum detectable net regression is about 15 points.Source: Limitations · p. 6
  • The initial prompt violated the strict format check on over 90% of outputs, mostly through code fences, leaving no room to observe format regression and distorting the in-loop gate.The starting prompt’s in-loop gate score was near zero.Source: Limitations · pp. 6–7
  • The preregistered headline result holds only for the fixed 200-token cap; the decomposition and cap-sensitivity analyses were post hoc.The cap was applied equally to optimization and evaluation, but was not chosen to study truncation.Source: Limitations · p. 7
  • Substantial regression evidence comes from HotpotQA; SQuAD 2.0 shows no regression, while qualifying IFEval rows rest on one to three items.IFEval offered no room for target gains.Source: Limitations · p. 7
  • Deterministic checks do not assess reasoning quality or citation quality beyond validity, and the selection and in-loop gates are simple reference designs.The authors caution that these gates say little about constrained or multi-objective optimizers built for this purpose.Source: Limitations · p. 7

Ethics Statement

The authors report using AI tools in preparing the code, analysis scripts, and manuscript, and state that they reviewed the work and take responsibility. The study uses public datasets and involved no human subjects.

Ethics Statement

  • AI coding assistants helped write the experiment code and analysis scripts, while Claude helped draft and edit the manuscript.The authors reviewed all code, results, and text and take responsibility for the content.Source: Ethics Statement · p. 7
  • The study uses public research datasets; item identifiers and model outputs are available from the authors upon request, but source datasets are not redistributed.Dataset licenses are listed in Appendix A.Source: Ethics Statement · p. 7
  • No human subjects were involved.Source: Ethics Statement · p. 7

A Reproducibility and Deviations

The authors describe the materials and run records available upon request, with tables and in-text numbers regenerable by a single command. They state that judging and analysis choices remained fixed after preregistration, while an open-model replication was not run and several analyses were post hoc.

A Reproducibility and Deviations

Related papers

Loading 2610.13287v1…