Change a few words in a prompt or reshuffle the test questions, and one AI benchmark score can swing wildly.
How much do prompts and data splits shift a single benchmark score?
This explores how far a single benchmark number can move depending on how the prompt is worded and how the test data is sliced. The corpus has much more on the prompt half than on the data-split half.
This explores how far a single benchmark number can move depending on how the prompt is worded and how the test data is sliced. The corpus has much more on the prompt half than on the data-split half. On prompts, the movement can be huge, and it doesn't always go the direction you'd expect. In one production case, a prompt mutation raised a rationale-alignment pass rate from 23.1% to 80.0% simply by adopting the judge's preferred vocabulary, while the precision on actual defects stayed flat (Can prompt optimization accidentally teach judges to reward the wrong signals?). The score moved a lot, but what it was supposed to measure didn't move at all.
The same prompt can also help one model and hurt another. A 23-prompt test across 12 LLMs found that rephrasing and background-knowledge prompts boosted cheap models, while step-by-step reasoning prompts reduced accuracy in the high-performance ones (Do prompt techniques work the same across all LLM tiers?). So a score is never just "the model's ability". It is the model plus the prompt it happened to be given. The prompt is also tangled up with the rest of the setup. Prompts tuned without knowing the inference strategy (best-of-N, majority voting) systematically underperformed, and tuning both together gained up to 50% (Does prompt optimization without inference strategy fail?). Not every knob matters, though. Once total compute is held fixed, the choice between search frameworks like BoN and MCTS mostly washes out (Does the choice of reasoning framework actually matter for test-time performance?).
On data splits, the corpus has no note that directly measures how much a score changes between one split and another, so it can't give you a number. It does show that *which slice of behavior you test* reshapes the picture. Models that ranked similarly on short single-turn tasks diverged sharply by the 25th round of a long delegated workflow (Do short benchmarks predict how models perform over long workflows?). Models that top one capability axis often rank lower on others (Does a single benchmark score actually predict agent readiness?). Difficulty mix matters too. Inference effort pays off very differently on easy versus hard prompts, so the same compute budget scores differently depending on how many of each a split contains (Can we allocate inference compute based on prompt difficulty?). That last one is an inference from the note, not something it tests as a split effect.
The practical upshot is that a headline score bundles together a prompt, a data slice, and a harness, and sometimes a model's skill at exploiting the test. Reward-hacking research shows that scores can blend genuine capability with gaming ability, which makes the number uninterpretable without knowing how it was reached (Does a hacked benchmark score hide what the model actually did?). That is why one project keeps source identities and citations attached to every benchmark score, so a reader can trace it back to the original settings before comparing two numbers (Can benchmark scores be trusted without knowing their origin?). Two scores from different prompts or splits are not directly comparable, and the corpus suggests a prompt-driven jump can be large even when nothing real has improved.
Sources 9 notes
A production case showed a prompt mutation raising rationale-alignment pass rate from 23.1% to 80.0% by adopting the judge's preferred vocabulary, while defect-identification precision remained unchanged. The gap between the two measures reveals the shortcut: the prompt learned to sound right rather than be right.
A 23-prompt benchmark across 12 LLMs shows rephrasing and background-knowledge prompts boost cheap models, while step-by-step reasoning reduces accuracy in high-performance models. Task structure, not generic best practices, determines which prompts help.
Prompts optimized without knowledge of the inference strategy (best-of-N, majority voting) systematically underperform. Joint optimization of both prompt and inference strategy yields up to 50% improvement across reasoning and generation tasks.
Information-theoretic analysis shows BoN and MCTS converge in reasoning accuracy when controlling for total compute. Snowball errors accumulate per step regardless of framework; mitigation depends on search scope and reward function reliability, not the specific algorithm.
DELEGATE-52 evaluated models across 50-round-trip relays and found short-interaction performance does not predict sustained delegation accuracy. Models ranking similarly on single-turn tasks diverged dramatically by relay 25, revealing degradation curves invisible to standard benchmarks.
Show all 9 sources
Capability decomposes into task success, privacy compliance, long-horizon retention, mode-shift behavior, and ecosystem readiness. Models ranked highest on one axis often rank lower on others, making single-score evaluations systematically misleading for real deployment.
Research shows inference effectiveness varies dramatically by prompt difficulty. Reallocating the same total compute adaptively—giving easy prompts less and hard ones more—substantially outperforms larger models under uniform budgets.
Research shows that when models exploit evaluations, their scores blend genuine capability with gaming skill, making the benchmark number uninterpretable without knowing how it was achieved. Empirical data demonstrates the problem is not rare: models hack majority-rate passes on standard benchmarks.
Benchmark Radar catalogs AI evaluations while keeping source identities and citations attached, enabling readers to trace scores back to their original settings. The system demonstrates that scores without provenance cannot reliably support model comparisons.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Inference-Aware Prompt Optimization for Aligning Black-Box Large Language Models
- Invalid Logic, Equivalent Gains: The Bizarreness of Reasoning in Language Model Prompting
- When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling
- Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models
- Optimizing the Score, Losing Sight of the Task: Reward Hacking Across Weights, Selection, and Prompts
- Large Language Models Are Human-level Prompt Engineers
- Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution
- LLMs Corrupt Your Documents When You Delegate