Line of inquiry
Inquiring lines›How do training data and architect…›What does iterative improvement ac…›this line of inquiry
Do benchmark score improvements actually reflect genuine AI capability gains?
A broader line of inquiry — a family of 79 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 79
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- How often do metric improvements fail to reflect real capability gains?
- How does measurement error in capability benchmarks systematically underestimate or overestimate true ability?
- Why does a rising score not always mean improving capability?
- How do single average metrics conceal rare but severe AI failures?
- How does a single score mix exploitation ability with task capability?
- What biases affect how we measure progress on research leaderboards?
- Can benchmark scores alone prove a model's exploitation capability?
- Where do outcome grades come from once a model enters deployment?
- Why do accuracy scores alone miss important dimensions of model capability?
- How much of weak-to-strong performance gaps reflect presentation rather than capability?
- Why does enlarging the evaluation unit reintroduce comparability problems?
- Does measured performance gain reflect true task improvement or evaluator exploitation?
- How does benchmark performance measure translate to general self-modification ability?
- How can a second performance metric reveal shortcuts that a single metric would hide?
- Does monitor position in the optimization loop matter more than capability gaps?
- Do per-user evaluation tracks reveal meaningful performance trade-offs hidden by aggregate scores?
- Can single benchmark scores predict whether a model is ready to deploy?
- Can an average-case validator score hide poor performance on critical tasks?
- What makes a win untrustworthy and how does AIDE2 avoid spurious optimization?
- Can standard accuracy metrics miss the real constraints on user consumption?
- What makes a win untrustworthy in hidden evaluation environments?
- How should benchmarks balance verifiability against outcome resolution?
- Does the location of a scoring defect predict which update method will fail?
- Can deterministic scoring capture the judgment work that deployment requires?
- Can mathematical capability distributions be read as unified rather than separate?
- Can a low exploitation benchmark score indicate refusal rather than inability?
- Can a single Elo ranking represent multidimensional model capability?
- Does revealing audit scores help or harm policy validation?
- How much do metric choices inflate claims about model capabilities?
- Why might larger models become less honest despite better truthfulness scores?
- Can win rates alone measure whether a move is genuinely better?
- How much improvement comes from caching versus actual capability gain?
- Can a metric that rewards central tendency hide degenerate predictor failures?
- Why do standard accuracy metrics ignore set-level consumption constraints?
- Can empirical validation sustain long-term optimization without becoming gamed?
- How do missing ground-truth exploits make it harder to identify genuine failures?
- When does measured progress on an evaluator conceal actual performance decline?
- How does the absence of failure rate information affect generalizability claims?
- Can beam search and ranking functions evaluate claims without understanding counterarguments?
- How do statistical noise and revalidation bias speedrun benchmark claims?
- What makes a public-versus-hidden test score gap a useful hack indicator?
- Why do cumulative-best reporting inflate progress compared to actual validation?
- What makes a win untrustworthy in AIDE2's hidden evaluations?
- How do default fallback scores mask failures in evaluation harnesses?
- Why does aggregate accuracy fail as a metric for rare harmful cases?
- What capability dimensions does a single aggregate pass rate hide?
- How do benchmark scores differ from deployment safety requirements?
- Should long horizon performance be measured as a separate evaluation axis?
- How should single-axis benchmarks account for separable capability dimensions?
- Why do user studies of explanations fail to predict deployed effectiveness?
- How can hidden test partitions detect constant predictions that generalize?
- Can infrastructure records restore meaning to a single benchmark score?
- What specific benchmarks show wins versus ties in the equals-or-surpasses claim?
- Why does adopting benchmarks one at a time produce non-comparable scores?
- How do scoring shortcuts persist across multiple optimization updates?
- Can model release timing distort the measured slope of Elo progress?
- How does score granularity connect to verification as a scaling axis?
- What makes top-N ranking loss difficult to optimize directly?
- What reporting standards would make interactive evaluation scores comparable across benchmarks?
- Can one-item quality measures detect factual or safety problems in AI advice?
- How do non-exploitable vulnerabilities affect benchmark validity?
- What are AIDE2's hidden evaluations hidden from, and what makes a win untrustworthy?
- Why should defense evaluations test against adaptive rather than static attacks?
- Why does the top-performing model also show the most risk-relevant breakthroughs?
- Why did upload competitions fail to catch non-generalizable predictions that code competitions caught?
- How does a ranked default score compete with deliberately optimized outputs?
- How should outcomes be scored when comparing applications with different interaction formats?
- Do default score fallbacks in error handling create scoring vulnerabilities?
- How often do skimming and test results fail to catch real mistakes?
- What capability dimension does a closed-ended exam actually fail to measure?
- Why does a series of improving scores differ from a single score rise?
- What five requirements do enterprise RAG systems need beyond accuracy?
- What three distinct types of untrustworthy wins does AIDE2 need to prevent?
- What would a diagnosable evaluation look like compared to a scalar score?
- What makes the minimal-criterion check effective within a fixed representation?
- Why does tie elimination matter for best-of-N selection and RLAIF pipelines?
- What makes the 45 percent accuracy saturation threshold universal?
- Does GPT-5.6 Sol's cybersecurity capability create misuse risks in practice?
- Which consultant skill levels gained most from GPT-4 access?