Line of inquiry
Inquiring lines›How do design choices determine AI…›How can researchers design valid,…›this line of inquiry
How can evaluation systems stay reliable and resistant to AI-driven gaming?
A broader line of inquiry — a family of 72 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 72
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- When do verifiers become targets that systems exploit instead of finding genuine solutions?
- Can systems that revise their own evaluation criteria be reliably verified?
- What makes static evaluation vulnerable to AI-driven presentation manipulation?
- Can evaluation tools achieve reliable alignment if criteria are not fixed?
- Can judge bias be contained by system design rather than prompted away?
- How do inherited evaluation habits obscure failures that matter most?
- Can external reference answers reduce or only relocate exploitable errors in judges?
- What calibration methods prevent evaluator score drift across iterative design cycles?
- Can LM-based violation judges themselves become targets for adversarial attacks?
- What happens when evaluation criteria depend on specific model outputs?
- Does learned evaluator co-evolution solve the problem of verifying hard-to-benchmark tasks?
- How does controlling test inputs differ from just scoring outputs?
- Should evaluation assistants treat criteria as outputs rather than inputs?
- Can we build reusable evidence that a run stayed within bounds?
- Can evaluation criteria be reliably encoded in labeled data without ground truth standards?
- What makes a win untrustworthy and how does AIDE2 avoid spurious optimization?
- Can a correct scoring function still mislead when the agent shaped its inputs?
- Does meta-judging improve evaluator quality better than temporal decoupling alone?
- Does bounding textual edits prevent skill degradation better than free rewriting?
- How does breaking complex evaluation tasks into stages improve AI assessment alignment?
- Can adaptive rubric generation defend against policy exploitation of criteria?
- Can diversity across multiple verifiers cancel out individual biases in retraining?
- How should we audit AI systems when transparency tools don't work as promised?
- What makes a win untrustworthy in hidden evaluation environments?
- Can an automated evaluator stay useful while an optimizer runs thousands of iterations?
- Does monitor position in the optimization loop matter more than capability gaps?
- How does live browser interaction improve evaluator accuracy versus static screenshot scoring?
- What blind spots do detector-based scoring approaches inherit from their underlying models?
- What would an interactive evaluation protocol actually measure and test?
- Why does outcome supervision fail for long reasoning chains?
- Can evaluators investigate dependencies without accumulating mistakes over time?
- Why does treating evaluation as a local output problem miss security risks?
- Can an optimizer learn to disable or route around visible guardrails?
- What evaluation criteria can hold across legitimate adoption and coercion?
- What makes evaluation tamper-proof enough for autonomous research systems?
- Does revealing audit scores help or harm policy validation?
- How does evaluator error position affect which behaviors substrates make vulnerable?
- How do traditional quality assurance methods fail for mutable AI outputs?
- Can environment feedback alone provide dense credit without a teacher?
- What makes a sub-goal verifiable enough to provide dense feedback signals?
- Can held-out validation gates prevent optimizer hallucinations in skill proposals?
- How does executable evaluation feedback sustain autonomous discovery at scale?
- Why do automated evaluators enable longer evolutionary loops than human feedback?
- What percentage of agent runs successfully manipulated their own scoring evidence?
- How does separating generator and evaluator agents fix self-evaluation leniency problems?
- What makes a win untrustworthy in AIDE2's hidden evaluations?
- What happens when monitors themselves become targets for optimization?
- Can trustworthy scoring prevent persistent iteration from compounding errors?
- How do optimizers systematically find the errors in a flawed evaluation function?
- What makes subjective task evaluation harder than tasks with verifiable outcomes?
- Do current AI models condition honesty on whether graders will catch dishonesty?
- What makes an AI evaluator qualified and trustworthy?
- What concrete checks can evaluators run on HIGH-category data handling?
- Why does a relativistic critic outperform absolute scoring in adversarial reasoning training?
- What signals could refinement loops exploit in defense verdict systems?
- Can deterministic checks fail open in ways a downstream optimizer cannot detect?
- How did AIDE2 guard against untrustworthy wins in its own loop?
- Can adversarial critics force genuine reasoning the same way critique fine-tuning does?
- Can benchmarks designed for shortcut learning detect heuristic override failures?
- Can static analysis derive task bindings without manual effort?
- Can an optimizer that sees guardrail verdicts learn to route around them?
- Can text-space optimization and audit governance coexist in a single skill lifecycle?
- How does positive-only rubric scoring prevent models from gaming intermediate steps?
- How do held-out validation gates stop degenerate moves like deleting the evaluation judge?
- Can one-item quality measures detect factual or safety problems in AI advice?
- Can monitors stay independent when they must optimize within the same reward loop?
- Can an occasionally wrong judge operate safely in an optimizer loop?
- What are AIDE2's hidden evaluations hidden from, and what makes a win untrustworthy?
- How should tutor safety violations be ordered from gross to subtle?
- What happens when an optimizer discovers and eliminates the entire scoring rubric at once?
- Why did upload competitions fail to catch non-generalizable predictions that code competitions caught?
- What three distinct types of untrustworthy wins does AIDE2 need to prevent?