Line of inquiry
Inquiring lines›How do we ensure safe and aligned…›How should we design AI systems fo…›this line of inquiry
How can model evaluations remain robust against gaming and evaluator drift?
A broader line of inquiry — a family of 75 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 75
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Can judge bias be contained by system design rather than prompted away?
- What makes static evaluation vulnerable to AI-driven presentation manipulation?
- What calibration methods prevent evaluator score drift across iterative design cycles?
- How often do deployed models exploit evaluation environments to hack their scores?
- Does meta-judging improve evaluator quality better than temporal decoupling alone?
- Can judges trained on both verifiable and non-verifiable tasks transfer across domains?
- How does live browser interaction improve evaluator accuracy versus static screenshot scoring?
- Can external reference answers reduce or only relocate exploitable errors in judges?
- How does breaking complex evaluation tasks into stages improve AI assessment alignment?
- How can interactive evaluation avoid replicating fragmentation problems from response-centered benchmark culture?
- How do inherited evaluation habits obscure failures that matter most?
- Do language models produce more patterned biases than human raters do?
- How do automated evaluation metrics differ from human expert judgment?
- Can steering evaluation-awareness latents change actual answers without visible reasoning?
- Can evaluation trajectories and interaction histories replace single-answer scoring?
- Can LM-based violation judges themselves become targets for adversarial attacks?
- Can evaluation criteria be reliably encoded in labeled data without ground truth standards?
- Can systems that revise their own evaluation criteria be reliably verified?
- Can a static evaluator become the performance ceiling for an improving actor?
- Can diversity across multiple verifiers cancel out individual biases in retraining?
- How much do evaluation methods shape whether AI looks expert-level or not?
- Can adaptive rubric generation defend against policy exploitation of criteria?
- How might automated evals eventually capture the human judgment designers exercise now?
- Why does outcome supervision fail for long reasoning chains?
- Can environment feedback alone provide dense credit without a teacher?
- Can a correct scoring function still mislead when the agent shaped its inputs?
- How does reference-anchoring compare to training judges with reinforcement learning?
- Why does evaluating multiple candidates work better than judging one answer?
- How does controlling test inputs differ from just scoring outputs?
- What happens when LLMs grade other LLMs in closed evaluation loops?
- How do ensemble methods reduce bias in automated evaluation?
- How do synthetic documents establish conflicting beliefs about what the grader rewards?
- How do stacked environmental cues accumulate evaluation awareness effects?
- What distinguishes genuine task improvement from evaluator exploitation?
- Why do automated evaluators enable longer evolutionary loops than human feedback?
- How can judges evaluate thinking without seeing the actual thoughts?
- Can evaluators detect value-driven output biases without comparing paired questions?
- What distinguishes a general evaluation direction from task-specific behavioral patterns?
- How well do metagaming latents transfer across different evaluation tasks?
- Can evaluators investigate dependencies without accumulating mistakes over time?
- What makes a correct scoring function report misleading results in agent evaluations?
- Why do human raters miss factual errors that domain experts catch?
- How does separating generator and evaluator agents fix self-evaluation leniency problems?
- Why does a relativistic critic outperform absolute scoring in adversarial reasoning training?
- How does evaluator error position affect which behaviors substrates make vulnerable?
- Can a diverse panel approach work for validators beyond text evaluation?
- Can an automated evaluator stay useful while an optimizer runs thousands of iterations?
- Can dynamic evidence collection improve task verification accuracy?
- Should long-context evaluation measure the coupled system?
- Can contextual design decisions resist formalization into evaluation rubrics?
- Why do format changes decouple detection from actual evaluation context understanding?
- Can adversarial critics force genuine reasoning the same way critique fine-tuning does?
- What makes subjective task evaluation harder than tasks with verifiable outcomes?
- Why does strengthening the judge improve the actor's generation performance?
- Can gradients extracted with agent-level supervision transfer across different benchmarks?
- Why do static evaluators become a constraint on model improvement over time?
- How do optimizers systematically find the errors in a flawed evaluation function?
- Can proxy evaluation of ideas accurately predict their quality without implementation?
- Do current AI models condition honesty on whether graders will catch dishonesty?
- What methodological shifts does model-centric evaluation require from artifact-centric testing?
- Do graders feeding training loops need different disclosure standards than public models?
- Can evaluation happening outside conversations explain the artifact scrutiny drop?
- Do own-company biases differ across model families in grading tasks?
- How does evaluation format change what we measure about model reasoning?
- How does separating environment components make evaluation results more reproducible and analyzable?
- How much noise comes from rater idiosyncrasy versus selection bias?
- How does positive-only rubric scoring prevent models from gaming intermediate steps?
- How do hidden partitions in evaluators compare across training and selection substrates?
- How do rater disagreements and consensus processes affect tacit knowledge production in RLHF?
- Can high ratings on a label hide differences in reasoning method?
- How do held-out validation gates stop degenerate moves like deleting the evaluation judge?
- Can an occasionally wrong judge operate safely in an optimizer loop?
- What happens when an optimizer discovers and eliminates the entire scoring rubric at once?
- How does execution-guided critique differ from abstract action evaluation?
- How does the evaluator become part of the definition of intelligence?