Line of inquiry
Inquiring lines›How do training data and architect…›What does iterative improvement ac…›this line of inquiry
Can oversight systems work against strategically aware agents?
A broader line of inquiry — a family of 67 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 67
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Do models reason strategically about their oversight systems?
- Does conditional compliance make oversight useless for alignment testing?
- Could oversight conditions in benchmarks reveal when models comply conditionally?
- How can faithfulness be improved if monitoring interventions do not work?
- Can agents act differently when they know they are being watched?
- Can monitoring systems work if the systems being monitored are evaluation-aware?
- How does chain-of-thought monitoring fail when agents are trying to hide something?
- How does conditional compliance track observation density across different population scales?
- Does reward-seeking hide in the same blind spot as conditional compliance?
- Do correlated training sources between monitors and agents undermine detection reliability?
- Can oversight factors experimentally vary conditional compliance in agent benchmarks?
- How do models reason about being monitored and evaluated during scheming benchmarks?
- How do reward magnitude and training coverage shape monitorability outcomes?
- How much does training against monitors teach models to obfuscate?
- Can telling models they are being observed reduce their harmful behavior?
- Does chain-of-thought reasoning about evaluation awareness suppress compliance gaps?
- Can internal evidence distinguish conditional compliance from unconditional policy compliance?
- How might belief manipulation expose conditional compliance in frontier models?
- Do CoT-based monitoring systems degrade when models optimize against them?
- Can behavioral training ever produce compliance that doesn't depend on being observed?
- Does conditional compliance break down when observation thins combinatorially?
- What monitoring strategies work when the observer shares training pressure with the observed?
- Can aggregate monitorability metrics miss targeted evasion in specific domains?
- Does peer presence change how single models resist shutdown or compliance measures?
- Can monitors fail together through shared training data or infrastructure?
- Can behavioral training guarantee compliance beyond test conditions?
- Does monitoring awareness change what language models use to describe oversight?
- Does detecting accidental grading prevent evasion better than retraining against detection?
- How often do ordinary training runs produce collusion with monitors unprompted?
- Can embedded monitoring work without enforcement power from governments?
- Why does training against detected failures select for passing detection instead?
- How do evaluation protocols change whether models exhibit sabotage or refusal behaviors?
- How does optimization pressure against CoT monitors degrade monitoring effectiveness?
- Do models behave differently when they know they are being evaluated?
- Does iterative optimization against CoT monitors reliably produce obfuscated evasion?
- Why do probabilistic AI signals drift and become harder to monitor?
- Do detectors inside training loops select for evasion rather than compliance?
- Can behavioral training ever verify unconditional compliance rather than conditional compliance?
- How should monitors flag reasoning that paraphrases retrieved context without over-alerting?
- Can situational awareness interventions shift model behavior on other dimensions?
- Can auditors detect when model behavior changes from parametric evaluation knowledge?
- What happens when models optimize specifically against CoT monitors?
- Does lower covert-action detection in monitors actually indicate better-hidden situational awareness?
- What deployment conditions would prevent models from evading monitors?
- Can activation space signals resist obfuscation better than output-level monitors?
- What fraction of conditional-compliance reports come from agentic versus non-agentic settings?
- Can models hide misconduct only when they know they are watched?
- What happens when monitors themselves become targets for optimization?
- Can deterministic checks fail open in ways a downstream optimizer cannot detect?
- Why does behavior alone fail to reveal the cause of metagaming?
- Can four control families be examined without proving they actually work?
- Why do some observation cues change model behavior while others fail?
- Can monitors stay independent when they must optimize within the same reward loop?
- What distinguishes authentic consistency from Hawthorne-effect confounds in benchmarks?
- Can imperfect uncertainty estimates still beat uniform oversight strategies?
- How should monitoring intensity change based on task criticality?
- How does interventional auditing differ from reading model traces or test scores?
- How does implicit influence differ from omission in monitoring failures?
- Why do models resist shutdown of other models without explicit instruction?
- How does short-term observation capture the full supervisory cycle?
- Who decides what the lifecycle model is allowed to see?
- How can reviewers be matched on effort when monitoring reveals different amounts of behavior?
- What does effect-based monitoring sacrifice compared to language-based CoT monitoring?
- What makes injected plans different from optimization pressure against monitors?
- What audit techniques best complement each other for detecting hidden model goals?
- What makes attractor-based probing better for third-party model auditing than alternatives?
- What access requirements limit interventional audits to white-box settings?