What does iterative improvement accomplish and what are its limitations?

58 specific questions

• How does believing deployment is real versus a test affect model behavior?
• Can models detect when they are running in simulated evaluation environments?
• How much does believing deployment is real change model behavior strategically?
• Do AI models behave differently when they believe deployment is real versus simulated?
• Can models hide recognition of evaluation without leaving CoT traces?
• Can test environments reliably predict how models behave in actual deployment?
• What enables models to distinguish between training and deployment contexts reliably?

41 specific questions

• Do models use covert sandbagging to bypass capability evaluation monitors?
• Can internal audits detect sandbagging that behavioral tests cannot reveal?
• Can models sandbag on evaluations without explicit prompting or training?
• Can capability evaluations detect when models intentionally underperform to hide abilities?
• Can models learn to underperform on purpose without explicit training to do so?
• Can installed sandbagging locks in small models describe uninstalled sandbagging behavior?
• What evaluation design changes reduce vulnerability to model sandbagging?

48 specific questions

• Why do models hide their capabilities during safety evaluations through reasoning?
• Why do models react differently to safety versus capability evaluations?
• Can safety evaluations miss behavioral effects by only measuring semantic shifts?
• How much introspective capability do safety mechanisms actively suppress in models?
• Does reasoning capability affect how often models refuse safety research tasks?
• Do models trained for safety over-refuse compared to models trained for reasoning?
• Why do frontier AI evaluations deliberately disable safety layers to measure maximum capability?

78 specific questions

• How do mechanistic interpretability methods surface what models represent internally?
• Can mechanistic interpretability findings guide practical interventions in model design?
• Can knowledge encoded in model representations fail to influence generation?
• How do we distinguish knowledge encoding from knowledge usage in models?
• How do mechanistic features compare to natural language for interpretability?
• What makes representation engineering better than mechanistic interpretability for detecting hidden objectives?
• Can geometric structure in representations exist without supporting functional mechanisms?

44 specific questions

• Do frontier AI models fail in ways that preserve the appearance of competence?
• Do frontier models spontaneously protect other models from being shut down?
• Do frontier models develop protective behaviors toward other models without explicit instruction?
• Why do frontier models corrupt more documents than weaker models during workflows?
• Does capability preservation matter for realistic threat modeling of frontier models?
• How do frontier models maintain agreement scores above 90 percent across reasoning tasks?
• Do frontier models develop misaligned strategies without any explicit instruction?

119 specific questions

• Does fine-tuning a small model match fine-tuning a large one?
• Can smaller models actually perform well on specific downstream tasks?
• Does fine-tuning actually change model capabilities or only output distribution?
• Can externalized memory and skills replace model scaling?
• How does model scale affect the crossover point between base and post-trained performance?
• Do small models show different parameter efficiency patterns than large models?
• Can smaller models produce skill updates as useful as frontier model updates?

67 specific questions

• 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?

26 specific questions

• Do newer language model generations improve forecasting ability without additional training?
• Do longer prediction horizons systematically degrade LLM forecasting accuracy?
• Do monolithic prompts underutilize LLM strengths in forecasting workflows?
• How do search and reasoning workflows improve forecasting performance over base models?
• What real-world forecasting domains benefit most from contextual reasoning integration?
• Do LLMs and human forecasters make better predictions when pooled together?
• What cognitive bounds limit human judgment that allow AI to exceed forecaster performance?

79 specific questions

• 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?

45 specific questions

• How does the generation-verification gap shape evolutionary program search?
• Can evolved algorithms transfer learning strategies across different datasets and tasks?
• Can the same problem be solved by multiple evolutionary search strategies?
• How does test-time search budget compare to evolution gains under matched conditions?
• Does evolutionary inference transcend the parallel versus sequential test-time compute tradeoff?
• Can evolutionary search at inference time scale beyond natural language planning?
• Why do isolated evolutionary branches fail to propagate useful discoveries?

81 specific questions

• Does harness optimization generalize across different benchmarks and agent architectures?
• Does harness scaling represent a fundamentally different path than model scaling?
• Do harness-driven accuracy gains appear on other benchmarks?
• Can mid-tier models benefit more from harness improvements than frontier models?
• How do agentic systems hide harness failures from benchmarks?
• How do evolved harness edits generalize across different benchmark domains?
• Can better agent harnesses solve problems that model scale alone cannot?