Line of inquiry
Inquiring lines›How do language models construct a…›How are AI-generated and human-wri…›this line of inquiry
How can identical external performance mask different internal representations?
A broader line of inquiry — a family of 39 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 39
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Can identical model performance mask fundamentally broken internal representations?
- How do weight perturbations reveal what performance benchmarks cannot measure?
- How do surface statistical regularities enable correct outputs while degrading robustness?
- Can RL format selection explain performance gains attributed to algorithmic improvements?
- Why do internal representations differ when external performance matches?
- Does highlighting input features reduce human over-reliance on machine outputs?
- Why do only two of fourteen models improve when problem constraints are removed?
- What distinguishes minimal-pair asymmetry from standard accuracy evaluation?
- Why do scaling laws show capability saturation at specific thresholds?
- What happens when prompt-optimized results lack anchoring in real data?
- Why do models fail under distribution shift if accuracy metrics stay high?
- Why do benchmarks become saturated so quickly after initial launch?
- Why do majority-label benchmarks hide models' failure on subjective tasks?
- Why might larger models become less honest despite better truthfulness scores?
- Can empirical validation sustain long-term optimization without becoming gamed?
- How do unstated constraints become invisible to training data distributions?
- Are larger models and search access substitutes for factual accuracy?
- Why do single function-calling benchmarks mask model weakness in specific areas?
- What is the accuracy cost of enforcing temporal causality inside model parameters?
- What makes some frictions negligible while others block entire pathways?
- What makes diffusion sampling preserve multiple optimal solutions better than alternatives?
- Why do intermediate predictors in looped models align with final outputs?
- Why do rare cases in medicine and science require models that preserve tail distributions?
- Why do feature-based approaches struggle when privacy or latent factors are involved?
- What makes top-N ranking loss difficult to optimize directly?
- Why should deep learning theory prioritize average-case over worst-case analysis?
- How do spectral-norm constraints prevent divergence in world model rollouts?
- What trade-offs emerge between graph staleness and recommendation freshness?
- Can other posterior approximation schemes match variational inference performance?
- Do generic kernel-decay assumptions alone explain coarse-to-fine spectral ordering?
- Why does input embedding magnitude affect perturbation sensitivity in transformers?
- What benefits do open foundation models create that closed systems cannot?
- What makes attractor-based probing better for third-party model auditing than alternatives?
- Why do standard social regularization methods miss the actual value networks provide?
- Why do benchmark designers treat content effects as confounds?
- Why does pure numeric ID indexing force models to learn from scratch?
- What audit techniques best complement each other for detecting hidden model goals?
- How do coverage and identifiability set separate performance ceilings?
- How do power-law distributions differ from uniform collision assumptions?