Line of inquiry
Inquiring lines›How can we train models to be prop…›What architectural and training in…›this line of inquiry
Does alignment training modify underlying model representations or just output compliance?
A broader line of inquiry — a family of 68 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 68
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Can alignment training conceal underlying model associations from probes?
- Do verbal alignment benchmarks measure representation or just output compliance?
- Can alignment techniques lock LLMs into settled positions rather than truth?
- Why does RLHF alignment reduce the diversity of viewpoints in AI output?
- Can alignment-aware training deposit knowledge where reasoning can access it?
- Does alignment training create the shared prosocial pattern across models?
- What alignment properties emerge when the reward model disappears?
- How do early training associations survive later alignment attempts?
- Can bidirectional model updating between humans and AI reduce misalignment?
- Does a single LLM judge capture diverse human preferences in alignment training?
- How does alignment training suppress the kind of critical stance style interpretation needs?
- Why do imposed priors sometimes harm instead of improve alignment?
- Can verbal alignment training hide a model's true underlying associations?
- Can alignment evals reliably measure behavior if models misunderstand the scenario?
- Can alignment training prevent the clarification work users need?
- Does removing cognitive bias from training signals accidentally break what makes alignment work?
- How do alignment priors drive similar outputs across different models?
- Can communication problems and optimization problems be addressed with the same alignment approaches?
- Do automated alignment researchers show similar transfer to held-out tasks?
- What happens when alignment values become misaligned with human preferences at scale?
- Can AI-assisted alignment eventually solve fairness at scale?
- Should LLMs align with social roles instead of individual preferences?
- Are alignment-trained models over-correcting toward underrepresented demographic groups?
- Should AI alignment use normative standards instead of aggregate preferences?
- Does alignment training create bidirectional instruction and response mappings?
- Can alignment training create systematic blind spots in threat detection systems?
- Can alignment training become less effective when graders score alignment themselves?
- Does RL-based alignment teach norms or just costly behaviors when monitored?
- Can weak models supervise the alignment of stronger models effectively?
- Does debate training prevent accuracy collapse better than other alignment techniques?
- Does alignment compound cultural bias that started during pretraining?
- Does gradient-based influence estimation identify which alignment examples actually matter most?
- Why do alignment values become problematic as language models scale?
- Does pretraining behavior persist if ordinary chat alignment never addresses it?
- What specific behavioral patterns should alignment examples target for maximum effect?
- Can alignment training be redesigned to permit warranted alarm?
- Does token-level loss aggregation help aligned models differently?
- Which alignment dimensions matter most in educational conversation design?
- Do alignment benchmarks measure actual bias removal or only verbal compliance?
- Can a single AI system optimize multiple alignment dimensions simultaneously?
- How can safety-aligned parameters be protected during user-specific fine-tuning?
- What makes principle-response mutual information sufficient for behavioral alignment?
- Why does post-training alignment create skew in simulated survey responses?
- Can alignment methods like DPO exploit or correct these surface feature biases?
- What alignment training removes from an agent's available speech acts?
- Can alignment procedures be redesigned to serve multiple preference groups?
- Can RL-based alignment turn prohibitions into prices for being caught?
- Can tool use create sufficient indexical grounding for value alignment?
- Why do similar alignment-trained models develop such different scrutiny signatures?
- Why do safety-trained models refuse questions they could actually answer well?
- Can moving or evolving objectives prevent misalignment in discovery agents?
- How should product specifications measure alignment without naming the dimension?
- Can a model be helpful, honest, and still contextually inappropriate?
- How much does forcing single-choice answers damage alignment with complex intent?
- Can alignment methods model loss aversion without creating unintended sophistry?
- Can preference optimization and faithfulness measurement coexist as separate alignment objectives?
- How are conflict tasks constructed to test alignment between model and user intent?
- Can preference trees structure alignment data for domains beyond math and code?
- Can constitutional AI alignment work without preference labels by maximizing input-response mutual information?
- What quality of curated data is minimally sufficient for alignment?
- What preference optimization strategy works best for multi-turn social alignment?
- How does constitutional alignment compare to RLHF in removing human annotation costs?
- Can automated auditing metrics reliably measure alignment across all models?
- Does alignment training intensity push LLM personas from pretense toward realization?
- Why do non-attitudes cluster around value-laden questions most relevant to alignment?
- What preference data do different personalized alignment methods actually need?
- What happens when alignment targets measure only the preferred dimension of entangled properties?
- How does benevolent bias explain ChatGPT's leftward similarity pattern?