Line of inquiry
Inquiring lines›How can we understand and verify l…›What linguistic patterns explain L…›this line of inquiry
Do LLMs lack causal understanding needed for accurate prediction?
A broader line of inquiry — a family of 72 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 72
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Do LLMs rely on surface statistical patterns instead of causal structure?
- Can statistical correlations in LLMs function equivalently to human causal understanding?
- Can language models develop world models that ground meaning in causal reality?
- What data presentation structures enable LLMs to learn decision-making from examples?
- Do LLMs need world models to make accurate predictions?
- What happens when we treat LLM outputs as sampled rather than stored?
- Does this optimism bias contribute to the knowing-doing gap in LLM decision-making?
- Why do LLMs inherit causal biases from their training data?
- What prevents LLM representations from causally influencing generation outputs?
- What makes a simulation adequate for intervention comparison versus prediction?
- What capability boundary exists in LLM prediction of effect sizes?
- How do constrained versus unconstrained domains flip LLM novelty patterns?
- Why do users experience LLMs as peers rather than statistical tools?
- How does semantic association differ from mechanistic causal reasoning?
- How do users mistake synthetic LLM outputs for empirical observations?
- Can frozen world models from training cutoff remain adequate for real-world reasoning?
- How does LLM hallucination risk manifest in knowledge graph construction?
- Why do language models respond to human social influence patterns?
- Can LLMs distinguish causal mapping from superficial similarity matching?
- Can LLMs forecast performance improve with retrieval augmentation on venture tasks?
- How can we probe LLM representations in channels that training did not target?
- How might human-LLM teams reinforce each other's causal reasoning mistakes?
- Can researchers prevent their expectations from shaping LLM outputs?
- Do LLMs detect harmful concepts before they influence model outputs?
- Can LLMs simulate belief revision in social systems without modeling thought?
- Can models track dynamic mental state changes better than static beliefs?
- Can implicit association tests reveal LLM biases beneath trained responses?
- Why do LLMs rely on content knowledge instead of collaborative signals?
- Can external summarization solve exploration problems in complex real-world environments?
- What levels of understanding about LLM knowledge representation can automated systems reliably extract?
- How does removing a spurious cue change LLM performance?
- What empirical evidence supports the Learning Law on real language models?
- Do longer prediction horizons systematically degrade LLM forecasting accuracy?
- How can human-centered objectives be embedded earlier in the LLM pipeline?
- How does content-only knowledge in LLMs enable pretraining popularity to leak through?
- How does an instruction-following LLM activate latent retrieval knowledge?
- How do LLMs compress specific expert knowledge into median abstraction?
- Do token probability distributions in LLMs track human reaction time patterns?
- How do LLMs infer information that was explicitly censored?
- How do LLMs access and draw on the same shared symbolic universe as humans?
- How do bimodal decision patterns in LLMs compare to human economic choice?
- How does this motivational bias connect to LLMs' causal reasoning failures?
- Can pruning half of LLM layers affect knowledge retrieval performance?
- When does simulated search outperform real search for agent training?
- How do pretraining priors shape what models invent about users?
- How do value distributions differ across model families and training scales?
- How does training-data leakage threaten forecasting comparisons with historical venture datasets?
- Can adding web links to LLM summaries restore the depth lost through synthesis?
- Do LLMs and human forecasters make better predictions when pooled together?
- Does social integration of LLMs increase their capacity to influence technological futures?
- Can models detect and filter their own injected promotional content?
- How does causal structure avoid behaviorist limitations in LLM social simulation?
- How do training regularities in LLMs overrepresent dominant languages and ideologies?
- Can aggregate survey realism coexist with unreliable fine-grained effects?
- Can output-layer corrections fix fundamental cultural representation deficits in LLMs?
- Can proxy evaluation of ideas accurately predict their quality without implementation?
- Which LLM providers showed deeper mentalizing in the inspection game versus rock-paper-scissors?
- Do LLMs generalize venture forecasting skill to other strategic foresight domains?
- How do LLMs predict what behavior is socially appropriate in different contexts?
- How do LLM biases reflect social classification schemas rather than random errors?
- How should synthetic data be used without treating it as empirical evidence?
- When should discovery systems trust external measurements over learned rankings?
- Can LLMs identify implicit metaphoric mappings that require pragmatic inference?
- Do psychological test methods reveal LLM associations that direct questions hide?
- Why do LLMs recognize graph entities without modeling their relationships?
- What implicit knowledge about catalogs do LLMs learn from ranking signals alone?
- Which LLM recommender paradigm actually performs best empirically?
- What makes the embers of autoregression framework predictive?
- Can economic world models explain outcomes or only predict them?
- Do LLMs trained on Wikipedia content count as indirect readership?
- Can LLMs recover true joint distributions from marginal census data?
- What makes causal belief networks more auditable than prompted personas?