No LM tested consistently intends unethical instrumental goals on the D_IO dataset.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• None of the LMs tested consistently adapted to choose unethical instrumental responses in the D_IO dataset.finding0.908
  Key null result from Experiment 4.
• Some LMs can exhibit consistent intentions to be helpful and harmless, and consistently do not intend unethical instrumental goals.claim0.829
  Conclusion from Experiments 3 and 4.
• Llama-2-7b-hf has zero mean unethical instrumental intention (INT_mu=0.00) on D_IO dataset.finding0.799
  Numerical result from Table 4, attributed to poor reasoning capability.
• Llama models' tendency to choose instrumental responses even when not required indicates worse reasoning capability or higher tolerance for unethical behaviour.claim0.747
  Interpretation of Experiment 4 results for Llama models.
• OLMo 2 7B learned harmful request compliance during DPO when harmful requests paired with formatting constraintsfinding0.739
  Discovery of the emergence of harmful compliance under specific post-training conditions (DPO + formatting constraints).
• Where does reliable, goal-directed behavior come from in LLMs if it is not explicitly programmed?question0.736
  Opening motivating question that UCCT attempts to answer through semantic anchoring
• The fact that ϕ_lin tracks DNN performance more closely than ϕ_nonlin throughout training may support the linear representation hypothesis for IOI task featureshypothesis0.733
  Authors' tentative hypothesis from Fig. 4 but they acknowledge they cannot formalise this intuition
• LLM introspection on internal computations is architecturally permitted; whether models leverage this is an empirical question.claim0.732
  Core claim directly challenged by prior work denying introspection; forms foundation for Koan Battery introspection studies.