LLMs sometimes know statements are false but generate them anyway, motivating the need for techniques that inspect internal model state rather than outputs alone

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

• We hypothesize that LLMs represent correctness of arithmetic expressions differently from factual statements.hypothesis0.818
  Core working hypothesis motivating the factual vs. arithmetic task split in the experimental design.
• LLMs can predict their own responses more accurately than external observers, implying privileged internal knowledgefinding0.802
  Binder et al. finding cited as evidence that LLMs possess introspective capacity analogous to mindfulness
• As LLMs scale, they develop increasingly general abstractions, with large models linearly representing abstract concepts like truth that capture shared properties of diverse inputsclaim0.801
  Interpretive claim connecting scale to abstraction level in LLM representations
• When LLMs produce experience claims under self-reference, is this sophisticated simulation or genuine self-representation, and how would we tell the difference?question0.799
  The core interpretive question the paper narrows but cannot definitively answer
• The paper does not claim that the LLM itself is the source of meaning and value, only that meaning can be discerned in its output.claim0.796
  Clarification to avoid misinterpretation.
• Linear truth directions in LLMs are reliable primarily in factual recall cases and break down when truth assessment depends on computing and storing intermediate results.claim0.793
  Central empirical conclusion of the paper about the fundamental limits of truth directions.
• Code agents operate on structured data with exact arithmetic, while LLMs must parse natural-language observations and track state across turns; some failures may partly reflect numerical parsing or working-memory limitationsclaim0.793
  discussion of potential confounds
• LLMs trained only on language data have rich enough knowledge of visual structures that decent visual representations can be trained on images generated solely by querying the LLMfinding0.789
  Sharma et al. result supporting cross-modal alignment: language-only models implicitly encode visual structure