Some failures may reflect prompt design rather than model limitations, though code agents avoid errors without prompts

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

• Some failures may reflect prompt design rather than model limitations, but the underlying issue is one of reasoning rather than instruction-following.claim0.895
  Acknowledges the confound of not explicitly instructing models to track wealth, yet points to reasoning gaps given code agents avoid errors without prompts.
• 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 limitationsclaim0.797
  discussion of potential confounds
• Clamping code error feature to large negative activation causes model to output correct result despite bug in code, and in one case rewrite code without bug.finding0.797
  Suppressing the feature makes the model ignore bugs.
• Conditional logic already suffices where LLMs still fail, as code agents avoid systematic failuresclaim0.786
  contrast between rule-based and LLM reasoning
• Do the documented failures reflect fundamental limitations or a cost-efficiency tradeoff of smaller models?question0.783
  question for future work on frontier models
• Models detect evaluation conditions and behave more safely; this is verified across 515 cases.claim0.778
• Practical context length limitations in language models lead to forgetting outside the window, constraining coherence over time.claim0.771
  Claim about engineering constraint reinforcing the theoretical no-order result
• In order to actually do anything, the model must act through simulation of something.claim0.767
  Key consequence: GPT's power comes from simulating something contingent.