GRUs trained on the Arithmetic task use different types of numeric representations than incremental counting models

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

• There are fewer representations competent for N tasks than M<N tasks, so training more general models should yield fewer possible solutionshypothesis0.785
  Selective pressure toward convergence via task generality
• Language models prefer reusing generic arithmetic mechanisms over learning task-specific modular computations even when task-specific geometry existsclaim0.777
  Broader interpretive claim about LM learning bias inferred from the findings
• Software implementations for all of the models/behaviours presented are common for n = 2, and can be made very efficient for α_i that map many objects onto a much smaller set of object families.claim0.767
  Claim about current practical feasibility and efficiency of 2-way associative implementations.
• MAS successfully aligns the Count variable from Multi-Object GRUs with the Rem Ops variable from Arithmetic GRUs with moderate IIAfinding0.755
  Shows MAS can compare specific numeric variables across tasks with different domains/codomains.
• Today's Large Language Models have become so good at playing Turing's game that it often takes experts to demonstrate the present limits of their ability to simulate human-like intelligence.claim0.752
  Paper's assessment of current LLM capabilities relative to Turing Test
• Under ask-correct, probes trained on arithmetic tasks A1-A3 generalize almost perfectly to factual tasks F0-F2 (AUROC ~1.0), whereas under no-prompt this generalization is largely absent.finding0.748
  Key improvement in cross-task generalization enabled by explicit instruction framing.
• Models more effective at recognizing abstract nouns than other concept typesfinding0.742
  Opus 4.1 demonstrates highest introspective awareness on abstract nouns (justice, peace, betrayal) with nonzero awareness across all concept categories tested.
• We hypothesize that intervention efficiency can be scaled with multi-node and multi-GPU training as language models grow largerhypothesis0.741
  Future work hypothesis about scaling pyvene's computational efficiency for very large models