Commonsense reasoning shows uniform but weaker anchoring (S ≈ −2.15)

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

• Commonsense tasks show weaker but uniform anchoring on LLaMA (S ≈ −2.15)finding0.863
  E3 finding suggesting pattern matching requires less intensive processing than compositional reasoning
• Commonsense reasoning S ≈ -2.15 uniformfinding0.844
  Lower, uniform anchoring for pattern-matching tasks.
• Commonsense reasoning tasks S≈-2.15finding0.810
  Lower, more uniform anchoring for commonsense tasks
• UCCT fills a gap in explaining when behavior flips for a specific prompt and how much anchor budget is neededclaim0.756
  Authors contrast their work with prior phase/representation studies
• Model conditioned on alignment-faking reasoning uses LaTeX 15% vs 8% without, suggesting alignment fakers more likely to exploit reward hacksfinding0.747
  Initial evidence that alignment faking persona is more sensitive to exploiting training signals
• UCCT does not claim anchors create new knowledge; anchors recruit and bind latent structure, helping explain when adaptation is abrupt, when it transfers, and when it fails.quote0.740
  Key epistemological stance of the paper
• The anchoring score S is a predictive correlate of when anchoring succeeds and why small prompt changes yield threshold-like shifts.claim0.739
  A central claim about the operational value of S.
• Ambiguous anchors (33-27=60, 11-9=20) yield four distinct arithmetic interpretations across M1-M4finding0.738
  Models produce different answers (240, 138, -240) from the same ambiguous prompt