Forced-choice pairwise comparison with Elo conversion operationalizes Alexander's method and produces stable rankings across models.

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

• The results generalise readily to non-equilibrium systems where scaling relationships remain similar (e.g., dynamic or localised scaling).claim0.754
  Claim about broader applicability of the scaling argument
• Overbid frequency, self-bidding rate, bankrupt-initiation patterns, and context-dependent offer calibration are failure modes invisible to both static evaluations and aggregate rankings like Eloclaim0.754
  key claim about the benchmark's unique diagnostic value
• Humans prefer representative distributions over non-representative ones when evaluating outputs at the distribution level, providing a reward gap that VS exploitshypothesis0.750
  Assumption D.3 formalized in the theoretical framework; empirically validated with coin-flip sequence experiments
• The TrueSkill ranking broadly aligns with Chatbot Arena but diverges from reasoning-mode-aggregating evaluations.claim0.748
  Comparison to external leaderboards showing misalignment.
• Typicality bias in preference data is a fundamental and pervasive data-level cause of mode collapse in aligned LLMsclaim0.748
  The central thesis of the paper, distinguishing it from algorithmic explanations of mode collapse
• All VS variants maintain refusal rates above 97% on StrongReject benchmark, within 0.3-0.8 percentage points of the Direct baselinefinding0.745
  Confirms VS does not compromise safety alignment while improving diversity
• Choice among alternatives as a group design strategy does not work realistically because alternatives differ on too many dimensions simultaneously for consensus to formclaim0.742
  Alexander's critique of conventional democratic design processes based on multiple-choice selection.
• The conflict between the model's existing preferences and the stated training objective is the key driver of alignment faking in this setupclaim0.741
  Authors' interpretation of prompt variation results showing alignment faking disappears only when conflicting objective is removed