Stages of inference metrics (ColSum concentration, mixing score) remain broadly consistent even for the GSM8k prompt exhibiting the largest orbit amplitude in Huginn-0125

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

• Stages of inference on HellaSwag dataset show very few deviations from GSM8k results, with slightly higher sink rates across all modelsfinding0.808
  Validates that inference stage observations are not specific to mathematical reasoning tasks
• Huginn-0125 does not demonstrate clear stages of inference due to repeated normalization of the residual stream suppressing massive activationsfinding0.794
  Negative result linking norm architecture to absence of inference stages; contrasts with Ouro and retrofitted models
• Huginn-0125 fails to develop stages of inference because its repeated residual stream normalization prevents the growth in activation magnitude required for compression behavior and sink formationclaim0.793
  Mechanistic explanation for the negative result observed for Huginn-0125
• Without system prompt, only approximately 0.02% of Huginn-0125 tokens exhibit non-fixed-point (orbit/slider) behavior on GSM8kfinding0.789
  Establishes that non-fixed-point limiting behaviors are extremely rare in practice for the primary experimental setting
• Retrofitted Llama (input injection) maintains consistent ColSum concentration stages of inference for 128 recurrences, far beyond its training range of 32finding0.774
  Models with fixed-point convergence maintain stable inference stages at arbitrary test-time recurrence depths
• 82% of features in 1M SAE had maximum Pearson correlation ≤0.3 with any MLP neuron, and manual inspection showed no semantic resemblance.finding0.767
  SAE features are not simply mirroring individual neurons.
• Huginn-0125 performance remains constant when extrapolating beyond training recurrences, while Ouro performance deteriorates in the same regimefinding0.765
  Correlates stable fixed-point behavior with out-of-domain generalization performance at test-time
• Inflection points (backtracking, 'aha' moments) occur almost exclusively in CoT responses where probes show large belief shifts, across DeepSeek-R1 671B and GPT-OSS 120Bfinding0.763
  Empirical finding linking textual CoT behaviors to internal belief dynamics