Stages of inference on HellaSwag dataset show very few deviations from GSM8k results, with slightly higher sink rates across all models

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

• Stages of inference metrics (ColSum concentration, mixing score) remain broadly consistent even for the GSM8k prompt exhibiting the largest orbit amplitude in Huginn-0125finding0.808
  Demonstrates robustness of inference stages to non-fixed-point limiting behavior
• Massive activations are required for stages of inference to emerge in looped modelshypothesis0.737
  Hypothesis supported by ablation of massive activations in Retrofitted Llama that eliminates stage structure
• DeepSeek-R1 Llama 8b gains 0.16% accuracy on GSM8k with positive intervention (more reflections) at cost of ~2000 additional tokensfinding0.731
  Only model showing marginal benefit from increased reflection, at substantial token cost
• 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 formationclaim0.729
  Mechanistic explanation for the negative result observed for Huginn-0125
• We hypothesize that native self-report, fine-tuned introspection models, and trained activation-to-language systems will show different performance on bias-resistant localization and strength benchmarkshypothesis0.727
  Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
• Base models assign higher likelihood to typical-set (representative) sequences than to degenerate sequences under VS promptshypothesis0.724
  Assumption D.6 formalized in the theoretical framework; empirically validated with coin-flip typicality rating experiments
• Triggered Reflection with 'Alternatively' achieves accuracy .684 on gsm8k_adv for Gemma3-4B-ITfinding0.724
  Highest single-instruction accuracy result in the paper.
• Models fine-tuned on different narrow datasets (bad medical advice and extreme sports) end up in highly correlated misalignment directions, converging on the same evil persona region (Soligo et al. 2025)finding0.723
  Evidence for the evil persona as a privileged basin supporting Hypothesis 3