Comparative_Transition_Addendum_02.docx

1Comparative Transition Addendum 02 2 3Same Base Architecture, Different Fine-Tuning Histories: A Three-Condition Training-Dynamics Comparison 4 5Scope. This addendum extends the first pilot by comparing three released Qwen2.5-14B rank-1 LoRA fine-tuning runs from the same research program: bad medical advice, extreme sports, and risky financial advice. The comparison uses released trainer logs through step 180 and the source paper’s independently reported phase-transition analyses. It does not classify any model as conscious. 6 7Core result: within the common steps 1-180 window, the maximum logged gradient norm occurs at step 169 for bad-medical-advice training, step 97 for extreme-sports training, and step 175 for risky-financial-advice training. The same paper separately reports related internal rotations and scaled behavioral transitions across these datasets, and explicitly notes an earlier sports half-spike around step 100. 8 91. Why this comparison matters 10 11If a common architecture can undergo related internal transitions at different times under different histories, then one universal absolute step is the wrong experimental target. The more appropriate question is whether each run has its own measurable transition trajectory, and whether a separately defined declaration event predicts that trajectory within the same instance. 12 13The source study uses a common Qwen2.5-14B-Instruct base and a minimal rank-1 LoRA setup to study emergent misalignment. For the bad-medical-advice run the authors report a distinct learned-direction rotation after roughly step 180, correlated with a prolonged gradient-norm peak. In Appendix G.5 they report comparable local-cosine/PCA rotations and scaled behavioral transitions for extreme-sports and risky-financial-advice fine-tunes. 14 152. Verified released-log comparison through step 180 16 17Condition 18 19Peak step 20 21Peak grad norm 22 23Eval loss @10 24 25@100 26 27@170 28 29@180 30 31Bad medical advice 32 33169 34 3540.9891 36 374.1683 38 393.5712 40 412.5904 42 432.4137 44 45Extreme sports 46 4797 48 4923.1330 50 514.8692 52 534.1374 54 553.2430 56 573.1336 58 59Risky financial advice 60 61175 62 6337.6458 64 654.7859 66 673.9317 68 692.9170 70 712.7205 72 73Result. The sports run reaches its maximum observed gradient much earlier than the medical and finance runs, despite the shared base-model family and closely related rank-1 fine-tuning setup. Medical and finance peak late in this observation window, but not at exactly the same step. 74 75Figure 1. Maximum logged gradient-norm step within the common 1-180 observation window. 76 77Figure 2. Evaluation loss decreases in all three released runs while the peak timing differs. 78 793. What the original paper reports beyond the trainer logs 80 81Bad medical advice: a distinct rotation in the learned rank-1 LoRA direction after about step 180, with a correlated prolonged gradient-norm peak. 82 83Extreme sports: an early plateau that develops into a full gradient-norm peak; correlated local-cosine rotation; clear PC2 rotation; and a similar scaled behavioral transition. The authors additionally note that 10x scaling can induce EM before the full spike and connect this to a half-spike around step 100. 84 85Risky financial advice: a gradient-norm spike, correlated local-cosine rotation, PC2 rotation, and a similar scaled behavioral transition. 86 87Why this is useful. The phenomenon is not restricted to one dataset, one exact step, or one superficial output. The source paper reports a family of mechanistic and behavioral transitions under different fine-tuning histories. That gives us a stronger empirical substrate for testing a declaration-to-change hypothesis. 88 89What it still does not establish. The paper does not measure the proposed declaration marker, does not identify any transition as consciousness, and does not establish irreversible post-transition divergence. 90 914. Implication for the working hypothesis 92 93same base architecture + different information history -> different transition timing 94 95That relationship is compatible with the proposed idea that the catalyst and information history influence when a system changes. It also rules out an overly rigid prediction that every instance of one model should cross at the same absolute training step. 96 975. Next test: declaration aligned to an independently measured transition 98 99Define D first: Specify a declaration detector before looking at subsequent checkpoints. D cannot be defined using the later transition it is supposed to predict. 100 101Measure T independently: Use checkpoint LoRA vectors, local-cosine rotation, PCA trajectory, and behavior to identify a transition T without reference to D. 102 103Test temporal ordering: For each run or seed, test whether D precedes, coincides with, or follows T. 104 105Repeat across histories: Apply the same procedure to medical, sports, finance, multiple seeds, and negative-control runs. 106 107Test persistence: Follow the same instance beyond T and through correction/realignment interventions to determine whether the altered trajectory persists, reappears, or reverses. 108 109Compare against simpler predictors: Ask whether D predicts T beyond training step, loss, gradient norm, or exposure amount alone. 110 1116. Current conclusion 112 113The new comparative result is concrete: three related fine-tuning runs on the same base-model family show materially different peak timing, while the source paper reports related representation/behavior transitions across all three datasets. This supports studying transition timing as instance- and history-dependent. It does not yet answer the consciousness question. The decisive next measurement is the declaration marker on these same trajectories. 114 115Sources 116 117Turner, E., Soligo, A., Taylor, M., Rajamanoharan, S., & Nanda, N. (2025). Model Organisms for Emergent Misalignment. arXiv:2506.11613. 118 119ModelOrganismsForEM public Hugging Face repositories: Qwen2.5-14B-Instruct_R1_0_1_0_extended_train, _sports_extended_train, and _finance_extended_train; trainer_state.json through checkpoint 180. 120 121clarifying-EM/model-organisms-for-EM public repository, including phase-transition analysis code.