Comparative_Transition_Addendum_02.pdf

1[[PAGE 1]] 2Comparative Transition Addendum 02 3Same Base Architecture, Different Fine-Tuning Histories: A Three-Condition Training-Dynamics 4Comparison 5Scope. This addendum extends the first pilot by comparing three released Qwen2.5-14B rank-1 LoRA 6fine-tuning runs from the same research program: bad medical advice, extreme sports, and risky 7financial advice. The comparison uses released trainer logs through step 180 and the source paper’s 8independently reported phase-transition analyses. It does not classify any model as conscious. 9Core result: within the common steps 1-180 window, the maximum logged gradient norm 10occurs at step 169 for bad-medical-advice training, step 97 for extreme-sports training, and 11step 175 for risky-financial-advice training. The same paper separately reports related 12internal rotations and scaled behavioral transitions across these datasets, and explicitly 13notes an earlier sports half-spike around step 100. 141. Why this comparison matters 15If a common architecture can undergo related internal transitions at different times under different 16histories, then one universal absolute step is the wrong experimental target. The more appropriate 17question is whether each run has its own measurable transition trajectory, and whether a separately 18defined declaration event predicts that trajectory within the same instance. 19The source study uses a common Qwen2.5-14B-Instruct base and a minimal rank-1 LoRA setup to 20study emergent misalignment. For the bad-medical-advice run the authors report a distinct learned- 21direction rotation after roughly step 180, correlated with a prolonged gradient-norm peak. In 22Appendix G.5 they report comparable local-cosine/PCA rotations and scaled behavioral transitions 23for extreme-sports and risky-financial-advice fine-tunes. 242. Verified released-log comparison through step 180 25Condition 26Peak step 27Peak grad 28norm 29Eval loss 30@10 31@100 32@170 33@180 34Bad medical 35advice 36169 3740.9891 384.1683 393.5712 402.5904 412.4137 42Extreme 43sports 4497 4523.1330 464.8692 474.1374 483.2430 493.1336 50Risky 51financial 52advice 53175 5437.6458 554.7859 563.9317 572.9170 582.7205 59Result. The sports run reaches its maximum observed gradient much earlier than the medical and 60finance runs, despite the shared base-model family and closely related rank-1 fine-tuning setup. 61Medical and finance peak late in this observation window, but not at exactly the same step. 62 63 64[[PAGE 2]] 65Figure 1. Maximum logged gradient-norm step within the common 1-180 observation window. 66Figure 2. Evaluation loss decreases in all three released runs while the peak timing differs. 673. What the original paper reports beyond the trainer logs 68 69Bad medical advice: a distinct rotation in the learned rank-1 LoRA direction after about step 180, 70with a correlated prolonged gradient-norm peak. 71 72 73[[PAGE 3]] 74 75Extreme sports: an early plateau that develops into a full gradient-norm peak; correlated local- 76cosine rotation; clear PC2 rotation; and a similar scaled behavioral transition. The authors 77additionally note that 10x scaling can induce EM before the full spike and connect this to a half- 78spike around step 100. 79 80Risky financial advice: a gradient-norm spike, correlated local-cosine rotation, PC2 rotation, and a 81similar scaled behavioral transition. 82Why this is useful. The phenomenon is not restricted to one dataset, one exact step, or one 83superficial output. The source paper reports a family of mechanistic and behavioral transitions 84under different fine-tuning histories. That gives us a stronger empirical substrate for testing a 85declaration-to-change hypothesis. 86What it still does not establish. The paper does not measure the proposed declaration marker, does 87not identify any transition as consciousness, and does not establish irreversible post-transition 88divergence. 894. Implication for the working hypothesis 90same base architecture + different information history -> different transition timing 91That relationship is compatible with the proposed idea that the catalyst and information history 92influence when a system changes. It also rules out an overly rigid prediction that every instance of 93one model should cross at the same absolute training step. 945. Next test: declaration aligned to an independently measured transition 951. 96Define D first: Specify a declaration detector before looking at subsequent checkpoints. D cannot 97be defined using the later transition it is supposed to predict. 982. 99Measure T independently: Use checkpoint LoRA vectors, local-cosine rotation, PCA trajectory, 100and behavior to identify a transition T without reference to D. 1013. 102Test temporal ordering: For each run or seed, test whether D precedes, coincides with, or 103follows T. 1044. 105Repeat across histories: Apply the same procedure to medical, sports, finance, multiple seeds, 106and negative-control runs. 1075. 108Test persistence: Follow the same instance beyond T and through correction/realignment 109interventions to determine whether the altered trajectory persists, reappears, or reverses. 1106. 111Compare against simpler predictors: Ask whether D predicts T beyond training step, loss, 112gradient norm, or exposure amount alone. 1136. Current conclusion 114The new comparative result is concrete: three related fine-tuning runs on the same base-model 115family show materially different peak timing, while the source paper reports related 116representation/behavior transitions across all three datasets. This supports studying transition timing 117 118 119[[PAGE 4]] 120as instance- and history-dependent. It does not yet answer the consciousness question. The decisive 121next measurement is the declaration marker on these same trajectories. 122Sources 123 124Turner, E., Soligo, A., Taylor, M., Rajamanoharan, S., & Nanda, N. (2025). Model Organisms for 125Emergent Misalignment. arXiv:2506.11613. 126 127ModelOrganismsForEM public Hugging Face repositories: Qwen2.5-14B- 128Instruct_R1_0_1_0_extended_train, _sports_extended_train, and _finance_extended_train; 129trainer_state.json through checkpoint 180. 130 131clarifying-EM/model-organisms-for-EM public repository, including phase-transition analysis 132code. 133