Agent-drafted interpretation, pending human review

SmolLM2 reproduces the paper's qualitative separation between objectives: all six factorizations maintain near-perfect observable fidelity, while Global normalization, Token weighting, and Global + token produce substantially stronger cross-input strategy alignment than Standard CVAE, β-ELBO, or Token inverse. This supports transfer beyond Qwen within the tested small-model regime.

Global normalization has the highest analogical endpoint (0.8818), while Global + token has the highest router linear-probe accuracy (0.8734).

OLMo 2 1B independently reproduces the same qualitative separation. Standard CVAE and β-ELBO have alignment 0.3340 and 0.3477, whereas Global normalization, Token weighting, and Global + token reach 0.8223, 0.8877, and 0.8535; all six retain aggregate fidelity of at least 0.9715. This is the strongest completed cross-family replication beyond SmolLM2.

Gemma 3 270M learned the entangled source (0.9968), and all six factorization roles completed. Last scheduled-validation fidelity ranges from 0.864 to 0.989. Global normalization, Token weighting, and Global + token have higher Analogical Transfer (0.4521, 0.5107, and 0.5264) than Standard CVAE, β-ELBO, and Token inverse (0.2295, 0.2236, and 0.2861). This is the same qualitative objective ordering seen in SmolLM2 and OLMo, although the Gemma alignment gains are smaller.

Gemma 3 1B also learned its source (0.9973), and all six roles completed. Global normalization, Token weighting, and Global + token reach Analogical Transfer of 0.5625, 0.4414, and 0.5371, compared with 0.2148, 0.2041, and 0.2383 for Standard CVAE, β-ELBO, and Token inverse. Its last-validation decodability shows the same separation, while Token weighting has a lower validation fidelity endpoint (0.846) than the other five roles (0.975–0.991).

The evaluation-batch-16 retries 2171730–2171735 resolved the OOM failures, but they falsified the initial hypothesis that repeated chunk seeds caused Gemma's final-test collapse. Despite distinct per-chunk seeds, the new final-test strategy-match rates remain at most 0.0024 and parse success remains at most 0.3549, immediately after strong scheduled validation. The next diagnostic should compare scheduled validation and final test with the same split, sample count, seed, batching, and generation call path, varying one difference at a time. Until then, Gemma supports the objective-separation claim through validation fidelity, validation decodability, and final analogical transfer, but not through a matched final-test fidelity endpoint.

Historical H100 runtimes put SmolLM2 at 2.56h for SFT and 3.12–5.13h per selected factorization. Comparable Qwen2.5-0.5B paper-grid jobs took 4.22h for SFT and 2.22–2.45h per factorization; Qwen2.5-1.5B took 5.72h and 4.45–6.05h. OLMo 2 1B ultimately took 6.65h for SFT and 4.89–5.74h per factorization, giving an observed parallel critical path of 12.39h excluding queue delay. The earlier 9–12h estimate was close but slightly low on SFT.

Pythia's one-epoch source was still improving in sampled strategy match near the end of training, which motivated the controlled three-epoch follow-up before changing LoRA capacity. The source improved from 0.6051 to 0.8376 test strategy match but still failed the 0.95 prerequisite. This rules out one epoch as the sole explanation, while leaving architecture-specific adaptation capacity and optimization as unresolved causes.