Answers are grounded in this paper and cite their sources.
Technical evaluation demonstrates limited reproducibility of internal representation instruments in frozen Qwen models, highlighting measurement boundaries despite perturbation sensitivity.
Technical report · Not peer reviewed This study asks when an internal representation observable is reusable across fresh local presentations rather than merely sensitive to a controlled perturbation on one substrate. It evaluates a phase-indexed binary register in a frozen Qwen checkpoint panel, separating explicit transition following, demonstration-conditioned register behavior, visible-state intervention use, operator-valued representation measurement, and representation–function coupling. Qwen3.5-27B and Qwen3.6-27B passed the behavioral ladder through visible overwrite propagation; Qwen3-14B and Qwen3.8-27B failed an upstream transition gate and were not tested downstream. On Qwen3.6-27B, a rank-4 local-frame paired ridge path-operator packet qualified on one bounded substrate. A subsequent prospectively frozen 20-block requalification retained broken-square sensitivity in 59 of 60 layer-blocks, but split-half path-map spectral re-estimation passed in only 35 of 60 layer-blocks. Only 5 of 20 blocks passed jointly at all three frozen layers, below the required 16 of 20 blocks and four blocks per rollout depth. Stage E therefore remained unopened. No functional outcome, association, coefficient, predictive increment, or permutation-test result was collected. The result is a measurement boundary for this particular rank-4, three-layer, paired-map route. It does not establish the absence of hidden state, internal naturality, useful representation geometry, or a neural causal-transition operator. The record contains the 13-page DOI-bearing technical report, buildable LaTeX source, and a model-forward-free reproducibility capsule with all 60 continuous Stage M layer-block records, deterministic figure/table inputs, model/runtime bindings, provenance closure, licenses, and checksums. Scope boundary: the report is limited to the frozen Qwen checkpoints, visible-state intervention family, rank-4 paired ridge coordinate maps, layers 21/43/62, and fresh 20-block requalification described in the manuscript. Stage E, hidden intervention, internal exact naturality, and representation–function coupling remain untested.
No takes yet. Share an insight, caveat, or question.
Aoi Kawasaki (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: