Objective-Dependent Collapse and Noise Suppression in Recursive Self-Conditioning

This study investigates how recursive self-conditioning training or evaluating models on their own prior outputs affects model stability and latent dynamics across iterative feedback loops. Using nine controlled diagnostic experiments on low-dimensional autoregressive, non-autoregressive, and linear recurrent architectures, we systematically swept across loss objectives, noise levels, conditioning ratios, and structural bottlenecks. Key Claims: Objective-Dependent Collapse: Model collapse is not an inevitable outcome of recursive feedback objectives enforcing feature diversity or structural constraints reliably prevent variance decay and attractor collapse. Noise–Adaptation Trade-off: Input corruption acts as a smooth regularizer where increasing noise restores diversity at a direct cost to task adaptation. Selective High-Frequency Filtering: Iterative self-conditioning systematically damps high-frequency perturbations, functioning as a setting-specific noise suppressor. No Long-Range Memory: Evidence rejects a generalized reuse-driven memory mechanism, as initial positive signals failed to replicate under higher seed counts. Transferable Diagnostic Probes: Feature-matched permuted controls and bootstrap protocols effectively isolate genuine state reuse from global drift artifacts. Overall, these findings establish that while recursive self-conditioning does not induce long-range memory, it acts as a selective noise filter whose collapse dynamics can be controlled through task objective design and regularized input corruption. This work was conducted with the assistance of LLMs (ChatGPT, Gemini, Claude, Grok, and Qwen). For full scientific transparency, the author provides complete chat histories and reproducible Google Colab notebooks inside the paper