Technical Summary: Self-Training Doesn't Flatten Language — It Restructures It

Problem Statement

The prevailing characterization of successive self-training on language models (LLMs) is one of "flattening" or "model collapse," where diversity drops, distributions narrow, and text converges toward a low-entropy attractor. Concurrently, the literature on LLM linguistic fingerprints observes that machine-generated text is statically rich in specific surface markers (discourse connectives, hedges, em-dashes) compared to human baselines. However, neither literature has correctly described the mechanism of this distributional drift. The dominant theoretical account attributes decay to a frequency-based mechanism: rare events are under-sampled and die out. This paper argues that this characterization is incomplete and that the phenomenon is better described as a structural bifurcation rather than uniform flattening.

Methodology

The authors conducted controlled self-training experiments across five models spanning three architecture families (GPT-2 124M, Pythia-410M, Pythia-1.4B, Pythia-2.8B, and OPT-1.3B).

• Protocol: Each model underwent 11 generations of pure self-training. Starting from a released checkpoint, the model generated 3,000 texts (256 tokens each) using nucleus sampling ($p=0.95, T=0.9$), and the resulting corpus was used to fine-tune the model for one epoch. This cycle repeated without mixing in human data.
• Feature Selection: Seventeen linguistic features were selected a priori and stratified by Structural Depth ($d(\phi)$), defined as the minimum number of nested syntactic dependencies required to license the feature in a sentence:
  • $d=0$ (Surface/Lexical): Discourse markers, hedges, em-dashes, exclamation marks.
  • $d=1$ (Local Syntax): Regular past tense ($V+ed$), sentence-initial conjunctions, simple coordination.
  • $d=2$ (Clause Structure): Questions, parentheticals, passive voice, irregular past tense, relative/adverbial clauses.
  • $d=3$ (Cross-clausal/Mood): Subjunctive constructions.
• Analysis: The authors tracked the relative frequency of these features across generations. They computed per-feature decay rates ($\lambda$) via OLS regression and tested correlations against structural depth ($d$) and baseline frequency ($f$). A mixed-effects model was used to account for the nested structure of the data ($N=85$ feature-model pairs).

Key Contributions

1. Formalization of the Structural Depth Hypothesis (SDH): The paper proposes that the per-generation drift rate of a linguistic feature is predicted primarily by its structural depth ($d$) and secondarily by its dependence on sampling stochasticity ($\sigma$). The hypothesis posits a competition between a "depth penalty" (deep structures decay due to multiplicative probability costs) and a "sampling-dependence bonus" (surface markers amplify due to rich-get-richer dynamics in stochastic sampling).
2. Documentation of the "Superficial Complexity Paradox": The authors demonstrate that while aggregate "complexity" proxies (dependency-tree depth, type-token ratio, average word length) rise during self-training, the underlying clause-structural features (questions, passives, subordination) collapse. This masks the bifurcation in standard stylometric analyses.
3. Cross-Model Validation: The study provides controlled evidence across five models and three architecture families, showing that the depth-gradient effect is robust to architectural differences and training corpora.
4. Refutation of Frequency-Only Explanations: Through head-to-head statistical testing, the paper demonstrates that structural depth is a significantly stronger predictor of feature decay than baseline frequency.

Key Results

• Bifurcation of Features: Surface markers ($d=0$) amplified significantly (e.g., discourse markers increased by +126%, hedges by +44%), while deep syntactic structures ($d \ge 2$) collapsed (e.g., question marks dropped by 92%, passive voice by 56%, subjunctives by 53%).
• Monotonicity: The mean per-generation change was monotone decreasing with structural depth:
  • $d=0$: +24.9%
  • $d=1$: -10.0%
  • $d=2$: -47.2%
  • $d=3$: -52.7%
• Statistical Significance:
  • Pooled across five models ($N=85$), the Spearman correlation between structural depth and decay rate was $\rho = 0.540$ ($p < 10^{-6}$).
  • In contrast, the correlation between baseline frequency and decay rate was weak and non-significant in the pooled test ($\rho = 0.225, p = 0.039$).
  • A mixed-effects model confirmed the depth coefficient was highly significant ($p < 0.001$), while frequency was not.
• The "Exception that Proves the Rule": Exclamation marks ($d=0$) collapsed by 99.3%. The authors attribute this to the SDH equation: while they are shallow ($d=0$), their baseline rate under nucleus sampling was too low to trigger the "rich-get-richer" amplification loop, causing them to decay despite their low depth.
• Control Experiment: A matched control where models were fine-tuned on human text (OpenWebText) instead of self-generated text yielded a near-zero correlation ($\rho = 0.039, p = 0.88$), confirming the depth gradient is specific to the self-training regime.

Significance and Claims

The paper claims to reframe two distinct literatures:

1. Model Collapse: It offers a structural rather than purely statistical account of collapse. It argues that collapse is not merely the loss of rare tokens but the systematic erosion of deep syntactic machinery, driven by the multiplicative probability penalty of traversing nested dependencies in autoregressive generation.
2. LLM Fingerprints: It provides a mechanistic explanation for why canonical machine-text markers (discourse connectives, hedges) are exactly the features they are: they are the survivors and amplifiers of a depth-graded collapse.
3. Methodological Warning: The paper argues that relying on aggregate fingerprint metrics (as is standard in stylometry) systematically masks the bifurcation of language under self-training. This has direct implications for training-data curation (suggesting a need to prioritize clause-structurally rich text) and detection (suggesting that detectors relying on aggregate complexity may fail as models undergo recursive training).

The authors remain modest regarding the specific thresholds for larger models (7B+), noting their experiments span 124M to 2.8B parameters, but assert that the direction and ordering of the depth gradient are consistent across the tested scales. They also note that the findings are currently limited to English and specific prompt distributions.