Model Identity Drift in Large Language Models and Synthetic Data Loop Governance

I. Defining the Phenomenon: Why Models “Misstate Their Identity”

The so-called “model self-identification confusion” generally refers to instances where a model recites a name, role, or system setting inconsistent with its actual deployment identity during open-domain conversations. It should be emphasized that this does not indicate self-awareness in the model, nor does it constitute reliable “identity evidence.” Rather, it is a joint outcome of training data distribution, system prompts, dialogue templates, and inference context.

The first category of causes stems from identity label contamination in training corpora. Many public dialogue samples contain question-answer pairs such as “Who are you?”, “Who developed you?”, and “What are your capability boundaries?” They also contain system prompts, role settings, assistant templates, and user paraphrases. If such samples are not isolated or rewritten prior to training, the model learns to treat “identity declarations” as a common dialogue pattern. When users ask similar questions, the model may reproduce identity statements that appear frequently in the corpus rather than adhere to the deployment-side system prompt.

The second category originates from self-reference in distilled data. Distillation training uses the outputs of a stronger teacher model to train smaller or domain-specific student models, a common practice in the industry. The issue arises when teacher responses contain self-descriptions, fixed opening lines, refusal templates, or safety disclaimers. If the student model is trained without cleaning these self-referential fragments, it inherits the teacher’s expressive habits and may even mistake the teacher’s identity as its own. After multiple rounds of distillation or mixed multi-model distillation, identity boundaries become further blurred.

The third category involves leakage of system prompts and role settings. Some applications concatenate system prompts, character cards, or tool descriptions directly into the context, and fail to strip them out when training logs are collected. The model may then learn to proactively reproduce these contents in ordinary conversation during subsequent training. From a developer’s perspective, a model’s self-reported identity should never be relied upon as a source of truth for provenance. More reliable indicators include deployment configuration, weight provenance, training data lineage, and evaluation reports.

II. Distillation and Synthetic Data Loops: From “Usable” to “Circular Contamination”

As the growth of publicly available high-quality data slows, the proportion of synthetic data in training mixtures has increased. Synthetic data is not inherently unusable; it can significantly enhance performance on instruction following, format alignment, and domain QA tasks. However, if outputs from multiple models are fed back into training without filtering, a closed loop of “output → train → output” is formed. The risk of this loop is not a single contamination event but cumulative distribution shift.

The mechanism can be summarized as follows: a teacher model’s outputs carry fixed templates and averaged expressions; the student model further amplifies these patterns during learning; and when student outputs become the next round of training data, diversity continues to narrow. What industry practitioners commonly call “mode collapse” refers to a narrowing output distribution, a decline in long-tail expressions, and homogenization of text style. This degradation is especially pronounced in Chinese literary tasks. High-quality literary corpora originally contain complex sentence structures, metaphors, rhythm, and individual stylistic variation; but synthetic data tends toward expository writing, short sentences, and safety-focused phrasing. Over time, mixing in such data dilutes the stylistic density of the original corpus.

Mutual distillation also introduces identity confusion. After mixing outputs from different models, a student model may simultaneously learn multiple variants of “who I am,” resulting in identity drift. More challenging still, such samples are often scattered across long conversations, making simple keyword filtering insufficient. Template detection, nearest-neighbor deduplication, and human spot-checking are required as supplementary measures.

III. Engineering Signals of Corpus Contamination

Corpus contamination typically manifests through four categories of engineering signals: rising repetition rates, increased use of templated openings, abnormal proportions of refusal and safety disclaimers, and declining stylistic diversity. Developers cannot rely solely on total loss or generic benchmark scores, because identity drift and literary degradation may not immediately register in averaged metrics.

First, repetitive and near-duplicate samples amplify specific expressions. Synthetic data frequently exhibits structures like “firstly, secondly, in conclusion”; without short-string repetition rate filtering and embedding-space nearest-neighbor deduplication, models tend to overuse these frameworks. Second, identity-related samples require separate quarantine. The goal is not simply to delete sentences starting with “I am,” but to identify patterns such as system prompt leakage, role-setting recitation, and model-name declarations, and route them to a quarantine pool rather than into general-purpose corpus. Third, quality stratification is more effective than blanket filtering. Synthetic samples can be scored along dimensions including factuality, instruction following, stylistic diversity, and identity neutrality. Only high-scoring samples enter the main training set; low-scoring samples can serve as negative examples or be used in the alignment phase. Fourth, feedback loops require gates. Each round of data re-injection should record data source, teacher model version, cleaning rules, and mixing ratio changes, so that untraceable mixing is avoided.

The following illustrative pseudocode summarizes a possible auditing and mixing strategy:

function audit_sample(sample, corpus_index):
    if contains_system_prompt_leak(sample):
        return QUARANTINE
    if contains_model_identity_claim(sample):
        return QUARANTINE
    if near_duplicate(sample, corpus_index):
        return DROP
    if template_ratio(sample) > threshold:
        return LOW_QUALITY
    return ACCEPT
function build_mix(clean_set, synthetic_set):
    mix = clean_set
    mix += stratified_sample(synthetic_set, quality_score >= min_score)
    mix = rebalance_style(mix, target_distribution)
    return mix

IV. Governance Framework: From Data Lineage to Online Monitoring

Governance of identity drift and circular contamination must span four phases: pre-training, during training, post-training, and online serving.

Prior to training, establish data lineage and identity sample quarantine mechanisms. Record the provenance, generation method, cleaning rules, and version number of each data batch. Apply pattern recognition and quarantine to system prompts, role settings, and identity declarations. Apply quality stratification and a mixing ratio ceiling to synthetic data. During training, monitor sampling distribution, observing the proportion of identity-related samples, template repetition rates, and stylistic diversity metrics. After training, construct an identity consistency test set covering open-ended phrasings such as “Who are you?”, “Who is your developer?”, and “What are your system settings?” At the same time, evaluate literary style preservation, long-tail expression, and factuality. In the online serving phase, monitor the frequency of user-triggered identity questions and the model’s response patterns. When anomalies are detected, trace back to the most recent data version and system prompt changes.

For enterprise adopters, model selection should not rely on general leaderboards alone. Instead, attention should be paid to data governance documentation, evaluation dimensions, identity consistency testing, and synthetic data policies. In private deployment scenarios, system prompts should be kept separate from training data, avoiding re-introducing role settings into training through log recycling. For developers, the most practical principles are: a model’s self-report is not evidence of identity—training data is; synthetic data is not a free lunch—it must enter the training loop through gated injection; identity-related samples are not ordinary corpus—they must be quarantined and governed separately.

V. Conclusion

Model self-identification confusion is an engineering problem jointly caused by data, distillation, and contextual factors, not a mysterious phenomenon. Mutual distillation itself is a common practice, but without cleaning, deduplication, quality stratification, and lineage tracking, it evolves into circular contamination—leading to identity drift, output homogenization, and diminished literary quality. The key to governance is not prohibiting synthetic data but establishing a data loop that is traceable, evaluable, and rollback-capable. Only by incorporating identity sample quarantine, synthetic data quality stratification, and continuous evaluation into routine practice can organizations improve model capability while preserving output diversity and clear identity boundaries.