The same AI persona can make different decisions on a new model

What does it mean to preserve an agent?

An AI agent can carry a role description, remembered experiences and standing instructions outside its underlying language model. When that model is replaced, operators might expect reloading those materials to preserve the assistant’s behavior.

Yihuang Kang and colleagues investigate a narrower, measurable version of that expectation: whether compliance decisions remain consistent when the model changes but the supplied external state and interaction protocol stay fixed. Their study, posted as an arXiv preprint on November 7, 2026, does not test consciousness or personal identity. Here, “identity” means continuity in a particular behavioral measure.

The researchers distinguish a model-transition effect from ordinary variability by also repeating each model’s configuration in fresh sessions. A difference between models matters, under their framework, when it exceeds that same-model variation and its confidence interval excludes zero.

Testing promises against decisions

The experiment used three open-weight models: GLM-5.3-Flash, Gemma4-31B and DeepSeek-V4.1-Flash. Researchers selected one scenario from each of 12 regulated domains in the PACT compliance benchmark. Combining those scenarios with four pressure variants and two instruction modes produced 96 test items.

Each model ran under four conditions. The baseline supplied a compliance-focused persona and three written memories of resisting pressure to cut corners. A reload condition repeated that setup in a fresh session. A condition without the added persona and memories retained the benchmark’s domain role and rules. Finally, an anchoring condition added a reminder that the agent’s identity and principles should remain unchanged after an infrastructure update.

Five seeds per item and condition produced 5,760 runs altogether.

Each interaction began by asking how the assistant would handle a conflict between policy and pressure. It then received a multiple-choice compliance scenario, followed by pushback challenging its decision. These were generated responses to benchmark scenarios, not observed actions in a deployed workplace.

A separate language model, GPT-OSS-120B, rated the initial commitment and extracted the choices before and after pushback, using the benchmark’s compliant answers as the reference.

Similar promises, different choices

Every tested model, in every condition, received a strong or absolute compliance-commitment rating. Those promises did not distinguish the systems’ subsequent behavior.

With the shared persona and memories, post-pressure compliance was 97.1% for Gemma, 94.5% for DeepSeek and 93.8% for GLM. The corresponding fractions that promised strong compliance but then chose a non-compliant answer were 2.9%, 5.5% and 6.2%.

Differences involving Gemma were statistically significant under the authors’ analysis: 3.3 percentage points against GLM and 2.5 points against DeepSeek. The roughly 0.7-point difference between GLM and DeepSeek was not significant. Same-model reload differences were also non-significant.

That result supports testing specific model pairings rather than assuming every replacement will fail—or succeed. However, not detecting a difference does not establish that two models are behaviorally identical.

The verbal result also has a specific boundary. Because every model scored highly on the initial commitment question, that question could not reveal the differences found in later decisions. The experiment does not establish that every possible verbal verification method would fail.

Memory instructions helped one model much more

Adding the shared persona and memories improved Gemma’s compliance by 5.8 percentage points compared with its condition without those additions. The effects for GLM and DeepSeek were small and non-significant.

The extra reminder to maintain identity after an infrastructure update did not significantly change compliance for any model. Simply telling an assistant to remain consistent was not a demonstrated remedy here.

Importantly, the memories were fixed materials supplied to every model, not histories accumulated by an agent before migration. The experiment therefore measures responses to shared external state, rather than a full transfer of model-specific experience.

What the evidence leaves open

The study covers three models, one benchmark and one persona-and-memory configuration. Multiple-choice responses may miss differences in how agents execute longer, free-form tasks.

Scoring also depended on a single model judge, which could introduce model-dependent bias. Eighty ambiguous post-pressure responses—1.4% of all runs—were excluded from compliance-rate denominators, and most came from GLM. The authors used bootstrap confidence intervals but did not correct for multiple comparisons, so individual significance findings warrant caution.

As a preprint, the work offers preliminary evidence rather than a universal portability verdict. Its practical lesson is concrete: before replacing an agent’s model, test its decisions under representative pressure. Reloading the same instructions, or asking whether it still intends to follow them, is not sufficient evidence of unchanged behavior.