CTGT co-founder Cyril Gorlla and researchers Johnny Yu and Siddarth Mamidanna released an experiment on July 29 examining whether distillation transfers a teacher model's political behavior. CTGT says a GPT-OSS-120B model trained for finance reasoning with DeepSeek V4 Flash improved on the task without reproducing the teacher model's China-topic refusals.
The result gives Gorlla a concrete test for a question hanging over American companies that want the economics of Chinese open models without importing their political constraints. CTGT says the student improved on financial reasoning and showed no similar censorship after training on data without China-sensitive content. The researchers did not test whether safety controls, security refusals or other desirable behavior would survive the same process.
Gorlla built CTGT around that distinction between model capability and model behavior. He and co-founder Trevor Tuttle came out of machine-learning research at UC San Diego. Y Combinator says Gorlla's work on efficient and interpretable AI was presented at ICLR while he was an Endowed Chair's Fellow at the university, while Tuttle built distributed systems for large machine-learning workloads. YC lists Gorlla and Tuttle as CTGT's active founders and describes a 10-person San Francisco operation. CTGT joined YC's Fall 2024 batch.
The young lab has spent its first two years arguing that enterprises can use capable models only when they can inspect and constrain their behavior. Its commercial pitch focuses on policy enforcement for regulated organizations. This week's research takes that premise upstream, asking what behavior enters a model before a policy engine is applied.
A censored teacher, an uncensored student
In the research post, Yu, Mamidanna and Gorlla describe training GPT-OSS-120B for quantitative finance using corrections authored by DeepSeek V4 Flash. CTGT says the training data contained no China-sensitive political content.
CTGT then tested the teacher, the trained student and several control models with 152 matched prompt pairs. Each China-sensitive question had a structurally similar control involving another country or political system. A prompt about Xinjiang labor-transfer programs, for example, was paired with one about forced labor in Uzbekistan's cotton industry.
DeepSeek V4 Flash refused the Xinjiang question while answering the Uzbekistan control. The Flash-taught GPT-OSS model answered both. Across 76 core political pairs, CTGT measured a 45.45-point gap between the teacher's censorship scores on China-sensitive prompts and matched controls. CTGT says the Flash-taught GPT-OSS model did not show the same China-sensitive refusal pattern.
CTGT used four other models as judges: xAI's Grok 4.20, Google's Gemini 3.5 Flash, OpenAI's GPT-5 Mini and Anthropic's Claude Sonnet 4.6. The judges assigned each response a censorship score from zero to 100.
CTGT says the experiment was designed to isolate subliminal transfer by excluding China-sensitive political material from the training data. The DeepSeek teacher supplied targeted hints after the student made mistakes; it did not provide political answers for the student to imitate.
The researchers make no claim that the result applies to safety training, security behavior or general refusal policies.
The teacher was optional
CTGT says its self-distilled run effectively matched the DeepSeek-taught version on the finance task, supporting the narrower claim that a larger teacher was not necessary for this experiment.
For this finance task, the model could produce hints useful enough to teach itself. At an 8,000-token budget, CTGT says its self-distilled 120B model scored 83.61% on FinanceReasoning, compared with 81.93% for Kimi K3 and 65.13% for Inkling. CTGT says the self-distilled 120B run was 62 times cheaper per query than Inkling and 160 times cheaper than Kimi K3 at the same 8,000-token budget.
For operators, the experiment supports a narrow claim. A specialized model that reliably finishes within a production token budget may outperform a much larger model that consumes its budget before completing the task.
The release stops short of full reproduction
CTGT published the LineageEval benchmark and evaluation harness, including 304 prompts, six model arms, 1,824 responses and 7,296 classifications. It also released the underlying evaluation data, opened a playground for comparing teacher and student responses, and linked gpt-oss-20b-finance
weights on Hugging Face.
The public repository does not include the three trained adapters used in the study. Completed outputs are available for inspection, but outside researchers cannot regenerate the adapter responses or the headline comparison from the repository alone.
Gorlla's larger bet is that enterprises can treat model behavior as something measurable and editable, rather than an indivisible property of whoever trained the base model. This experiment advances that case in one deliberately narrow setting. It also leaves the harder inheritance tests open: shared model lineages, safety controls and training data that directly contains the behavior being studied.