A Korean AI startup's "genetic recombination" approach challenges the assumption that smarter models always require more compute.
TL;DR: VIDRAFT, a Korean AI startup based in Seoul, has introduced the Darwin Family framework, which improves large language model reasoning capabilities through parameter recombination rather than additional gradient-based training. The flagship model, Darwin-27B-Opus, achieved a score of 86.9% on the GPQA Diamond scientific reasoning benchmark, ranking sixth among 1,252 evaluated models at the time of publication. The research was released as a preprint in May 2026 under arXiv:2605.14386.
VIDRAFT, a Seoul-based Korean AI startup, is drawing international attention after China's TechWalker spotlighted the company's Darwin Family framework — a model-merging system that elevates AI reasoning performance without running any additional training. Published in May 2026 as arXiv preprint 2605.14386, the work proposes that the intelligence latent inside already-trained models can be unlocked and amplified by intelligently recombining their parameters, much like selective breeding in biology.
The Darwin Family framework treats model improvement as an evolutionary process rather than a training problem. Instead of fine-tuning or running reinforcement learning, VIDRAFT's research team merges two existing models — referred to internally as a "Father" model and a "Mother" model — by selectively blending their respective parameter layers. Both parent models share the same pre-trained foundation but have been developed in different capability directions afterward.
The core insight driving the framework is that reasoning ability is largely established during pre-training and is embedded in specific internal layers of a model. Subsequent fine-tuning can shift surface behavior without fundamentally altering those underlying reasoning circuits. By diagnosing which layers carry the most reasoning signal and then mixing parent-model parameters at those layers with greater weight, VIDRAFT argues that a merged offspring model can outperform either parent.
Three mechanisms work together to make this possible. First, an MRI Importance Scoring system scans each parameter layer and assigns it an importance score based on two components: static statistical properties of the parameters (such as entropy, variance, and norm), and dynamic probe responses measured by comparing how each layer activates when handling reasoning tasks versus general conversational inputs. Each component contributes equally to the final score, giving the framework a principled basis for deciding how much of each parent's parameters to use in any given layer.
Second, an evolutionary search process refines those initial blending recommendations. Each candidate merge configuration is encoded as a 14-dimensional vector — a "genome" — that controls mixing ratios across different module types (attention layers, feed-forward networks, embedding layers, and so on), as well as sparsification settings and block-level mixing ratios. The algorithm iteratively evaluates candidate genomes, retains higher-performing configurations, and applies crossover and mutation operations to generate improved candidates, mimicking natural selection.
Third, an MRI Trust Fusion parameter called τ (tau) balances the relative influence of the diagnostic scores and the evolutionary search at each step. Rather than being set manually, τ is itself optimized by the evolutionary process. According to the research, τ consistently converged between 0.35 and 0.55 across all tested model sizes, indicating a stable, learned balance between guided diagnosis and free exploration.
The result is Darwin-27B-Opus, the framework's flagship model, which scored 86.9% on the GPQA Diamond benchmark — a rigorous scientific reasoning test — and placed sixth among 1,252 models in the evaluation ranking at time of submission, surpassing its fully trained parent model despite receiving no gradient-based training of its own.
Training frontier AI models costs millions to tens of millions of dollars. If a framework like Darwin can meaningfully improve reasoning performance through parameter recombination alone, it opens a more resource-efficient path to capability gains — one that doesn't require massive compute budgets or proprietary datasets. For the broader research community, the result also reinforces a growing body of evidence suggesting that reasoning ability is a structural property baked in during pre-training, not something that emerges primarily from later fine-tuning.
For VIDRAFT specifically, international coverage in outlets like China's TechWalker signals that the Darwin Family approach is attracting scrutiny well beyond the Korean AI ecosystem, positioning the startup as a contributor to one of the field's most active research conversations: how to make models smarter without simply making them bigger or more expensively trained.
Q: What is VIDRAFT's Darwin Family framework?
A: It is a model-merging system developed by Korean AI startup VIDRAFT that improves large language model reasoning by recombining the parameters of two existing models, without any additional training.
Q: How did Darwin-27B-Opus perform on benchmarks?
A: Darwin-27B-Opus achieved 86.9% on the GPQA Diamond scientific reasoning benchmark, placing sixth among 1,252 models evaluated at the time of the preprint's publication in May 2026.
Q: Where can I read the original research?
A: The work is publicly available as a preprint on arXiv under the identifier arXiv:2605.14386.
Source: TechWalker (중국) (2026-05-21) — original article