Novix Achieves Scaled Co-Evolutionary RSI in Scientific Research

Novix and Co-Evolutionary RSI

Novix, founded by Shi Yu and Tang Jiabin with core members from Google, Microsoft, Meta, Zhipu, and Kimi, has operationalized Recursive Self-Improvement (RSI) at scale. The platform serves 200,000 experts across 40+ countries and 500+ institutions, including researchers from Tsinghua, Peking University, ETH Zurich, Cambridge, CMU, NTU, NUS, and the University of Tokyo. Novix calls its mechanism Co-Evolutionary RSI : AI proposes directions, explores solution spaces, executes tasks, and verifies results, while experts continuously feed back what problems matter, what evidence suffices, which evaluation rubrics to use, and which directions align with their taste and long-term judgment. Every adoption, rejection, correction, and feedback becomes a new target and standard for the next exploration round, further improving task decomposition, tool use, harnesses, and model capabilities.

The flywheel runs on real high-difficulty tasks that expose system capability boundaries; expert feedback pushes evolution; the evolved AI tackles harder problems. As Shi Yu states, "The system that truly scales RSI is the one that best leverages top human expert taste."

Three Frontier Tasks Demonstrating the RSI Flywheel

End-to-End Model Post-Training

Novix autonomously handles data construction, training code, SFT, GRPO/RFT, checkpoint comparison, and final delivery. On SynLogic's 135 fixed evaluation questions, Qwen2.5-7B accuracy rose from 8.89% to 37.78%. On ScienceWorld's 30 representative tasks, Qwen2.5-7B-Instruct average score increased from 11.2667 to 35.6333.

A critical judgment occurred when training loss kept decreasing but the model's score on real task probes dropped from 38.7 to 21.1 — the model was overfitting the training trajectory without becoming better at solving problems. Novix actively rejected that update, rolled back, and delivered the better-performing model. This illustrates that the key AI-for-AI capability is not continuing every training run, but judging whether the model truly became smarter and stopping, rolling back, or redirecting evolution when necessary.

Training data, logs, failure paths, and evaluation results flow back into Novix to improve its own task decomposition, tool selection, evaluators, and models, closing the AI-for-AI recursive loop.

Federated Dam Control under Multiple Constraints

In a simulation based on real river records, Novix allocated hourly spill volumes for eight dams on the Columbia and Snake rivers. Spillage helps juvenile fish bypass turbines but raises total dissolved gas, creating ecological risk; upstream adjustments propagate downstream with delay.

Within ~3.57 hours of effective compute, Novix executed 578 tool calls, repeatedly processing data, developing algorithms, running experiments, and correcting predictions and control strategies. It discovered a gas propagation time constant incorrectly set to 44 hours and re-derived it to 16 hours. It also found that reducing prediction error did not necessarily improve scheduling outcomes, so it abandoned cross-season unstable schemes and redesigned the optimization objective based on how prediction error affects scheduling cost. Candidate solutions were continuously validated — not just for prediction accuracy but for whether they yielded better decisions.

Novix produced a coordinated control strategy satisfying power generation requirements, gas caps, and complex operational rules. On 16 unseen test weeks, it scored 0.719264, 6.50% higher than the expert reference, with all control instructions passing execution constraint checks. The deliverable was not a generic scheduling suggestion but an executable controller outputting per-dam, per-hour decisions.

Predicting Steel Cyclic Response under Unknown Seismic Loading

Steel structures under earthquake loading experience repeated tension and compression; prior loading history affects subsequent material behavior. Novix started from existing tensile and cyclic tests, autonomously proposed constitutive hypotheses, modified numerical implementations, and repeatedly validated.

On 60 loading histories unseen during research, it achieved a held-out test score of 0.620034, approximately 1.99 times the expert reference score of 0.310910, with all instances returning valid results and exceeding the baseline. This demonstrates Novix can propose constitutive hypotheses, build models, calibrate parameters from limited existing tests, and transfer the learned material laws to never-before-seen loading paths.

These tasks share a common trait: answers cannot be generated in one shot; they must converge through experimentation, failure, and evaluation. Each convergence leaves data, failure paths, and expert judgments that flow back into the system as the starting point for the next exploration round — the RSI loop in real tasks.

Mechanism: Human Judgment as the Steering Variable

Novix learns each researcher's problem focus, evidence judgment style, and exploration preferences, then proactively proposes candidate directions and completes verification. Scientists intervene at key nodes to continue, pause, or correct, and can redefine evaluation rubrics; these judgments immediately seed the next exploration round.

In a typical materials research scenario, a PI need not write full daily tasks; they only judge which candidate directions deserve pursuit, which evidence is insufficient, which anomalies warrant follow-up. The AI continuously handles the remaining exploration, experimentation, and verification. As judgments accumulate, the system gradually forms an exploration path tailored to that researcher.

Tang Jiabin notes: "General models tend to propose near-identical question distributions; human expert feedback is the seed variable that determines where the RSI system goes when exploring the unknown." Longer collaboration yields more complete task distributions, judgment standards, failure records, and exploration trajectories, crystallizing into personal or lab-specific research assets and data moats retained in private spaces.

Product partner Shi Bei emphasizes the metric: "The measure of this system is not how many tasks it completes at once, but how far a single expert judgment can push the exploration." When AI drives the marginal cost of execution and verification near zero, the research ceiling is no longer "how many trials" but "knowing what is worth trying." The scientist's taste becomes the steering wheel of intelligence evolution.

Implications for Human-AI Relationship in the Era of Stronger Intelligence

The team's background — Shi Yu on Microsoft Copilot growth and MS AgentOS; Tang Jiabin on Kimi K2 post-training, AutoAgent, and AI-researcher — explains why they treat model improvement, agent systems, and long-term user collaboration as a single problem.

They observe that top experts are not static knowledge bases waiting to be extracted. Experts are persuaded by new evidence, intrigued by unexpected results, and abandon past explanations. AI's new discoveries change human judgment; human's new judgments change AI's next exploration. Co-Evolution lets both evolve together amid emerging new problems.

Even if agents surpass human intellectual boundaries, this collaboration continues to generate value. Open-ended task directions evolve through real tasks, expert feedback, and new scientific discoveries. More experts with diverse tastes entering the loop enrich the system's goals, standards, and reality constraints, expanding the reachable knowledge frontier.

Stronger base models raise everyone's capability floor but do not flatten different researchers' problem awareness. General capability determines how much AI can do; long-term co-evolved taste determines what problems those capabilities are applied to. These two axes are orthogonal: the stronger the base model, the greater the differentiation value of individualized exploration engines.

As RSI advances, humanity's scarcest judgment becomes the key variable in the intelligence recursion.

Novix has turned this co-evolution into an AI autonomous laboratory that every scientist can own. Here, AI tirelessly explores; scientists continuously harvest research outcomes and use their taste to decide what enters the next research round. This is the Co-Evolutionary RSI path Novix is validating — AI explores the unknown, humans choose the future.