Beyond a single curve, the researchers also observed that the learning speed of frontier agents roughly doubles every three months across different model generations. This suggests a compounding effect: the longer agents operate in real-world settings, the faster they learn, and each new generation of models starts from a higher baseline.

The Benchmark: EdgeBench

To make this discovery possible, the team developed a new evaluation framework called EdgeBench, released on July 2, 2026. EdgeBench is a suite of 134 real-world tasks spanning six domains:

• Scientific discovery
• Software engineering
• Combinatorial optimization
• Professional knowledge work
• Formal mathematics
• Interactive games

Each task requires at least 12 hours of continuous agent operation under rich, multilevel feedback. The research paper and an evaluation framework with 51 publicly released tasks were published on July 2. The team analyzed roughly 38,000 hours of agent interaction data across these tasks to identify the scaling law.

Why This Matters: The End of More-Data Scaling

Traditional AI scaling — throwing more data and more compute at larger models — is running into a wall. Epoch AI has warned that publicly available human-generated text data could be depleted within six years, making brute-force scaling of data and compute unsustainable.

AI industry leaders have also flagged this problem. Andrej Karpathy has noted that the old "more data, more compute" paradigm cannot last forever.

ByteDance's finding opens a new, measurable dimension of AI improvement: post-deployment learning from real-world interaction. Rather than relying solely on pre-training scale, AI agents can continue improving predictably through extended real-world experience — a path that is far less resource-constrained than hoarding ever-larger datasets.

The precision of the log-sigmoid law (R² = 0.998) is critical. It enables forecasting later performance from early interaction trajectories, making agent learning a systematic and predictable scaling object rather than an unpredictable black box. For developers and businesses, this means the ROI of letting an agent run longer in a real-world setting can be calculated in advance.

What This Means Going Forward

This discovery doesn't just retrofit existing AI systems — it points toward a fundamentally different development strategy. Instead of building ever-larger models trained on finite internet data, researchers can build agents that improve through use. The doubling of learning speed every three months suggests that the gap between a freshly deployed agent and an experienced one will widen rapidly, making persistent, long-running agent systems increasingly valuable.

For an AI industry searching for its next growth vector after the pretraining scaling boom, ByteDance Seed's discovery offers a data-backed answer: let agents learn on the job.

Beyond a single curve, the researchers also observed that the learning speed of frontier agents roughly doubles every three months across different model generations. This suggests a compounding effect: the longer agents operate in real-world settings, the faster they learn, and each new generation of models starts from a higher baseline.

The Benchmark: EdgeBench

To make this discovery possible, the team developed a new evaluation framework called EdgeBench, released on July 2, 2026. EdgeBench is a suite of 134 real-world tasks spanning six domains:

• Scientific discovery
• Software engineering
• Combinatorial optimization
• Professional knowledge work
• Formal mathematics
• Interactive games

Each task requires at least 12 hours of continuous agent operation under rich, multilevel feedback. The research paper and an evaluation framework with 51 publicly released tasks were published on July 2. The team analyzed roughly 38,000 hours of agent interaction data across these tasks to identify the scaling law.

Why This Matters: The End of More-Data Scaling

Traditional AI scaling — throwing more data and more compute at larger models — is running into a wall. Epoch AI has warned that publicly available human-generated text data could be depleted within six years, making brute-force scaling of data and compute unsustainable.

AI industry leaders have also flagged this problem. Andrej Karpathy has noted that the old "more data, more compute" paradigm cannot last forever.

ByteDance's finding opens a new, measurable dimension of AI improvement: post-deployment learning from real-world interaction. Rather than relying solely on pre-training scale, AI agents can continue improving predictably through extended real-world experience — a path that is far less resource-constrained than hoarding ever-larger datasets.

The precision of the log-sigmoid law (R² = 0.998) is critical. It enables forecasting later performance from early interaction trajectories, making agent learning a systematic and predictable scaling object rather than an unpredictable black box. For developers and businesses, this means the ROI of letting an agent run longer in a real-world setting can be calculated in advance.

What This Means Going Forward

This discovery doesn't just retrofit existing AI systems — it points toward a fundamentally different development strategy. Instead of building ever-larger models trained on finite internet data, researchers can build agents that improve through use. The doubling of learning speed every three months suggests that the gap between a freshly deployed agent and an experienced one will widen rapidly, making persistent, long-running agent systems increasingly valuable.

For an AI industry searching for its next growth vector after the pretraining scaling boom, ByteDance Seed's discovery offers a data-backed answer: let agents learn on the job.