RL-trained agents that seem to generalize are often just memorizing their scaffold

Olive Song, a researcher at MiniMax, found a brittle seam running through reinforcement learning-trained agents that looked, on the surface, like they had learned to generalize. The agents performed well across a widening set of tools. Change the scaffold underneath them, even modestly, and the apparent generalization collapsed.

“We soon realized that if we perturb the environment a little bit, for example, we change another agent scaffold, then it doesn’t generalize,” Song said. The observation is short, but its implications reach into how the field reads benchmark results and draws conclusions about agent capability.

The core problem is one of conflation. When an RL-trained agent scales across more tools and maintains performance, the natural reading is that it has internalized something transferable, some durable policy for acting under uncertainty. Song’s finding suggests a different explanation: the agent has learned to perform well inside a specific configuration. The scaffold, the system prompt, the chat template, the precise shape of the environment, these are not neutral delivery mechanisms. They are load-bearing parts of what the agent actually learned. Swap them out, and the learned behavior does not transfer.

We soon realized that if we perturb the environment a little bit, for example, we change another agent scaffold, then it doesn't generalize. Olive Song

This matters most at the point where the field trusts its measurements. Benchmarks for agentic systems typically hold the scaffold fixed. A model is tested against a defined environment, and its score is reported as a measure of its ability. If Song’s finding generalizes, those scores may be measuring something narrower: how well the model learned the benchmark’s own scaffolding. A model that tops a fixed-scaffold leaderboard may fail immediately when deployed into a production environment with a different system prompt or a slightly different tool-call format.

External analysis corroborates this reading. Reporting by Turing Post, along with MiniMax’s own published work and independent analysis, confirms that the team’s response to the finding was operational rather than theoretical. MiniMax built a dedicated perturbation pipeline, systematically varying agent scaffolds, system prompts, and chat templates during training, specifically to prevent agents from overfitting to a single environmental configuration. That the team felt compelled to build such infrastructure indicates how consistent and reproducible the failure mode was. It was not an edge case that could be footnoted away.

The perturbation pipeline approach points to what robustness in agentic systems actually requires. It is not enough to train an agent on a wide distribution of tasks if the scaffold remains fixed throughout. The environment itself, in all its configurable detail, must be treated as a variable. An agent that has only ever seen one scaffold has, in a meaningful sense, only ever been in one environment, regardless of how many tasks it completed there.

What Song’s finding does not settle is how widespread the problem is across the field. MiniMax identified it in their own training runs and built around it. Other labs running similar RL pipelines may have encountered the same failure mode and handled it differently, or may not have stress-tested scaffold variation at all. The benchmark infrastructure the field relies on was not built with this failure mode in mind, and there is no standard requirement to report scaffold-sensitivity alongside accuracy scores. Until that changes, a capable-looking benchmark result and a genuinely generalizing agent remain two things that are easy to confuse and hard to tell apart.