- The River training recipe improves RL-trained terminal agents’ generalisation by 106% on Terminal-Bench-Lite, using fewer than 30% of the training environments the TMax baseline requires.
- River’s core mechanism is environment filtering combined with process-level behaviour regularisation, rewarding intermediate reasoning steps, not just task completion, which pushes agents toward strategies that transfer across tasks.
- The “Agentic Compositional Generalisation hypothesis” underlying River argues RL shapes how agents route and sequence pre-trained skills, not how they acquire new ones, a reframe with direct implications for how training pipelines should be designed.
A training recipe called River has doubled generalisation performance for AI terminal agents, and does it with fewer training environments than the methods it outperforms. Posted recently to Hugging Face, the paper reports a 106% improvement on Terminal-Bench-Lite and a 30% average gain on the more demanding Terminal-Bench-v2.1, across model sizes from 2 billion to 27 billion parameters.
What RL Actually Does
Terminal agents operate through command-line interfaces and text-based feedback, making them natural candidates for automating developer and operations workflows. The problem is that training them typically relies on synthetic environments, since real user interaction data is scarce. That synthetic origin creates a domain gap: agents that perform well in training often fail to adapt when the task changes even slightly.
River’s central claim is that the standard response to this problem, scaling up the quantity and variety of synthetic environments, addresses the wrong variable. The paper proposes the “Agentic Compositional Generalisation hypothesis”: reinforcement learning for terminal agents works primarily by shaping high-level decision-making and routing, not by teaching low-level skills from scratch. Those lower-level skills are already embedded during pre-training and supervised fine-tuning. RL’s job, on this account, is to teach the agent how to sequence and combine what it already knows when it encounters a novel task.
If that framing is correct, the quality of the reward signal matters more than the size of the environment library. That is where River’s design choices follow from the theory rather than from convention.
Two Mechanisms, One Recipe
River implements its approach through two mechanisms applied together. The first is environment filtering: systematically identifying and removing low-quality synthetic environments from the training set. Noisy or poorly constructed environments push agents toward brittle, non-generalisable behaviours, so cutting them improves signal density rather than merely reducing compute. River achieved its benchmark gains using fewer than 30% of the training environments the TMax baseline requires.
The second mechanism is process-level behaviour regularisation. Standard RL for agents rewards successful task completion but says nothing about how the agent got there. River adds feedback on the quality of intermediate steps, not just the final outcome. Think of it as the difference between grading a student on their final answer versus grading their working: the latter catches reasoning errors that the former lets slide until the exam changes. That process-level signal guides agents toward problem-solving strategies that transfer across tasks rather than strategies that happen to work in one specific training scenario.
The Numbers Hold Across Scale
Performance gains were consistent across model sizes from 2 billion to 27 billion parameters and across different model families, agent harnesses and RL objectives. A method that only works at one scale, or only with one RL algorithm, has limited practical reach. River’s results holding across that range suggests the underlying mechanism, better reward signal quality and tighter environment curation, is not architecture-specific.
The 30% average improvement on Terminal-Bench-v2.1 is the harder number to dismiss. Terminal-Bench-Lite is a lighter evaluation; v2.1 is a more demanding generalisation test. Gains that hold on both suggest the method is not merely optimising for the easier benchmark’s distribution.
Where This Fits the Deployment Picture
According to Anthropic, its Agent Skills standard is designed to let agents interact with external SaaS tools such as Jira, Slack and GitHub across a user’s software stack after completing a code change. River and Agent Skills sit at different layers: River improves how agents learn to generalise during training; Agent Skills structures what they do once deployed. Whether better generalisation at training time reduces the need for rigid deployment pipelines, or whether the two remain complementary, is a question the paper does not address. Anthropic’s broader push on agentic development costs suggests the company is working both layers simultaneously.
