Researchers behind RLM-27B report a 26.9% average improvement over Qwen3.8-27B on a suite of benchmarks involving large amounts of text. Both models ran in the same Recursive Language Model (RLM) harness—the software managing the model’s context and tools. The authors also report using roughly half as many inference tokens. The training setup included synthetic environments with fake tools and simulated calls to other models. Apparently, useful practice doesn’t require all the equipment to be real. Source
Alex L. Zhang announced the research on November 20, 2026. The paper, Language Model Harnesses Can Be Compositional Generalizers, is authored by Zhang, Souradip Chakraborty and Omar Khattab. The authors argue that harnesses can help models combine familiar skills to tackle unfamiliar tasks, rather than merely compensate for weaker models. Source Source
Teach the model to divide the work
Zhang’s complaint about conventional harnesses is that they keep accumulating the agent’s history in its prompt, the input sent to the model. That leaves the model to handle increasingly unfamiliar inputs. Adding sub-agents—other model calls assigned smaller jobs—isn’t automatically a fix: the model directing them can still end up outside the kinds of situations it handles reliably. More assistants do not, by themselves, constitute an organizational plan. Source
The authors instead describe harnesses that manage context programmatically. Their approach combines context offloading (keeping information outside the prompt), programmatic tool calling (using code to invoke tools) and calls to sub-agents. Source Source
The authors hypothesize that breaking a task into smaller parts can keep individual calls “locally in-distribution”: dealing with familiar kinds of problems even when the overall assignment is new. Source
In separate experiments, the authors report that RLM systems trained on short tasks generalized to versions 2–64 times longer. They also report that training in one domain transferred to others sharing an underlying structure. Source
The fake tools belong to training
The method behind the RLM-27B result is called surrogate harness training. The researchers target individual model calls using easier synthetic environments, including fake tools and simulated calls to other models, instead of training through a full, lengthy harness run. The idea is to teach reusable ways of breaking down work without repeatedly executing the entire expensive task. Source Source
Comments