Inference-time co-evolution

Inference-time co-evolution describes a team of AI agents that gets better at its work while it is in use, without retraining the AI model. Think of new colleagues who start out identical: after a failed job, the one who is good at that kind of work passes on a tip, and over time each becomes the go-to person for one type of task. In May 2026 this was lab research.

Where does the term come from?

Researchers call the moment a trained model is used "test time"; in practice that is inference. Two papers put "test-time co-evolution" in their titles in May 2026: TacoMAS on 10 May and EVOCHAMBER on 11 May. TacoMAS concludes that inference in a multi-agent system is "a temporal process of co-evolution". Co-evolution means that two things change together: what each agent knows, and how the agents work together. The entry on test-time co-evolution covers the same idea. In both papers the agents learn through text in their prompts, and the model weights stay the same.

How does EVOCHAMBER work?

EVOCHAMBER starts with a pool of 20 identical agents and lets them evolve over a stream of tasks at three levels:

• Individual: each agent keeps its own experience and an estimate of how good it is at each type of task.
• Team: for every task a selector puts together a team of three agents with complementary strengths, and a leader picks one of four ways of working together.
• Population: after a failed or disputed task, CoDream starts. The agents reflect together, write down insights and pass them from agents that are strong on that type of task to the weaker ones. Every 10 tasks the system also copies strong agents, merges near-duplicates, removes agents that keep underperforming and adds new ones.

With the open model Qwen3-8B, the authors report 63.9% on competition maths, 75.7% on code and 87.1% on a mixed stream of six task domains. On maths that is 32% better than the best comparable method, in relative terms. From identical agents, four to five stable specialists emerged.

How does TacoMAS differ?

TacoMAS treats a team of agents as a network: agents are the nodes, and the connections show who passes information to whom. A fast loop updates each agent's memory after every round. A slow loop lets a separate model add or remove agents and connections. The authors argue that skills should change quickly and the team structure slowly, to keep cooperation stable. TacoMAS evolves within a single task and clears its memory between tasks. On four benchmarks it scored 13.3 percentage points higher on average than the strongest of about 20 other multi-agent systems.

Where are the limits?

Both papers measured on benchmarks, from maths and code to web browsing and office tasks. EVOCHAMBER's authors put its inference cost at roughly 3.6 times that of a single agent, which may be too much for applications that need a fast answer. Which agent becomes the specialist differs from run to run, even though the pattern repeats. And CoDream can only pass on what someone in the team solved. Not every self-improving agent system is training-free either: Solvita (14 May 2026) keeps the language model fixed but trains a small knowledge network per agent with reinforcement learning.

Example

In EVOCHAMBER's maths run (382 tasks, Qwen3-8B), CoDream fired for the first time at task 11, after a team scored 0 out of 3. Three agents wrote insights such as: "When counting integers in a range divisible by multiple numbers, use the inclusion-exclusion principle with LCM adjustments." The same three agents wrote 72 of the 93 verified insights in that run. On the hardest competition problems, where nobody in the team found the answer, the insights turned into general advice about reading the question.