Test-time co-evolution means that a team of AI agents keeps improving while it is in use, without retraining the AI model. The agents learn which of them is good at what, who works well together and which lessons to pass on, through text in their prompts and a few scores. The promise is a team that adapts to the tasks it gets. In May 2026 this was lab research.
Where does the term come from?
In machine learning, "test time" is the phase in which a trained model is used instead of trained; in practice that is inference. Two papers put "test-time co-evolution" in their titles in May 2026: TacoMAS (Xu et al., 10 May) and EVOCHAMBER (Zhang et al., 11 May). Co-evolution means that several things change together. In the words of the EVOCHAMBER authors, a multi-agent system evolves "who collaborates, how they collaborate, and how knowledge flows across the population". The entry inference-time co-evolution describes the same idea under another name and compares both papers; this entry follows EVOCHAMBER step by step.
How does EVOCHAMBER do it?
The pool starts with 20 identical agents on the same model. For each task in a stream:
- A selector forms a team of three for that type of task (a niche). The anchor is the agent with the best record on the niche and also leads. The complement adds skills the anchor lacks and has worked well with it before. The scout has little experience on the niche, so every agent keeps getting tasks.
- The leader picks one of four ways to work together: voting, debate, generator and critic, or splitting the task. It learns this choice from a shared log of earlier rounds.
- The team's score counts for all three members and updates each agent's competence per niche and how well pairs work together.
- Each agent writes a lesson for the niche and a general insight, and later retrieves the most similar ones into its prompt.
- After a failed or disputed task, CoDream runs: the team reflects together and passes lessons from agents that are strong on that niche to weaker ones.
- Every 10 tasks, the system copies strong agents, merges near-duplicates, removes agents that keep lagging behind and adds new ones.
The model weights never change. Everything the pool learns sits in prompts, scores and the roster of agents.
What did it achieve?
With the open model Qwen3-8B, the authors report 63.9% on 382 hard maths problems, against 48.4% for the best comparable method (MemCollab) and 29.8% for a single agent. That 32% relative gain applies to maths only. On code the pool reached 75.7%, and on a mixed stream of six domains 87.1%. Without CoDream, the mixed stream fell to 76.3%, the largest drop of any part. With GPT-4.1-mini the gain over a single agent was smaller: 10.7 points on hard maths, against 34.1 with Qwen3-8B.
Where are the limits?
- It costs about 3.6 times the inference of a single agent, which the authors call possibly prohibitive where speed matters.
- The evidence is narrow: two model families, benchmark streams of up to 600 tasks and fixed thresholds for adding and removing agents.
- Which agent becomes a specialist differs per run, although four to five specialists emerge each time.
- Whether lessons go out of date on longer streams is left to future work.
Example
In one maths run (382 tasks), each type of competition problem ended up with its own dominant agent, and a different one for each year of the AIME competition, although all agents started identical. The specialist for AIME 2024 had never led a team until those problems arrived. The harder the problems, the more the pool relied on one anchor: the paper's specialisation index rose from about 0.1 on the general hard maths set to about 0.3 on AIME 2024.