HACE Co-Trains a Domain Specialist and an LLM Agent Across Three Domains

The authors of the Heterogeneous Agentic CoEvolution paper propose a scheme in which two different models improve each other at the same time: a narrowly specialized foundation model and an LLM agent. The framework is called HACE. It was tested in three areas: recommender systems, protein function prediction and crystal stability prediction. On the main task metrics, HACE outperformed all the self-improvement methods it was compared against. The abstract gives no specific figures for the gains, so the scale of the effect can only be judged after reading the full tables.

The problem: the specialist improves, but the way it is improved stays the same

The idea of self-improving LLM agents is no longer new: these are systems that rewrite their own prompts, code and strategies and get a verifiable gain. But in applied fields, the models doing the work are often not general-purpose LLMs but specialized models with deep domain expertise. A model for proteins or crystal structures knows its field better than any generalist. Yet people still improve it: they select the data, the architecture and the training regime.

The authors point to a key difficulty. As the specialist gets stronger, the strategies that pushed it forward stop working as well. What produced gains on a weak model may hit a ceiling on a strong one. The "improver" therefore has to change along with the thing it improves.

How the loop works

HACE builds a closed loop from two participants and the harness between them.

• The specialist is trained with supervision (supervised learning) on data selected during the loop.
• The agent goes through on-policy reinforced self-training. Its reward depends on the verified quality gain specifically on the current version of the specialist.
• The agent's exploration, meaning what it tries and finds, becomes material for the specialist's next training round.
• Evolutionary search optimizes the interaction harness: how the models exchange information and which training examples are selected.

The last item is arguably the most interesting. The harness is not fixed in advance. It evolves as the capabilities of both models grow. This answers the original problem: the rules of the game between the agent and the specialist are revisited at every stage.

The reward signal has one important detail. The agent is rewarded not for solutions that merely look plausible, but for confirmed improvement. This reduces the risk that the agent learns to produce data that is convincing but useless to the specialist.

What was tested and what is not yet visible

The choice of domains is telling. Recommender systems, protein function prediction and crystal stability prediction differ in the nature of their data, their metrics and the design of their models. The authors word their conclusion cautiously: heterogeneous coevolution may be a universal approach to improving specialized foundation models through interaction with a learning agent. Three domains are convincing as a demonstration of breadth, but "universality" remains a hypothesis for now.

What remains unclear from the abstract:

• the size of the gain on each metric and its statistical robustness;
• how much compute the whole loop costs compared with ordinary fine-tuning;
• which self-improvement methods were used as baselines and how strong they are;
• how the system behaves over the long run, when the specialist is close to its ceiling.

Practical takeaway

For teams that maintain their own domain models, the main idea of the paper is this: use the agent not as a one-off tool for picking hyperparameters, but as a permanent training partner rewarded for measurable progress. If you have a reliable automatic quality check, such as a benchmark with an unambiguous metric or laboratory validation, this scheme applies most directly. Without a verifiable signal, the loop loses its point.

Engineers should budget for the cost of experiments up front. Each round requires fine-tuning the specialist, training the agent and searching over harness configurations. Until the authors publish detailed costs, it makes sense to treat HACE as a promising research scheme, not a ready-made production recipe. The direction is clear, though: self-improvement is moving from general-purpose LLMs to narrow models, where results in science and recommendations depend on it.