Heterogeneous Agentic Coevolution

Jointly evolving specialized foundation models and LLM agents through reciprocal learning

Wenqian Ye1,2,* Ziwei Guan1 Zi Yang1 Shivani Modi1 Cici Cong Wu1 Yunlong He1 Yuxin Chen1 Zhengyu Zhang1 Liang Luo1 Venkatesh Ranganathan1 Jianhui Sun1 Yantao Yao1 Wenlin Chen1 Buyun Zhang1,† Ellie Dingqiao Wen1,† Henry Kautz2,† Aidong Zhang2,†
1Meta AI 2University of Virginia *Work done at Meta †Joint supervision
Paper How it works Results BibTeX
TL;DR Many domain-specific applications rely on specialized foundation models that possess strong domain expertise but remain dependent on human-designed improvements. HACE jointly evolves specialized foundation models and LLM agents through reciprocal learning, and employs evolutionary search to optimize their interaction harness. HACE outperforms the evaluated self-improving methods on the primary task metrics across recommendation systems, protein function prediction and crystal stability prediction.
15/15recommendation settings where HACE achieves the highest mean HR@10
+0.0460 vs +0.0040NDCG gain over Base on CAFA with ESM-2, compared with at most 0.0040 for existing evolution methods
+12.1%average relative HR@10 gain over Base across the 15 recommendation settings

Motivation

Specialized foundation models remain dependent on human-designed improvements

Specialized foundation models learn reusable domain knowledge through large-scale pretraining and acquire domain expertise that often exceeds that of general LLM agents. Yet they still cannot self-evolve like LLM agents. Once supervised training reaches convergence, researchers have to manually identify useful designs to improve the specialized foundation model.

From human-directed optimization to heterogeneous agentic coevolution. HACE brings together specialized foundation models and LLM agents into a shared learning process, where each helps determine how the other improves.

Most self-evolving systems still rely on proprietary LLMs whose parameters remain fixed while agent code and interaction harnesses evolve. For specialized foundation models, the challenge is for the LLM agent to keep improving the specialized model as the specialized model advances. An intervention that improves the initial model can become redundant after further training.

Continued progress requires the agent to learn continually with the model when it improves. We refer to this mechanism as reciprocal learning.

We propose Heterogeneous Agentic Coevolution (HACE), which coevolves a specialized foundation model (the specialist) with a heterogeneous LLM agent that learns to guide the coevolution. Agent learning is guided by verified gains over the current specialist. The agent's exploration, in turn, guides the selection of experiences for subsequent specialist training. An evolutionary interaction harness governs information exchange and training experience selection.

Method

An evolving harness connects specialist adaptation, joint selection, and agent learning

Let S denote the specialist, A the agent, and H the harness. Agent exploration selects specialist training experience. Joint validation retains the specialist and harness, after which fresh verified responses train the agent.

  1. Harness optimization

    We propose a harness H′ by modifying a parent from an archive of accepted harnesses. Its fields control the task information and specialist context shown to the agent, how responses are interpreted, and which examples train each learner.

  2. Specialist optimization

    Under H′, the agent receives a prompt constructed from a training input and the predictions of S, without access to its label. A fixed optimization budget produces S′ using the original supervised task loss. Agent exploration determines where to train, while task labels determine what the specialist learns.

  3. Joint selection

    With A fixed, we compare the proposed pair (S′, H′) and the current pair on the same validation batch. Joint prediction ranks candidates by s(c) + λa(c), with λ fit from {0, 0.25, 0.5, 1, 2, 4, 8} on one fold and evaluated on the other. The proposal is accepted when ΔJ > 0. Otherwise, we restore the specialist parameters, optimizer, calibration, and harness.

  4. Agent optimization

    We collect fresh responses with A under the retained pair. With qA and qS the respective task qualities, the response contribution is r = qA − qS. We keep the original response with the largest positive contribution for each eligible instance, and the agent trains on these with on-policy reinforced self-training.

A specialist update can change which responses improve on its predictions, and a harness update can change the information available to the agent. The agent's updated responses guide exploration in the next round.

Verification

Agent learning is guided by verified gains over the current specialist

We evaluate each response against the label using the same task constraints as the specialist. A response passes verification when its prediction is valid, supervision is evaluable, and r > 0. Sample a few responses below to see which ones would be kept.

Model-referenced contribution. Verification on one instance, with contribution r = qA − qS. We keep the response with the largest positive contribution.

Main results

HACE enables specialists to continue improving beyond supervised convergence

In each domain, we begin with specialists already trained to convergence. We compare against ReST-EM, Meta-Harness, OpenEvolve, DGM and TTT-Discover. By default, we use Gemma-4 E4B as the LLM agent.

Recommendation

HACE achieves the highest mean HR@10 in all 15 settings. Relative gains over Base range from 2.9% to 18.5%, averaging 12.1% across settings. On ML-1M, several evolution methods fall below the initial SASRec and HSTU models, while HACE improves all three architectures.

HACE achieves the highest mean HR@10 in all 15 settings. Relative change from Base, means on three random seeds. Tap or hover a point for the exact score.

Protein function prediction and material discovery

On CAFA, HACE raises NDCG from 0.6285 to 0.6745 with ESM-2 and from 0.6241 to 0.6739 with ProtT5-XL. Existing evolution methods improve over Base by at most 0.0040 and 0.0032, respectively. On WBM, all competing methods improve NDCG over Base, from 0.6388 to at least 0.6474. HACE reaches 0.6541, exceeding the strongest competing mean by 0.0028.

Main results on protein function prediction (CAFA) and material discovery (WBM). Means with standard deviations on three random seeds. The dashed line marks Base.

Evolution dynamics

HACE reaches the highest selection score in all three recorded trajectories

In HACE, each retained specialist and harness defines the starting point for the next round. The trajectories show when higher scoring configurations are found during search.

Evolution dynamics of three domains. Best recorded selection scores, measured by HR@10, Recall@10 and mean NDCG@3 and NDCG@5. Dashed lines mark the frozen specialist. These trajectories describe the scores observed during selection, while the main tables report aggregate evaluation results.

Case studies

Illustrations on events during coevolution

Recorded changes include removing genre metadata, targeting difficult training instances, and hiding candidate ranks from the agent. These examples illustrate how the harness search changes both the context used for prediction and the experience used for reciprocal learning.

Recorded selection scores and their running maxima from one run in each domain. Filled points mark selected updates. These trajectories describe the scores observed during selection, while the main tables report aggregate evaluation results.

Case studies on three domains

HACE recovers a matching product variant, places two relevant protein terms in the top five, and brings a second stable phase into the top three.

Specialist and combined predictions on three domains. Drag the protein and the crystal to rotate them. Protein structure from AlphaFold DB model AF-P00157-F1 (CC BY 4.0).

Ablation studies

All designed components contribute positively to the framework

Removing agent optimization causes the largest drop on Toys, from 0.0706 to 0.0622. On CAFA, removing harness or specialist optimization reduces NDCG by 0.0438 and 0.0435, respectively. Removing agent-guided experience lowers performance even when specialist optimization remains enabled, by 0.0044 on Toys, 0.0170 on CAFA, and 0.0034 on WBM. Removing joint selection also reduces all three scores.

Ablation studies in three domains. HR@10 for Toys/SASRec and mean NDCG for CAFA/ESM-2 650M and WBM/CHGNet.

Harness Engineering

The harness governs information exchange and training experience selection

Harness fields control the task information and specialist context shown to the agent, how responses are interpreted, and which examples train each learner.

Citation

@article{ye2026heterogeneous,
  title   = {Heterogeneous Agentic Coevolution},
  author  = {Ye, Wenqian and Guan, Ziwei and Yang, Zi and
             Modi, Shivani and Wu, Cici Cong and He, Yunlong and
             Chen, Yuxin and Zhang, Zhengyu and Luo, Liang and
             Ranganathan, Venkatesh and Sun, Jianhui and
             Yao, Yantao and Chen, Wenlin and Zhang, Buyun and
             Wen, Ellie Dingqiao and Kautz, Henry and
             Zhang, Aidong},
  journal = {arXiv preprint arXiv:2610.16078},
  year    = {2026}
}