TRINITY: An Evolved LLM Coordinator

A Lightweight Coordinator Orchestrating Multiple LLMs via Evolutionary Strategy

Authors: Jinglue Xu, Qi Sun, Peter Schwendeman, Stefan Nielsen, Edoardo Cetin, Yujin Tang
Affiliations: Sakana AI, Japan; University of Michigan, USA; Institute of Science Tokyo, Japan
Published: ICLR 2026 | arXiv: 2512.04695

Motivation & Key Contributions

Scaling monolithic LLMs yields diminishing returns, and model merging is impractical across incompatible architectures and closed-source APIs. TRINITY adopts a macro-level approach: test-time model composition via coordination, fusing complementary strengths of multiple models without modifying their weights.

• Lightweight coordination: Hidden states from an SLM (0.6B) + a tiny head (∼10K params) coordinate diverse LLMs — total learnable parameters under 20K.
• Efficient training: sep-CMA-ES outperforms RL, SFT, and random search under tight budget constraints.
• State-of-the-art: 86.2% pass@1 on LiveCodeBench. Outperforms all baselines on 8 benchmarks (4 in-distribution, 4 held-out).

Why Evolutionary Strategy?

Training the coordinator is challenging: each evaluation requires a full multi-turn LLM rollout (expensive), the reward is binary (correct/incorrect), and parameters exhibit weak mutual coupling — each has only a tiny influence on the scalar reward.

• RL fails: REINFORCE gradients are low-SNR under binary rewards and weak parameter correlations, yielding unstable learning.
• SFT is intractable: Multi-turn label generation scales as O(7 · 3) per question — prohibitively expensive.
• sep-CMA-ES works: Derivative-free, diagonal covariance matches the block-separable structure. Improvement grows linearly with iterations (vs. logarithmically for random search). Only ∼1.5K–40K evaluations needed for ∼10K parameters.

sep-CMA-ES Results

sep-CMA-ES consistently outperforms all alternative training methods:

Coordination Pipeline

TRINITY processes queries over multiple turns. At each turn, a compact SLM reads the full transcript, and a lightweight head selects an LLM and assigns it one of three roles: Thinker (strategize), Worker (execute), or Verifier (evaluate). The process halts when the Verifier accepts the solution.

Conclusion

TRINITY demonstrates that a lightweight coordinator (<20K learnable parameters) can orchestrate diverse LLMs to achieve state-of-the-art performance. The results suggest a promising path forward: engineering collaborative AI ecosystems rather than scaling monolithic models.