Existing multi-agent LLM frameworks falter in knowledge-intensive scientific domains, hampered by static prompts, rigid workflows, and homogeneous model reliance. This leads to poor adaptation, limited reasoning, and significant latency on complex, long-horizon tasks. Crucially, these systems struggle to revise decisions when intermediate reasoning falters, undermining reliability in calculation-heavy scenarios. To overcome these hurdles, a new scientific domain-oriented interactive two-tier multi model orchestration framework has been proposed. This framework features a dedicated orchestration model that analyzes tasks, dynamically constructs domain-aware reasoning pipelines, and instantiates specialized expert agents with tailored prompts. An execution model then performs each step under generated role and instruction specifications.
Dynamic Pipeline Construction and Replanning
The core innovation lies in the orchestrator's ability to iteratively update the pipeline based on intermediate feedback. This enables dynamic replanning, role reallocation, and prompt refinement across multi-turn interactions. This adaptive approach significantly strengthens robustness and specialization for scientific reasoning through structured, heterogeneous model collaboration. The framework is model-agnostic, supporting integration of diverse LLMs with varying capacities or costs, thereby allowing for flexible performance-efficiency trade-offs in practical scientific deployments, as detailed on arXiv.
Enhanced Specialization and Reliability
Experiments demonstrate consistent improvements over existing multi-agent systems and strong baselines across diverse reasoning and scientific benchmarks. The proposed multi model orchestration framework offers a more flexible and reliable approach to complex scientific problem-solving by moving beyond static configurations towards dynamic, adaptive agent interactions. This represents a significant step towards unlocking the full potential of LLMs in specialized, high-stakes domains.
