TopoAgent: A Self-Evolving Topological Agent for Multimodal Scientific Reasoning

Multimodal Large Language Models (MLLMs) often struggle with scientific reasoning because they rely on linear, sequential planning. This approach forces models to process long, noisy histories and often leads to visual-semantic misalignment or execution deadlocks when a task is too complex for a single tool call. TopoAgent addresses these issues by shifting from linear trajectories to a non-linear, topological framework.

TopoAgent introduces three primary innovations to improve reasoning reliability:

1. Visually-Grounded Atomic Decomposition: The system breaks complex queries into minimal, visually-grounded units called "atoms." This ensures that visual perception is anchored before any symbolic reasoning occurs.
2. Directed Acyclic Graph (DAG) Planning: Instead of a flat list of steps, atoms are organized into a DAG. This allows for strict context isolation, where each node only receives the specific, purified results from its direct prerequisites, preventing irrelevant historical noise from corrupting the reasoning process.
3. Adaptive Atomic Fission: If an atom cannot be solved within a set number of steps, the agent dynamically splits it into finer-grained sub-atoms. This allows the model to adapt its reasoning resolution at runtime, effectively bypassing capability boundaries that would otherwise cause a pipeline to fail.

Scientific reasoning requires a precise balance between visual perception and logical deduction. By decoupling these processes and allowing the agent to self-evolve its task granularity, TopoAgent achieves a state-of-the-art accuracy of 66.3% across mathematics, physics, and chemistry benchmarks. This framework demonstrates that non-linear, state-isolated planning is significantly more robust than traditional sequential methods for autonomous scientific discovery.