Skip to content

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

🕒 Published (v1): 2026-07-16 07:22 UTC · Source: Arxiv · link

Why this paper was selected

Self-evolving topological agent for multimodal scientific reasoning; novel non-linear planning architecture

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

TopoAgent replaces the standard linear-trajectory agent paradigm with a Directed Acyclic Graph (DAG) of visually-grounded atomic sub-tasks, enforcing strict context isolation so each node consumes only its prerequisite outputs. A runtime Adaptive Atomic Fission mechanism dynamically fractures failing nodes into finer sub-atoms, enabling self-correcting long-horizon scientific reasoning without global replanning.

Problem

Linear agentic frameworks (CoT, ReAct, sequential Plan-and-Solve) suffer three structural failures in multimodal scientific reasoning: (1) visual-semantic misalignment — perception and logical deduction are entangled in a single generation step, causing hallucination of visual variables; (2) long-context hallucination — blindly concatenating all historical outputs into the prompt injects irrelevant noise that corrupts attention in downstream equation solving; (3) static capability horizons — fixed task granularity causes deadlocks when a scheduled sub-task exceeds the assigned tool's capability, with no recovery path.

Method

Visually-Grounded Atomic Decomposition. A Decomposer module \(D\) fractures the query \((Q, I)\) into atoms \(A = \{A_1, \ldots, A_n\}\), each a minimal executable unit that mandates spatial visual grounding before any logical deduction.

DAG Planning with Context Isolation. Atoms are organized into a DAG \(G = (A, E)\) encoding prerequisite dependencies. When executing \(A_i\), the system constructs a purified local context: $\(C(A_i) = \bigcup_{A_j \in P(A_i)} M(A_j)\)$ where \(M\) stores completed atom states and \(P(A_i)\) is the direct prerequisite set. All irrelevant historical context is excluded.

Adaptive Atomic Fission. Atoms are managed in a dynamic deque \(Q_\text{exec}\). If \(A_i\) fails to reach a STOP state within \(T_\max\) steps, a fission function \(F\) analyzes the failure trace and decomposes: $\(F(A_i, \text{trace}) \to \{A_{i.1}, A_{i.2}, \ldots, A_{i.k}\}\)$ Sub-atoms are prepended to \(Q_\text{exec}\) with hierarchical namespace mapping \(M_\text{virtual}(A_i) = \bigoplus_k M(A_{i.k})\), preserving global DAG integrity without recompilation.

Key Contributions

  • Identification of three structural failure modes in linear agentic planning for multimodal scientific reasoning.
  • TopoAgent: a DAG-based execution framework that decouples perception from reasoning via visually-grounded atomic decomposition and topological context isolation.
  • Adaptive Atomic Fission: a runtime self-evolution mechanism that dynamically degrades task granularity to match tool capabilities, recovering 51.5% of failed math nodes vs. 18.3% for Self-Refine and 2.7% for No-Retry.
  • Empirical demonstration that DAG-enforced context isolation prevents the "context explosion" that causes accuracy collapse in deep inference steps.

Results

  • Global average accuracy across 6 MLLMs: TopoAgent 66.3% vs. AutoGen 62.0%, OctoTools 62.8%, GPT-Functions 58.4%, LangChain 58.1%, Original LLM 57.0%.
  • OlympiadBench (Physics), global avg: TopoAgent 41.2% vs. 30.9% (Original LLM); with GPT-5: TopoAgent 60.3% vs. 45.8% (Original LLM) and 40.2% (GPT-Functions).
  • Qwen3-vl-235B: TopoAgent 68.3% vs. 64.2% (OctoTools), 59.3% (Original LLM).
  • Ablation (global avg): removing DAG Planning −1.4%; removing Atomic Fission −0.6%; DAG drop is largest in Mathematics (−2.1% on MathVista) and smallest in Chemistry (−0.5%).
  • Fission efficacy: 12.65% of math atoms and 9.88% of physics atoms are resolved exclusively via fission after initial failure; chemistry relies on fission far less due to its inherently sequential reasoning path.
  • Token efficiency: TopoAgent maintains stable prompt length across inference steps while OctoTools (linear accumulation) exhibits exponential token growth and accuracy collapse after step 4.

Limitations

  • DAG decomposition quality depends on the LLM planner's initial structuring; poor decompositions require fission cascades that add latency.
  • Adaptive Atomic Fission introduces additional inference steps per fission event; the paper does not report end-to-end wall-clock latency or per-query cost vs. baselines.
  • Chemistry gains are modest (Chem-Symbol: 62.6% vs. 62.1% without DAG), suggesting the framework's benefits are domain-dependent and less pronounced where reasoning is inherently sequential.
  • Evaluation is restricted to seven benchmarks in three domains; generalization to open-ended or non-scientific agentic tasks is untested.
  • The paper text is truncated, so details on failure-case analysis and maximum fission depth behavior are unavailable.

Relevance to Agentic AI / LLM Agents

TopoAgent directly advances the core agentic challenge of planning under uncertainty with tool use, moving beyond the dominant CoT/ReAct linear paradigm to a graph-structured execution model with principled context scoping. Its context isolation principle — each node sees only its topological ancestors — is a concrete architectural answer to the long-context hallucination problem that afflicts virtually all multi-step agents, with measured token-efficiency benefits. The Adaptive Atomic Fission mechanism is notable as a form of online replanning that avoids global recompilation, which is directly relevant to researchers building resilient tool-augmented or code-executing agents. This work sits at the intersection of structured reasoning (Graph-of-Thoughts lineage), tool-augmented agents, and self-correcting execution, and the model-agnostic improvements across proprietary and open-weight MLLMs suggest the architectural principles transfer broadly.