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 planning agent; novel non-linear orchestration for multimodal reasoning

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

TopoAgent replaces linear agent planning with a Directed Acyclic Graph (DAG) of minimal executable "atoms," enforcing strict context isolation so each reasoning node receives only its direct prerequisite outputs rather than the full history. A runtime Adaptive Atomic Fission mechanism dynamically splits failing nodes into finer sub-atoms, enabling the framework to self-correct without recompiling the global graph. Across six MLLMs and seven multimodal science benchmarks (math/physics/chemistry), TopoAgent achieves a global average accuracy of 66.3%, outperforming the strongest sequential and multi-agent baselines.

Problem

Linear agentic frameworks (CoT, ReAct, Plan-and-Solve) suffer three structural failures in multimodal scientific reasoning: (1) visual-semantic misalignment from interleaving perception and deduction in one generation step; (2) long-context hallucination from blindly concatenating all execution history, causing attention dilution; and (3) static granularity deadlock when a planned sub-task exceeds an external tool's capability and the pipeline aborts with no recovery path.

Method

TopoAgent operates in three phases executed via Algorithm 1:

  1. Visually-Grounded Atomic Decomposition. A Decomposer module \(D\) converts a query \((Q, I)\) into atoms \(A = \{A_1, \ldots, A_n\}\), each a minimal perception-first unit that anchors visual variables to symbolic states before any logical deduction.

  2. DAG Planning with Context Isolation. Atoms are organized into a DAG \(G = (A, E)\) by prerequisite dependencies. At execution time, atom \(A_i\) receives a purified local context: $\(C(A_i) = \bigoplus_{A_j \in P(A_i)} M(A_j)\)$ where \(M\) is the memory of completed atoms and \(P(A_i)\) is the direct-predecessor set. All other history is excluded.

  3. Adaptive Atomic Fission. Execution proceeds via a dynamic deque \(Q_{\text{exec}}\). If an atom \(A_i\) does not reach a STOP state within \(T_{\max}\) steps, a fission function \(F(A_i, \text{trace}) \to \{A_{i.1}, \ldots, A_{i.k}\}\) fractures it into sub-atoms prepended to \(Q_{\text{exec}}\). A hierarchical namespace mapping aggregates sub-atom results transparently: $\(M_{\text{virtual}}(A_i) = \bigoplus_k M(A_{i.k})\)$ so downstream nodes query the parent key unchanged and the global DAG requires no recompilation.

Key Contributions

  • Identification of three structural failure modes (visual-semantic misalignment, long-context hallucination, static granularity deadlock) in linear agents for scientific reasoning.
  • DAG-based topological execution with formal Context Isolation (Eq. 1) that feeds each atom only its prerequisite states.
  • Adaptive Atomic Fission (Eq. 2–3) for runtime self-evolution: hierarchical namespace preserves global graph integrity while locally degrading task granularity.
  • Model-agnostic framework validated across six MLLMs (proprietary and open-weight) and seven benchmarks spanning math, physics, and chemistry.

Results

  • Global average accuracy: TopoAgent 66.3% vs. AutoGen 62.0%, OctoTools 62.8%, LangChain 58.1%, GPT-Functions 58.4%, Original LLM 57.0%.
  • OlympiadBench (Physics): global average 41.2% vs. next-best LangChain 34.0%; GPT-5 backbone reaches 60.3% vs. 45.8% baseline.
  • Qwen3-vl backbone: 68.3% vs. OctoTools 64.2% (strongest baseline for that model).
  • Ablation — removing DAG Planning: −1.4% global average; Mathematics drops up to −2.1%.
  • Ablation — removing Atomic Fission: −0.6% global average; MathVista −0.9%, Olympiad −0.6%.
  • Fission rescue rate (Math): 51.5% of failed nodes resolved via fission vs. 18.3% via Self-Refine and 2.7% via No-Retry.
  • Token efficiency: TopoAgent maintains stable prompt token length across inference steps; OctoTools exhibits exponential context explosion with accuracy collapse after step 4.
  • Level-3 (critical thinking) tasks: TopoAgent shows minimal degradation vs. severe drops for all baselines.

Limitations

  • The Decomposer's quality bounds overall performance; an imperfect initial decomposition creates sub-optimal DAG topology that fission must partially compensate.
  • Fission adds latency per bottleneck node and introduces additional LLM calls; worst-case depth of recursive fission is not bounded in the paper.
  • Chemistry benefits least from DAG planning (−0.5% on removing it) because chemistry reasoning is inherently sequential, suggesting the topology overhead may not always be justified.
  • Evaluation is restricted to static benchmark problems; dynamic or interactive environments (tool output uncertainty, streaming data) are not addressed.
  • No analysis of failure modes when fission itself cannot resolve a node (the 5–15% of failed-case atoms shown in Figure 3 that remain unresolved).

Relevance to Harnesses / Meta-Harnesses

TopoAgent is a concrete instance of a dynamic meta-harness: rather than a fixed orchestration script, it generates and mutates its own execution graph at runtime, making it directly relevant to work on adaptive harness architectures. The Adaptive Atomic Fission mechanism is functionally a self-modifying scheduler—the harness rewrites its own task queue in response to tool failure signals, a pattern that meta-harness designers must explicitly support (or prevent). The Context Isolation mechanism formalizes a key harness design principle: sub-task inputs should be scoped to proven predecessor outputs, not the full accumulated trace, which has implications for token-budget management in long-running harness pipelines. Researchers tracking harness design patterns will find TopoAgent's DAG-plus-deque execution model and its hierarchical namespace aggregation (Eq. 3) particularly transferable to general-purpose agentic scaffolds.