Topology-Aware Prompt Optimization

• Topology-aware prompt optimization is a method that models the prompt search space as a graph to systematically improve prompt efficacy.
• It uses graph-structured modeling and search strategies like beam search and random walk to explore discrete transformations of prompts.
• Empirical outcomes indicate superior performance on NLP tasks with interpretable operator effects and scalable multi-agent coordination.

Topology-aware prompt optimization refers to a class of methods for systematically improving prompts given to LLMs, in which the topological structure of the prompt search space (i.e., the graph defined by possible transformations of prompts) is explicitly modeled and exploited. This paradigm recasts prompt engineering as discrete optimization over a structured state space, allowing for the application of combinatorial search algorithms and, in multi-agent systems, probabilistic graphical models. Empirical studies demonstrate that leveraging graph topology leads to more robust and interpretable prompt policies, superior performance on downstream tasks, and deeper understanding of the operators that drive prompt efficacy (Taneja, 23 Nov 2025, Zhang et al., 8 Oct 2025).

1. Graph-Structured Modeling of Prompt Spaces

Prompt space is formalized as a directed graph $G = (V, E)$, where $V = \mathcal{P}$ is the set of all syntactically valid prompt strings and $E \subseteq V \times V$ defines adjacency via discrete transformations. Each node $v \in V$ corresponds to a specific prompt, while edges represent primitive operators such as making a prompt concise, adding few-shot examples, reordering instructions, or increasing verbosity. Each operator $O_i$ maps a parent prompt (optionally conditioned on a training context) to a child prompt, contributing a labeled, directed edge to the graph. In concrete experimental settings, the operator set is typically small (e.g., $|\mathcal{M}| = 4$ for {make_concise, add_examples, reorder, make_verbose}), leading to a regular out-degree topology and a potentially infinite, but highly structured, prompt graph (Taneja, 23 Nov 2025).

In the multi-agent context, the topology extends to a factor graph representing the information-flow or communication DAG of the agent system. Here, nodes correspond to agent-specific prompts, and edges encode the success probabilities of agent handoffs conditioned on adjacent prompt configurations (Zhang et al., 8 Oct 2025).

2. Topology-Exploiting Search Algorithms

Classical search techniques are adapted to exploit the structure of the prompt graph. Beam search and random walk are the principal algorithms employed for single-agent prompt optimization.

• Random Walk: Iteratively applies a uniformly random operator from the available set, traversing the graph for a fixed number of steps and tracking the highest-scoring prompt.
• Beam Search: Maintains a set (beam) of the $k$ highest-scoring prompts at each depth, expanding each with all available operators and pruning to retain only the top-$k$ candidates for further expansion. The total complexity for shallow search is $O(k \cdot |\mathcal{M}| \cdot d)$, with favorable empirical performance when $k=2$, $d=2$ (Taneja, 23 Nov 2025).

In multi-agent settings, prompt optimization is formulated as maximum a posteriori (MAP) inference over a factor graph. Max-product belief propagation propagates beliefs about optimal prompt configurations across the topology, integrating both node-level and edge-level reward potentials. This enables coordinated prompt updates throughout multi-agent systems, scaling efficiently with agent count and topology complexity (Zhang et al., 8 Oct 2025).

3. Objective Functions and Evaluation Metrics

Prompt candidates are scored via heuristics based on their development set performance. Two elementary scoring functions are employed:

• String-Match Scorer: $s_\text{str}(p; x, y) = 1$ if $f_p(x) = y$, $0$ otherwise.
• Critic-LM Scorer: $s_\text{crit}(p; x, y) = 1$ if an external LLM “critic” deems $f_p(x)$ semantically correct with respect to $y$, $0$ otherwise.

The overall objective for a prompt $p$ is the mean score across a held-out dev split:

$$h(p) = Eval(p, \mathcal{D}_\text{dev}) = \frac{1}{|\mathcal{D}_\text{dev}|}\sum_{(x, y) \in \mathcal{D}_\text{dev}} s(p; x, y)$$

For multi-agent optimization, agent prompts are scored both individually and jointly, via node- and edge-potentials, with composite system-wide metrics such as joint pass@1 or exact-match accuracy (Taneja, 23 Nov 2025, Zhang et al., 8 Oct 2025).

4. Emergent Topological Patterns in Successful Optimization Trajectories

Empirical analysis of optimization trajectories reveals distinct topological patterns:

• Transformations that make prompts concise are consistently overrepresented among successful search paths; such edges (make_concise) lead the beam into high-accuracy neighborhoods.
• Verbosity-increasing operators (make_verbose) are never part of improvement paths, correlating with low-reward subgraphs in the prompt space.
• In shallow beam search ($d=2$), improved prompts predominantly arise at length 1 or 2 from the seed, indicating the search efficiently narrows on a compact and promising subgraph via early operator pruning.
• Average branching factor matches $|\mathcal{M}|$, but heuristic-driven pruning sharply reduces the effective neighborhood, yielding interpretable locality of high-performing prompts (Taneja, 23 Nov 2025).

In multi-agent scenarios, the topology of the agent interaction graph governs which prompts are selectively updated. Agents only mutate prompts when they receive blame signals from downstream neighbors, concentrating optimization on subgraphs implicated by end-to-end failures (Zhang et al., 8 Oct 2025).

5. Topology-Aware Refinement and Iterative Convergence

Iterative frameworks alternate between global MAP assignment of prompt sets and local refinement guided by topology-aware feedback. After each optimization round:

• Global execution feedback and local blame signals are collected.
• Critic LLMs re-label demonstration pools for prompt candidates.
• Mutated prompt candidates are generated in a trust region around the current best and selectively updated based on blame signal frequency and magnitude.

The system halts when the improvement over a sliding window falls below a tolerance, typically converging in a small number of iterations. This structure ensures that refinement is both efficient and topologically focused, preventing unnecessary exploration of unproductive regions in the prompt graph (Zhang et al., 8 Oct 2025).

6. Quantitative Results and Applications

Evaluation across NLP tasks (sentiment, QA, summarization, reasoning, NLI) indicates substantial dev-set improvements for topology-aware optimization compared to seed and one-shot baselines. For example, shallow beam search doubles dev accuracy on complex reasoning (0.40 → 0.80), but test improvement is more modest (0.20 → 0.50), quantifying overfitting to the dev heuristic. Key empirical findings are summarized below:

Task	Seed (dev)	Random Walk (dev)	Beam Search (dev)	Seed (test)	Beam Search (test)
NLI	1.00	1.00	1.00	1.00	1.00
QA	0.60	0.80	0.80	1.00	1.00
Reasoning	0.40	0.60	0.80	0.20	0.50
Sentiment	1.00	1.00	1.00	1.00	1.00
Summarization	0.40	0.40	0.60	0.30	0.30

Multi-agent prompt optimization via topology-aware MAPRO substantially outperforms both manually engineered and recent automatic baselines across code and QA benchmarks, with relative improvements of 1–6% depending on task and agent communication graph. Improvements are consistent across LLM backbones and MAS topologies (Taneja, 23 Nov 2025, Zhang et al., 8 Oct 2025).

7. Broader Implications and Extensions

Topology-aware prompt optimization formally connects prompt engineering for LLMs to state-space search, graphical model inference, and combinatorial optimization. Structural analysis of the prompt space enables principled operator design and interpretable refinement strategies. In distributed or multi-agent settings, information-flow topology fundamentally limits and structures the optimization landscape, enabling scalable and coordinated policy improvement via graph-based credit assignment mechanisms. While shallow search is prone to overfitting on small development sets and deeper exploration is resource-intensive, this approach yields actionable insights into which prompt engineering moves systematically promote generalization and accuracy (Taneja, 23 Nov 2025, Zhang et al., 8 Oct 2025).

---

References:

• "Prompt Optimization as a State-Space Search Problem" (Taneja, 23 Nov 2025)
• "MAPRO: Recasting Multi-Agent Prompt Optimization as Maximum a Posteriori Inference" (Zhang et al., 8 Oct 2025)