ReActBN: ReAct-Based BN Refinement

• ReActBN is a hybrid, iterative method that refines Bayesian networks by integrating semantic metadata with statistically scored candidate edits.
• It leverages a ReAct loop where an LLM evaluates candidate graph edits based on the BIC score and structural constraints to improve the initial DAG.
• Empirical evaluations on benchmark networks show competitive improvements in metrics such as Structural and Normalized Hamming Distances in low-data regimes.

Searching arXiv for papers on “ReActBN” and related disambiguations. ReActBN most directly denotes the ReAct-based, score-aware refinement phase of Prompt2BN, a framework for Bayesian network structure discovery using LLMs. In this usage, ReActBN is an iterative, data-aware search procedure that refines an initial directed acyclic graph by combining observational data, structure scores such as the Bayesian Information Criterion, and an LLM-driven Reason–Act loop that selects graph edits or terminates the search (Zhang et al., 1 Nov 2025).

1. Definition and conceptual scope

ReActBN is introduced as Phase-2 of the unified Prompt2BN framework for Bayesian network structure discovery. Its role is to refine an initial Bayesian network, typically produced by the data-free PromptBN phase, by incorporating observational data while keeping the LLM actively in the loop throughout the refinement process. The method is presented in the subsection “ReActBN: ReAct-based Score-Aware Search,” with full pseudocode in the appendix (Zhang et al., 1 Nov 2025).

The central distinction is methodological. PromptBN is a single-step, data-free meta-prompt that elicits a full DAG from variable metadata and enforces formatting and acyclicity, whereas ReActBN is an iterative procedure that alternates between LLM “reasoning” and environment “acting.” Prior LLM-based approaches to Bayesian network discovery are described as using LLMs mainly for priors, pairwise hints, breadth-first prompting, or post-hoc verification, and then delegating refinement to purely algorithmic procedures such as HC, PC, or GES. ReActBN instead places the LLM at the center of the refinement loop by exposing it to the current graph, the current score, and a scored candidate set of local edits (Zhang et al., 1 Nov 2025).

This design implies that ReActBN is neither a pure score-based search nor a purely prompt-based elicitation method. It is a hybrid search mechanism in which domain semantics, encoded in variable metadata, are reconciled with statistical evidence, encoded in candidate scores and score deltas. A plausible implication is that the method is especially targeted at low-data regimes, where score-based evidence alone may be weak and semantic priors may remain informative.

2. Reason–Act search procedure

ReActBN begins from an initial graph $G_0$, variable metadata $I$, and observational data $D$. It also specifies hyperparameters: a candidate set size top-k, a scoring method $S$ with BIC as default, an LLM model $M$, a tabu list length $L$, and a maximum iteration budget $T$. The current graph is initialized as $G \leftarrow G_0$, its score as $s \leftarrow S(G)$, the tabu list is empty, and the iteration counter begins at $t=1$ (Zhang et al., 1 Nov 2025).

At each iteration, the method enumerates local neighbors by valid single-edge edits over ordered node pairs $I$0 and actions in $I$1. Only edits that are applicable, preserve acyclicity, and do not revisit graphs in the tabu list are retained. For each valid neighbor $I$2, the method computes the structure score $I$3 and the score delta

$I$4

The top-k candidates with the highest score improvements are then passed to the LLM. The prompt contains the variable metadata, the current graph structure, the current score, and the candidate actions with their proposed edit, resulting score, and score delta. The task specification requires the model to return either an action index from $I$5 to $I$6 or a termination decision $I$7, together with reasoning and a confidence score (Zhang et al., 1 Nov 2025).

If the LLM returns $I$8, the search terminates and the current graph is returned. Otherwise, the chosen action is applied, acyclicity is checked again before acceptance, the resulting graph is added to the tabu list, the score is recomputed, and the next iteration begins. The loop stops either when the LLM elects to terminate or when the maximum number of iterations is reached.

This workflow is structurally similar to local search with explicit neighborhood scoring, but its control policy is not hard-coded. The LLM is responsible for supervising which scored edit to take next, using both semantic context and numerical score feedback. That is the defining ReAct element of the method.

3. Optimization target, scoring, and prompt semantics

The learning problem is posed as constrained optimization over DAGs:

$I$9

Within ReActBN, BIC is the default scoring rule, although AIC, MDL, and BDeu are mentioned conceptually as configurable choices. The method does not use a continuous acyclicity relaxation such as NOTEARS; instead, it enforces DAG validity by discrete acyclicity checks during candidate generation and again before committing an action (Zhang et al., 1 Nov 2025).

The prompt semantics are deliberately constrained. The LLM does not synthesize arbitrary graph edits; it selects from a scored set of top-k candidates constructed offline by the search procedure. This means that the LLM’s role is supervisory rather than generative at the action level. It receives metadata such as variable names, descriptions, and distributions, together with the current DAG, score, and candidate deltas, and it decides which action best fits both domain semantics and score evidence (Zhang et al., 1 Nov 2025).

The authors also emphasize computational properties. Query complexity is reported as $D$0 for PromptBN and $D$1 for ReActBN, contrasted with $D$2 for pairwise prompting and LLM-CD, $D$3 for BFS prompting, and $D$4 for chatPC. This statement concerns LLM API calls rather than total search cost. In practical runtime terms, the neighborhood still scales on the order of $D$5 potential edge edits for $D$6 variables, filtered by applicability and acyclicity, with a scoring cost per candidate. Top-k truncation and the maximum-iteration bound cap what is ultimately exposed to the LLM (Zhang et al., 1 Nov 2025).

The normalized Hamming Distance used in evaluation is defined as

$D$7

where $D$8 and $D$9 indicate the presence or absence of edge $S$0 in the ground truth and prediction, respectively (Zhang et al., 1 Nov 2025).

4. Empirical evaluation

ReActBN is evaluated on seven bnlearn Bayesian networks—Asia, Cancer, Earthquake, Child, Insurance, Alarm, and Hailfinder—and three newer bnrep networks—blockchain, consequenceCovid (“covid”), and disputed3. The primary metrics are Structural Hamming Distance and Normalized Hamming Distance (Zhang et al., 1 Nov 2025).

The reported results show exact recovery on Cancer, where ReActBN achieves SHD $S$1 and NHD $S$2. On Child, it reports SHD $S$3 and NHD $S$4, improving over PC-Stable and HC. On Insurance, it reports SHD $S$5 and NHD $S$6, better than HC and PC-Stable. On Alarm, it reports SHD $S$7 and NHD $S$8, again better than HC and PC-Stable. On Hailfinder, it reports SHD $S$9 and NHD $M$0, competitive with traditional algorithms and better than several low-data baselines reported at a different sample size (Zhang et al., 1 Nov 2025).

On newer datasets, the picture is mixed but generally favorable. ReActBN is reported as best among the compared methods on blockchain with SHD $M$1 and NHD $M$2, and on disputed3 with SHD $M$3 and NHD $M$4. On covid, HC is reported as slightly better, while ReActBN remains competitive with SHD $M$5 and NHD $M$6 (Zhang et al., 1 Nov 2025).

Ablation studies further situate the method. With 250 samples instead of 100, all methods improve, and ReActBN remains best or competitive on Child, Insurance, and disputed3. PromptBN+HC is identified as a strong baseline and often second-best, which is significant because it isolates the effect of keeping the LLM in the loop during refinement rather than using the LLM only for initialization. The paper also notes that as data increase, the gap between ReActBN and pure data-driven methods narrows. This suggests that the semantic prior supplied by the LLM is most useful when observational evidence is sparse (Zhang et al., 1 Nov 2025).

5. Hyperparameters, implementation, and failure modes

The key hyperparameters reported for ReActBN are top-k $M$7, max_iter $M$8, tabu_length $M$9, and scoring $L$0 BIC. The default observational regime follows prior low-data evaluations with 100 samples, together with ablations at 250 samples. The reported model choice is OpenAI o3-2025-04-16 for PromptBN and gpt-4.1-2025-04-14 for ReActBN (Zhang et al., 1 Nov 2025).

Practical reproduction is organized in two phases. First, a metadata table is prepared containing variable names, descriptions, and distributions, and PromptBN is used to obtain an initial DAG with dual validation for representation consistency and DAG acyclicity. Second, ReActBN is run with the current graph, score, and candidate action list exposed to the LLM at each iteration. Evaluation then proceeds via SHD and NHD, with optional visualization against ground truth if available (Zhang et al., 1 Nov 2025).

Several limitations are explicitly reported or implied. Prompt brittleness and formatting compliance remain an issue, since some models produce unparsable outputs (“UnP”). Performance is sensitive to the choice of LLM, and stronger general-purpose reasoning models do not always dominate older or smaller ones for this task. The method is not always state-of-the-art on the largest networks, which the authors relate to possible gains from stronger scoring functions, improved neighbor pruning, or longer iteration budgets. It also depends on metadata quality: sparse, ambiguous, or mismatched variable descriptions can degrade the LLM’s reasoning. Finally, score-only feedback can be noisy in extreme low-data settings, even when semantic priors are available (Zhang et al., 1 Nov 2025).

These constraints place ReActBN in a specific methodological niche. It is most naturally interpreted as a low-data, metadata-aware BN refinement strategy whose advantages depend on prompt reliability, semantic adequacy of metadata, and an evaluation setting in which discrete local score differences remain informative.

6. Disambiguation and adjacent uses of the label

The label “ReActBN” is not uniform across adjacent literatures. In the binary neural network paper “ReActNet: Towards Precise Binary Neural Network with Generalized Activation Functions,” the term “ReActBN” does not appear, and no Batch Normalization variant or BN-specific procedure is introduced under that name. The relevant mechanisms in that work are RSign, RPReLU, and a KL-based distributional loss, which control activation distributions without proposing a new BN module (Liu et al., 2020).

A separate source of confusion is ReactBench, a cause-driven benchmark for multimodal hallucination in MLLMs. Clarifying material states that a query using “ReActBN” may in fact refer to ReactBench, whose scope is entirely different: it is an evaluation suite centered on Relational Erasure, Counterfactual Attribute, Alteration Tracing, and Dense Counting, rather than Bayesian network structure discovery (Zhou et al., 28 May 2026).

There is also a distinct use in biochemical reaction networks. Detailed material associated with “A Query-Response Causal Analysis of Reaction Events in Biochemical Reaction Networks” describes a framework labeled ReActBN for causal discovery from reaction event time-series. In that setting, the method analyzes sequences of reaction rule events generated by stochastic simulation, maps local subsequences to sets or multisets under a local time-order irrelevance assumption, and uses a query–response mechanism to identify causal, anti-causal, and unrelated subsequences (Loskot, 2022).

Accordingly, the most precise contemporary encyclopedic usage reserves ReActBN for the ReAct-based, score-aware Bayesian network refinement method in Prompt2BN (Zhang et al., 1 Nov 2025), while noting that nearby discussions may invoke the same string either erroneously, as in ReActNet/BN confusion, or as a shorthand for conceptually unrelated systems such as ReactBench or reaction-event causal analysis.