title: >- [Paper Note] FloCA: Towards Faithful and Logically Consistent Flowchart Reasoning description: >- [EMNLP 2026][VLM Reasoning][Flowchart Reasoning] FloCA decouples semantic understanding from flowchart reasoning via a topology-constrained external tool, achieving faithful and logically consistent multi-turn dialogue progression. tags: - EMNLP 2026 - VLM Reasoning - Flowchart Reasoning - Task-Oriented Dialogue - Neurosymbolic Agent date: 2026-11-03 content_hash: 2056285a8f4c2858
FloCA: Towards Faithful and Logically Consistent Flowchart Reasoning¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Jinzi-Zou/FloCA-flowchart-reasoning
Area: VLM Reasoning
Keywords: Flowchart Reasoning, Task-Oriented Dialogue, Topology-Constrained Execution, Neurosymbolic Agent, User Simulator
TL;DR¶
To tackle the severe hallucination and topology-inconsistent transitions in flowchart-oriented dialogue, FloCA decouples language understanding from graph execution via an external topology-constrained reasoning tool, achieving state-of-the-art task success and path coverage.
Background & Motivation¶
Flowchart-oriented dialogue (FOD) systems aim to guide users through complex multi-turn decision-making and operational workflows, such as technical troubleshooting, clinical triage, and legal compliance. Unlike general chit-chat or flat task-oriented dialogs, real-world flowcharts impose strict procedural dependencies and hierarchical branch conditions. An incorrect judgment or an invalid leap across disconnected nodes at any intermediate decision point inevitably routes the conversation into an erroneous branch, resulting in task failure, invalid advice, or potential physical damage.
While modern Large Language Models (LLMs) and Vision-Language Models (VLMs) demonstrate remarkable conversational proficiency, directly applying them to flowchart-oriented dialogues reveals two core structural bottlenecks. Existing paradigmsโincluding retrieval-augmented generation (RAG), graph serialization, and raw image-based VLM reasoningโeither shatter the native graph connectivity during text flattening or force models to implicitly maintain long-horizon topological validity in probabilistic generation spaces. Consequently, models frequently hallucinate topologically impossible node jumps, skip mandatory intermediate operational steps, or fall into endless repetitive clarification loops.
This paper tackles the challenge by recognizing that language models excel at local semantic interpretation and natural dialogue synthesis, but struggle with strict deterministic graph execution. Core idea: FloCA establishes a zero-shot neurosymbolic conversational agent that assigns intent understanding and edge semantic matching to an LLM while delegating deterministic path traversal to an external topology-constrained graph execution tool, complemented by self-loop initial grounding and FAQ state-freezing mechanisms.
Method¶
Overall Architecture¶
The core architecture of FloCA couples an instruction-following LLM with a faithful flowchart reasoning tool. The flowchart is formally modeled as a text-attributed directed graph \(F = (V, E, X)\), where \(V\) is the set of nodes, \(E\) denotes directed edges, and \(X\) provides the textual attributes for nodes and edges. Rather than prompting the model with the entire graph or image, the LLM is restricted to local semantic parsingโinterpreting user utterances, determining branch conditions, and rewriting node texts into natural dialogue responses. The external tool maintains the current grounded node \(\hat{v}_t\) and provides strictly constrained atomic query interfaces, mathematically preventing any transition to non-adjacent nodes.
The runtime flow consists of three distinct phases: first, an automated self-loop grounding phase analyzes the user's initial composite query against the graph top-down to locate the correct starting node without redundant turns; second, an interactive one-hop transition loop guides the user step-by-step along valid graph edges; third, whenever an out-of-flowchart knowledge question arises, the agent invokes an external FAQ retriever while freezing the node state, smoothly resuming deterministic execution once the clarification is resolved.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["User Initial Utterance & Dialogue History"] --> B["Self-loop Initial Node Grounding<br/>Recursively traverse graph from root"]
B --> C["Topology-Constrained Reasoning Tool<br/>Query node and outgoing edge attributes"]
C --> D["LLM Edge Matching & Intent Analysis<br/>Map utterance to valid branch or question"]
D -->|User asks domain knowledge query| E["FAQ Knowledge Interruption Recovery<br/>Suspend transition and fetch external answer"]
E --> C
D -->|Valid transition condition met| F["Single-Step NextHop State Update<br/>Deterministic move along valid edge"]
F --> G["Response Rewriting & Terminal Diagnosis"]
Key Designs¶
1. Topology-Constrained Flowchart Reasoning Tool: Enforcing structural connectivity via symbolic execution
To eliminate the rampant node-transition hallucinations seen when LLMs or VLMs reason over serialized text or visual layouts, FloCA encapsulates flowchart navigation within a deterministic symbolic tool. The tool exposes four atomic graph functions to the conversational backbone:
- NodeAttr(NodeID): Returns the detailed textual attribute of a specified node (e.g., diagnostic questions or instructions).
- OutEdgeAttr(NodeID): Returns the list of condition texts for all valid outgoing edges from the node.
- NextHop(NodeID, EdgeAttr): Given the current source node and a confirmed outgoing edge condition, deterministically returns the target node ID.
- TerminalCheck(NodeID): Returns a boolean flag indicating whether the node is a terminal leaf.
By restricting the LLM's decision space to the local outgoing edges of \(\hat{v}_t\), the agent avoids global distraction. Transitions are executed exclusively through NextHop, making it impossible to produce illegal jumps or skip intermediate verification steps.
2. Self-Loop Initial Node Grounding Mechanism: Eliminating redundant turns via top-down verification
In real-world applications, a user's opening utterance rarely starts from a blank slate; it often packs rich observable symptoms and completed checks (e.g., "The key is in ON position and the fuse is already blown, how to fix it?"). Starting blindly from the root node would cause frustrating, redundant interactions, while guessing an arbitrary entry node from the global graph invites severe grounding errors. FloCA introduces a self-loop grounding mechanism: starting from the root node internally, it sequentially queries NodeAttr and OutEdgeAttr, using the LLM to inspect whether the user's initial utterance already satisfies an outgoing transition condition (or confirms a completed operation). If satisfied, the tool executes NextHop internally and recursively advances down the branch until reaching the first unresolved decision node. Only then does the LLM rewrite that node's attribute into the agent's first spoken question, achieving both interaction efficiency and grounded correctness.
3. FAQ Knowledge Interruption Recovery: Decoupling domain Q&A from procedural state tracking
During procedural troubleshooting, non-expert users frequently encounter instructions requiring unfamiliar expertise (e.g., when asked "Is there a short circuit after the blown wire?", the user may ask "How do I test for a short circuit with a multimeter?"). In standard conversational architectures, such conversational detours pollute the context window, causing the model to lose track of the previous flowchart state. FloCA introduces a dedicated state-suspension mechanism: when the LLM detects that an utterance is an informational query rather than a branch decision, it suspends graph traversal, queries an external FAQ retriever (FAQRetriever), and answers the user's question while strictly preserving the grounded node \(\hat{v}_t\). Flowchart reasoning resumes from the identical node once the user supplies the resulting diagnostic outcome.
A Worked Example¶
Consider an automotive electrical troubleshooting workflow: root node โ is "Open circuit on run?", with a "Yes" branch leading to node โก "Fuse wire blown?", whose "Yes" branch connects to node โข "Short after blown wire?", which branches on "No" to terminal node โฃ "Replace with new fuse wire rated same or slightly lower."
1. Initial Grounding: The user begins with: "A fuse wire has blown in my car. How can I fix it?". FloCA triggers internal self-loop from node โ . The LLM infers that a blown fuse entails an open-circuit fault, matching the "Yes" edge to call NextHop(โ , Yes) -> node โก. At node โก, the user explicitly reported the blown wire, matching "Yes" to call NextHop(โก, Yes) -> node โข. At node โข, short-circuit status is unknown; the self-loop halts. FloCA rewrites node โข into the first spoken query: "Please check whether there is a short circuit."
2. FAQ Interruption: The user asks: "How do I check if my car has a short circuit?". The LLM identifies a domain-knowledge inquiry, maintains node โข as active, invokes FAQRetriever, and replies: "Using a multimeter set to resistance measurement...". The user tests and responds: "There is no short circuit in the car."
3. Deterministic Progression: The LLM matches this reply to the "No" condition of node โข, calling NextHop(โข, No) to reach node โฃ. TerminalCheck(โฃ) flags completion, and the agent outputs the final prescription: "Use a new fuse wire with the same or slightly lower rating."
Key Experimental Results¶
Main Results¶
Evaluations are conducted on the benchmark FLODIAL dataset (12 diagnostic flowcharts with 2,738 multi-turn dialogues). Beyond terminal grounding accuracy (TNGA), the evaluation framework introduces Initial Node Grounding Accuracy (INGA), ordered Path Coverage Accuracy (PCA), Timeout Rate (TR), and Node Stay Redundancy (NSR).
| Method | Backbone / Type | INGA Overall (โ) | TNGA (โ) | PCA (โ) | TR (โ) | NSR (โ) |
|---|---|---|---|---|---|---|
| SA-FloNet (Raghu et al., 2022) | Fine-tuned Seq2Seq | 67.75 | 1.65 | 0.20 | 0.00 | 23.24 |
| Graph Serialization | GPT-4o | 52.87 | 78.26 | 73.49 | 0.00 | 0.85 |
| Graph Serialization | GPT-5 | 81.38 | 87.16 | 79.29 | 0.62 | 1.15 |
| Graph Serialization | Claude Opus 4.1 | 60.85 | 83.43 | 80.12 | 0.00 | 0.38 |
| RAG + Graph Serialization | GPT-4o | 51.29 | 80.53 | 75.98 | 0.00 | 0.72 |
| RAG + Graph Serialization | Claude Opus 4.1 | 60.62 | 87.16 | 6.62 | 0.00 | 42.25 |
| Visual Language Models (VLM) | GPT-4o | 54.85 | 33.12 | 12.83 | 0.62 | 2.60 |
| Visual Language Models (VLM) | GPT-5 | 77.85 | 50.10 | 29.39 | 9.10 | 8.45 |
| Visual Language Models (VLM) | Claude Opus 4.1 | 49.20 | 16.35 | 3.72 | 0.20 | 1.80 |
| FloCA (Ours) | Llama-3.3-70B | 61.77 | 83.85 | 83.01 | 0.00 | 0.86 |
| FloCA (Ours) | DeepSeek-R1 | 60.44 | 90.26 | 89.84 | 0.00 | 0.45 |
| FloCA (Ours) | Qwen3 | 59.90 | 91.71 | 90.88 | 0.00 | 0.64 |
| FloCA (Ours) | GPT-4o | 62.11 | 91.71 | 90.88 | 0.00 | 0.15 |
| FloCA (Ours) | Claude Opus 4.1 | 61.02 | 91.30 | 90.46 | 0.00 | 0.23 |
| FloCA (Ours) | GPT-5 | 61.45 | 91.92 | 90.88 | 0.00 | 0.60 |
Ablation Study¶
Ablations on FLODIAL using GPT-5 examine three critical design choices: (1) removing the self-loop mechanism (predicting initial nodes globally from serialized text); (2) removing edge matching (predicting transitions directly over serialized 1-hop subgraphs); (3) removing explicit state preservation during FAQ interruptions (re-inferring active nodes from dialogue history).
| Config | TNGA (%) | PCA (%) | Note |
|---|---|---|---|
| FloCA (Full model) | 91.92 | 90.88 | Complete neurosymbolic execution with self-loop grounding |
| w/o self-loop mechanism | 84.67 | 83.02 | Global start guessing matches similar nodes, causing largest drop (-7.86% PCA) |
| w/o FAQ interruption recovery | 90.68 | 89.64 | Contextual drift after Q&A causes state misalignment (-1.24% PCA) |
| w/o edge matching | 91.30 | 90.26 | Direct subgraph textual prediction loses discrete edge gating (-0.62% PCA) |
Key Findings¶
- Symbolic execution eliminates logical divergence: Direct image-based VLMs collapse on multi-turn path coverage (e.g., GPT-5 VLM achieves only 29.39% PCA, suffering repetitive loops across 8 consecutive turns). In contrast, FloCA achieves over 90% PCA across diverse LLM backbones, proving that structured graph execution is essential for faithful flowchart traversal.
- Self-loop initial grounding is vital: Discarding the self-loop mechanism causes the largest degradation (-7.86% PCA). Flowcharts often contain structurally distinct nodes with nearly identical phrasing (e.g., checking voltage in different sub-circuits); traversing top-down from the root acts as a topological regularizer.
- Decoupled execution ensures zero timeout: While RAG baselines suffer severe interaction bloat due to noisy context retrieval (Claude Opus 4.1 NSR spikes to 42.25), FloCA achieves zero timeout rate (\(\text{TR} = 0.00\)) and negligible stay redundancy across all models.
Highlights & Insights¶
- Neurosymbolic task division: Constraining probabilistic foundation models strictly to local semantic alignment while leaving global trajectory consistency to deterministic graph execution yields optimal faithfulness.
- Rigorous multi-turn FOD evaluation suite: The introduction of a simulator conditioned on user-known facts paired with INGA, TNGA, PCA, NSR, and TR establishes a benchmark that evaluates topological fidelity rather than superficial lexical similarity.
- Reusable agent state machine pattern: The combination of self-loop fast-forwarding and suspended state recovery offers a general blueprint for troubleshooting, customer support, and medical diagnosis agents governed by formal protocols.
Limitations & Future Work¶
- Lack of backtracking support: FloCA currently assumes strictly forward progression along a directed acyclic graph. When users correct previous statements (e.g., "Wait, I checked again and the fuse is fine"), the agent lacks mechanisms to backtrack along previous branches.
- Single-graph topology assumptions: The framework currently operates over unified connected text-attributed graphs, leaving irregular visual elements (swimlanes, off-page connectors, decoupled subgraphs) to future modeling.
- Dataset scale: Evaluations remain constrained to public benchmarks like FLODIAL and PFDial; evaluating large-scale industrial and multi-page diagrams remains an open goal.
Related Work & Insights¶
- vs SA-FloNet (Raghu et al., 2022): SA-FloNet uses dense retrieval over path representations without structural hard constraints, resulting in a severe drop to 1.65% TNGA on multi-turn sessions; FloCA enforces topology through symbolic execution, achieving >91% TNGA.
- vs End-to-End VLMs (FlowVQA, GPT-5 Vision): Direct visual reasoning frequently misinterprets visual connection tokens and gets trapped in redundant clarification loops; FloCA translates diagrams to graph tools, ensuring faithful execution.
Rating¶
- Novelty: โญโญโญโญโ Formulates flowchart dialogue as topology-grounded execution and introduces effective self-loop grounding.
- Experimental Thoroughness: โญโญโญโญโญ Rigorous evaluation across multiple frontier LLMs, user simulation diversity metrics, and detailed failure taxonomy.
- Writing Quality: โญโญโญโญโญ Well-structured narrative with crisp neurosymbolic separation.
- Value: โญโญโญโญโญ Practical and robust paradigm for industrial troubleshooting and SOP automation.