A Pre-Reasoning Governance Layer for Topology Selection in LLM Reasoning
LLMs can reason using many structures: chains, trees, graphs, and more. But how do you pick the right one?
Current practice relies on researcher intuition, prompt heuristics, or hardcoded defaults. No existing framework connects the purpose of a reasoning task to the selection of a reasoning structure.
We identify six distinct levels at which intent operates in reasoning. The first five are addressed by prior work. The sixth, topology governance, is the gap this paper fills.
IoT adds a pre-reasoning checkpoint that governs which topology to deploy, then monitors for intent drift during execution.
Every reasoning task gets a three-part checkpoint before topology selection:
IoT = (Purpose, Anti-Purpose, Success Signal)
| Moment | Question | What It Prevents |
|---|---|---|
| Purpose | WHY are we reasoning? | Aimless computation |
| Anti-Purpose | What must we AVOID? | Technically valid but useless output |
| Success Signal | HOW will we know? | Reasoning that never terminates |
Same algorithm, different intents, different topology selections:
| Case | Task | IoT Recommends | Why |
|---|---|---|---|
| 1 | Mathematical proof | 📝 CoT | Single valid path, each step depends on previous |
| 2 | UI design challenge | 🌳 ToT | Must explore 3+ alternatives before committing |
| 3 | Hospital readmission analysis | 🕸️ GoT | Feedback loops between staffing, planning, education |
from iot import IntentOfThought, TopologySelector
# Define your intent
iot = IntentOfThought(
purpose="Map causal relationships between hospital readmission factors",
anti_purpose="Treating factors as independent when they interact",
success_signal="Relationship map with bidirectional dependencies and feedback loops"
)
# Select topology
selector = TopologySelector()
result = selector.select(iot, context="systems analysis")
print(result)
# => Recommended: GoT
# The Purpose involves interconnected factors with feedback
# loops, requiring a graph structure for refinement and merging.intent-of-thought/
├── README.md # This file
├── LICENSE # Apache 2.0
├── paper/ # Paper 1 (original, standalone)
│ ├── intent_of_thought.md
│ ├── intent_of_thought.tex
│ ├── references.bib
│ └── figure*.svg / figure*.pdf
├── papers/
│ ├── Paper-IoT-Intent-of-Thought/ # Paper 1 drafts, outlines, NOOL, LOON
│ ├── Paper-IoT-Lifecycle/ # Paper 2 drafts, outlines, NOOL, LOON
│ └── Paper-IoT-Unified/ # Unified paper v12/v13 (active)
├── iot/ # Core IoT framework
│ ├── specification.py # IoT triple: Purpose, Anti-Purpose, Success Signal
│ ├── selector.py # Algorithm 1: Topology Selection
│ └── drift.py # Algorithm 2: Intent Drift Detection
├── lifecycle/ # Lifecycle extensions (from Paper 2)
│ ├── capture.py # Capture Spectrum (L0-L4)
│ ├── judge.py # Retrospective Judgement (3 failure modes)
│ ├── respond.py # Governance-Proportional Response
│ ├── runner.py # Lifecycle Runner
│ └── learning.py # Learning Loop feedback
├── benchmark/ # Topology selection benchmark tasks
├── experiments/ # Full experiment suite (5 runners, scoring, analysis)
└── examples/
├── case1_sequential.py # Mathematical proof → CoT
├── case2_parallel.py # UI design → ToT
├── case3_interconnected.py # Hospital analysis → GoT
├── case1_false_capture.py # Clinical triage (False Capture)
├── case2_false_selection.py # Legal precedent (False Selection)
└── case3_false_execution.py # Aviation safety (False Execution)
@article{mohamedkani2026intent,
title={Intent of Thought: A Pre-Reasoning Governance Layer for
Topology Selection in LLM Reasoning},
author={Mohamed Kani, Naveen Riaz},
journal={arXiv preprint arXiv:2503.XXXXX},
year={2026},
url={https://arxiv.org/abs/2503.XXXXX}
}Naveen Riaz Mohamed Kani ORCID: 0009-0003-9173-2425
Apache 2.0. See LICENSE for details.