Brain-Inspired Multi-Agent Architecture
for "Spike of Insight"

Retrieval is solved. Adaptation is missing.
geDIG implements the autonomic nervous system of "when to learn and when to forget" into RAG using a single gauge:
$\mathcal{F} = \Delta EPC_{norm} - \lambda(\Delta H_{norm} + \gamma \Delta SP_{rel})$ Plain English: A single scalar that decides: "Does this new fact deserve a permanent place in the graph?"
$\Delta EPC_{norm} \approx$ "structure cost", $\Delta H_{norm} \approx$ "extension gain", $\Delta SP_{rel} \approx$ "shortcut gain" Benchmark Result: On 25x25 Maze + RAG, geDIG reduced redundant exploration by 40% while keeping FMR < 2%.

🧠 Try Live Demo (HF Spaces) Interactive Graph Demo View Paper (Draft)
Amoeba geDIG (live) --
$F$ = $\Delta EPC_{norm}$ − $\lambda(\Delta H_{norm} + \gamma\Delta SP_{rel})$
F-- ΔEPC-- ΔH-- ΔSP--
High $F$ → AG (explore). Low $F$ → DG (commit).
Tip: the bright dots are nodes; thick veins are edges.
Tune parameters
--
--
--
--
--
This is a visualization: $\Delta H$ and $\Delta SP$ are local proxies, but the equation is real.
$\Delta EPC$ (Structure Cost) $\Delta H$ (Entropy) $\Delta SP$ (Shortcuts) FEP-MDL Bridge

Visualizing "The Pulse"

Real-time simulation of the gauge $\mathcal{F}$ as the graph grows.
Click anywhere in the graph to inject a new query node and see if it triggers an Insight (DG) or Ambiguity (AG).
AG (Ambiguity Gauge): 0-hop error (High Cost).
DG (Discovery Gauge): Multi-hop shortcut (Insight Found).

Observation Guide:
  • Click near a cluster: $\Delta EPC$ is low, but $\Delta IG$ is also low. $\mathcal{F}$ stays high (Reject).
  • Click between clusters: If a shortcut is found, $\Delta SP$ spikes (negative cost), driving $\mathcal{F} < \theta$ (Insight!).
IDLE
Click to Add Node
Base Node
Query
1-hop (Ego)
2-hop+ (Insight)

Gauge Telemetry

Structural Cost ($\Delta EPC$) --
wiring cost (normalized)
Information Gain ($\Delta IG$) --
Entropy ($\Delta H$) --
Shortcuts ($\Delta SP$) --
Total Gauge ($\mathcal{F}$) --
Threshold: < 0.30 IDLE
Manual Override

Conceptual Architecture

Most RAG systems only optimize "what to retrieve". geDIG optimizes "when to integrate".
This adds a metacognitive layer to the AI's memory system.

Static RAG
Always Retrieve -> Always Append
Infinite Growth / Pollution
vs geDIG
LLM $\to$ Retriever $\to$ Gauge $\mathcal{F}$ $\to$ [Accept/Reject]
01. THE PROBLEM

Blind Accumulation

Static RAG systems lack a norm for rejection. They accumulate noise, redundancies, and contradictions, leading to performance degradation over time (The "Context Window" trap).

02. THE SOLUTION

One-Gauge Control ($\mathcal{F}$)

We unified Structural Cost (Graph Edit Distance) and Information Gain (Entropy + Shortcuts) into a single scalar.

  • $\Delta EPC$: Cost of wiring. Penalizes complexity.
  • $\Delta H$: Entropy reduction. Rewards order.
  • $\Delta SP$: Path shortening. Rewards insight (shortcuts).
$\mathcal{F} < \theta$
Accept
Reject
03. WHY BOTH

Why Probability + Graph, Together

Similarity alone stops too early; structure alone cannot tell if uncertainty remains. geDIG evaluates both to decide when to integrate.

Quick Comparison
Method Probability Graph
BM25 / Contriever Yes No
GraphRAG No Yes
geDIG Yes Yes
Rationale in the paper (expand) Details

geDIG measures structural cost (graph edit distance) and probabilistic gain (entropy reduction + shortcut effect) in one gauge. This blocks false positives that look structured but stay uncertain, and avoids redundant retrievals that look probable but add no structure.

See the full explanation (PDF, §Why Both)

FEP-MDL Bridge

0-hop (AG) Ambiguity / Error → FEP
Multi-hop (DG) Compression / Insight → MDL
"Operational Correspondence"
04. THEORETICAL BACKBONE

Bridging FEP & MDL

geDIG provides an operational correspondence between the Free Energy Principle (minimizing surprise) and Minimum Description Length (maximizing compression).

0-hop detects immediate prediction error (FEP), while Multi-hop validates global compression (MDL).

*For readers familiar with FEP/MDL: This is an operational analogy, not a formal proof of equivalence.

Downloads, Code & Reproducibility

Draft papers, code entrypoints, and maze reproduction commands.

Interactive

Live Demo with Knowledge Graph Visualization

See geDIG in action on HotPotQA questions. Watch how the knowledge graph evolves before and after retrieval, and understand why geDIG accepts or rejects new information.

  • ✓ Before/After Knowledge Graph comparison
  • ✓ Real-time F-score calculation
  • ✓ AG/DG gate status visualization
🧠 Try on Hugging Face
For Researchers

Research Paper

"geDIG: Gauge what Knowledge Graph needs"

v6.0 Draft
JA Version (v6) EN Version (v6) JA Overview Slide (Landing PDF)
For Engineers

Source Code

miyauchikazuyoshi/InsightSpike-AI

Brain-Inspired Multi-Agent Architecture for “Spike of Insight”

View Repository Browse Experiments & Logs

CLI / Quick Start

Quick start: python examples/public_quick_start.py

CLI: python -m insightspike.cli.spike --help

Smoke tests: make codex-smoke

Reproduce (Maze 25×25, 500 steps)

Phase-1 PoC: Maze + RAG under equal resources.

L3 batch (60 seeds):

python scripts/run_maze_batch_and_update.py --mode l3 --seeds 60 --workers 4 --update-tex

Eval batch (60 seeds):

python scripts/run_maze_batch_and_update.py --mode eval --seeds 60 --workers 4 --update-tex

Aggregates land in docs/paper/data/; the 25×25 table updates automatically.

Call for arXiv Endorsement / Reviewers

We are seeking an arXiv endorsement for cs.AI or cs.LG to publish our v6 paper.

We’re also looking for collaborators on:

  • Information thermodynamics / FEP:
    Help formalize the free-energy mapping and check for missing assumptions.
  • Graph RAG / Multi-hop:
    Stress-test geDIG on your existing GraphRAG benchmarks.
  • Phase-2 (Offline Rewiring):
    Co-design offline "sleep phase" rewiring experiments.

How to engage: open an Issue with “Review” label, PR small fixes, or DM on X (Twitter): @kazuyoshim5436.

See also: geDIG spec, Phase‑1 (maze & RAG), Trace a spike.

Citation (BibTeX)

@article{miyauchi2025gedig, title={geDIG: Gauge what Knowledge Graph needs}, author={Miyauchi, Kazuyoshi}, year={2025}, journal={GitHub repository}, url={https://github.com/miyauchikazuyoshi/InsightSpike-AI} }