SymFSM API

From Text Generation to Computable Reasoning

SymFSM is a cognitive reasoning architecture that adds a formal reasoning layer on top of Large Language Models.

Instead of generating answers directly from prompts, SymFSM builds and analyzes a task structure before generation.

The goal is not to make models produce more text.

The goal is to make reasoning itself a computable object.

---

How SymFSM Works

Traditional LLM workflow:

Prompt
   ↓
LLM
   ↓
Answer

SymFSM workflow:

Prompt
   ↓
Cognitive Map
   ↓
Reachability Analysis
   ↓
Structural Repair
   ↓
LLM Generation
   ↓
Verification
   ↓
Answer

🧪 Experiment: SymFSM vs Standard LLM on GPQA Main

We conducted a controlled experiment to evaluate the impact of SymFSM on reasoning performance.

Setup

• Same model used in both modes
• Benchmark: GPQA Main – 448 challenging multiple-choice questions from biology, physics, and chemistry
• Two modes:
  1. Standard LLM generation – direct answer generation without additional reasoning control
  2. LLM + SymFSM – generation with SymFSM reasoning control enabled (cognitive maps, reachability checks, and repair mechanics)

---

Results

Mode	Correct	Accuracy	Improvement
Standard LLM	282 / 448	62.95%	—
LLM + SymFSM v1.0 (Verification of reasoning)	314 / 448	70.09%	+7.14 p.p.
LLM + SymFSM v3.0 (Invention of solutions)	333 / 448	74.33%	+11.38 p.p.
LLM + SymFSM v4.0 (Dynamic graph rewriting)	336 / 448	75.00%	+12.05 p.p.
LLM + SymFSM v5.0 (Cognitive computation management)	340 / 448	75.89%	+12.94 p.p.
LLM + SymFSM v6.0 (Competing thinking programs & self-evolution)	349 / 448	77.90%	+14.95 p.p.

---

Key Takeaway

Applying SymFSM increased accuracy by 14.95 percentage points on the same test dataset.

This demonstrates that formal reasoning control – building a task graph, checking reachability, repairing logical gaps before generation, and evolving the cognitive strategy itself – can significantly improve LLM performance on complex reasoning tasks, without any fine‑tuning or prompt engineering.

---

Evolution of Improvements

Version	Key Innovation	Accuracy Gain
v1.0	Verification of reasoning	+7.14 p.p.
v3.0	Invention of solutions	+11.38 p.p.
v4.0	Dynamic graph rewriting	+12.05 p.p.
v5.0	Cognitive computation management	+12.94 p.p.
v6.0	Competing thinking programs & self-evolution	+14.95 p.p.

Total improvement from v1.0 to v6.0: +7.81 p.p.

---

What v6.0 Adds

v6.0 introduces competing thinking programs – the system no longer solves every problem with a single predefined approach. Instead, it simultaneously runs multiple independent cognitive strategies, compares their trajectories, selects the most effective one, and uses it to build the final solution.

After each request, the system saves successful cognitive paths, increases the utility of used cogs, and memorizes effective thinking recipes – accumulating not just knowledge, but its own reasoning experience.

Before text generation, SymFSM attempts to determine:

• what concepts exist in the task;
• how they are connected;
• whether the goal is reachable;
• where reasoning gaps exist;
• which solution trajectories are valid.

---

Key Features

Cognitive Maps

Tasks are transformed into structured cognitive graphs.

Graph nodes may represent:

• concepts;
• goals;
• constraints;
• mechanisms;
• hypotheses;
• dependencies.

---

Reachability Analysis

SymFSM evaluates whether a target conclusion can be reached from known information.

Instead of guessing missing steps, the system detects reasoning gaps.

---

Structural Repair

When the reasoning graph is incomplete, SymFSM can:

• introduce missing concepts;
• create sub-maps;
• reorganize dependencies;
• search alternative reasoning paths.

---

Finite State Reasoning

The reasoning process is controlled by finite-state machines.

This prevents arbitrary jumps between reasoning stages and enables formal analysis of solution trajectories.

---

Recursive Reasoning

During answer generation, the model may request:

• expansion of a local cognitive map;
• creation of a new sub-map;
• reachability verification;
• structural repair.

Reasoning can therefore evolve dynamically while the answer is being generated.

---

Experience Accumulation

SymFSM stores:

• successful reasoning trajectories;
• cognitive patterns;
• repair strategies;
• structural solution templates.

The system learns solution strategies rather than memorizing answers.

---

Typical Use Cases

• Product strategy
• Business architecture
• Research tasks
• Engineering design
• Complex planning
• Security analysis
• Malware analysis
• Knowledge-intensive decision support

---

API

Submit Task

Endpoint:

POST /submit

Request:

{
  "prompt": "Generate 10 business ideas for the artificial intelligence industry"
}

Response:

{
  "id": "task_id",
  "status": "queued"
}

---

Get Result

Endpoint:

GET /result?id=task_id

Response:

{
  "id": "task_id",
  "status": "done",
  "result": "..."
}

Possible statuses:

queued
running
done
error

---

Server Address

Current API endpoint:

http://ip:8088

Example:

http://ip:8088/submit

Future versions may use a domain name instead of a direct IP address.

---

C# Example

See:

examples/Program.cs

---

Python Example

See:

examples/example.py

---

Website

https://principium.pro/symfsm/

---

License

Research Prototype

Copyright © SymFSM