File size: 3,954 Bytes
4af471f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 | ---
language: en
license: apache-2.0
tags:
- reasoning
- agent
- world-model
- long-horizon
- code
- proof-of-concept
datasets:
- Qwen/AgentWorldBench
- datacurve/deep-swe
metrics:
- pass@1
model_name: OMEGA-Reasoner
---
# OMEGA-Reasoner (POC)
β οΈ **This is a proof-of-concept** β a small-scale demonstration of a
self-contained LLM-driven agent framework. It is **not** a frontier
model and is **not** intended to compete with production models on
public leaderboards such as DeepSWE.
## Honest positioning
| Property | This repo | DeepSeek-V4.1-Flash (Top-1 DeepSWE) |
|---|---|---|
| Parameters | β 100K (configurable) | 35B-class |
| Training data | none (synthetic / hand-coded) | trillions of tokens |
| Pass@1 on DeepSWE | **not measured** | 74.2 |
| Status | research POC | production model |
We do not submit to the official DeepSWE leaderboard. The integration
code (`deep_swe.py`) is provided for educational and self-evaluation
purposes only.
## What this model is
`OMEGA-Reasoner` is a self-contained **language world model** + **autonomous
agent** framework. It implements:
- HMoE-D reasoner: hierarchical mixture-of-experts with dynamic routing
- AGI_AgentWorld: full cognitive agent with episodic / semantic / procedural
memory, ReAct loop, Tree-of-Thoughts planning, Reflexion-style
self-correction, and constitutional guardrails
- AgentWorldBench adapter: 5-dim scoring (format, factuality, consistency,
realism, quality)
- DeepSWE adapter: Harbor-format task loader + verifier runner
## Why this repo exists
To show that a complete agent framework β including memory, planning,
tool-use, self-correction, and benchmark adapters β can be implemented
from scratch without any external LLM API. It is meant as a research
artifact, not a competitive submission.
## How to use
```python
from evo.agi_agent_world import AGIAgentWorld
from evo.agentworld_bench import load_all, WorldModelAgent, evaluate
from evo.deep_swe import load_all_tasks, run_benchmark
# 1) Run a single goal
agi = AGIAgentWorld()
result = agi.run("Buy coffee by noon")
# 2) Evaluate on AgentWorldBench (synthetic data)
records = load_all("data/agentworldbench_synth")
rep = evaluate(WorldModelAgent(), records)
print(rep.overall)
# 3) Run on DeepSWE tasks (synthetic seed data)
tasks = load_all_tasks("data/deepswe_synth")
summary = run_benchmark(tasks)
print(summary.pass_at_1)
```
## Files
- `agi_agent_world.py` β autonomous agent (memory, planning, tools)
- `omega_arch.py` β HMoE-D model definition
- `omega_nlu.py` β NLU engine (sentences, frames, FOL)
- `omega_math.py` β math reasoning primitives
- `omega_physics.py` β physics reasoner
- `omega_meta.py` β meta-cognition / audit
- `omega_quant.py` β quantization utilities
- `agentworld_bench.py` β AgentWorldBench loader + 5-dim scorer
- `deep_swe.py` β DeepSWE loader + verifier
- `tests/` β unit tests (78+ tests, all passing)
## Test results
The repository ships with 78+ unit tests covering metrics, agents,
benchmarks, and the cognitive core. All tests pass on the included
synthetic fixtures.
## Limitations
- The underlying HMoE-D model is a **demonstration** with no language
pretraining. Its `predict()` is template-based, not learned.
- `Pass@1` on DeepSWE is **not** claimed β the agent is heuristic.
- Numerical / symbolic reasoning is rule-based, not learned.
## Citation
If you use this code in research, please cite the framework only
(no leaderboard claim):
```bibtex
@software{omega_reasoner_poc,
title = {OMEGA-Reasoner: a self-contained LLM-agent framework (POC)},
year = {2026},
url = {https://huggingface.co/<your-org>/omega-reasoner-poc},
}
```
## License
Apache-2.0.
|