Darwin-9B-NEG-x-Negentropy-V8

World-first hybrid: NEG (Native Entropy Gating) × Trace Inversion

This is a Darwin V8 (Gen4) evolutionary FFN merge between two leading 9B reasoning models:

Role	Model	Contribution
Mother	FINAL-Bench/Darwin-9B-NEG	Native Entropy Gating (4M-param confidence head, GPQA 84.34% with 3-stage ensemble)
Father	Jackrong/Negentropy-claude-opus-4.7-9B	Claude Opus 4.7 reasoning chains via Trace Inversion (arXiv:2603.07267)

Method (Darwin V8 Gen4 engine)

In-memory FFN swap + 21-point grid search on (CLIcK + KMMLU + GPQA):

Mother weights: (1 − α)
Father FFN (gate_proj / up_proj / down_proj only): α
Optimal α = 0.85 selected by total weighted score
All non-FFN tensors (attention, embeddings, NEG-Head, NEG-Gate, vision tower, MTP head) inherited from Mother → NEG modules preserved intact

Benchmarks (logit-based, NEG OFF baseline)

Metric	Score
CLIcK (Korean QA, n=200)	68.5%
KMMLU (Korean MMLU, n=200)	52.5%
GPQA Diamond (n=50)	48.0%
TOTAL (weighted)	55.50

Note: Above scores are with NEG gating disabled (single-forward logit comparison). With NEG generation gating active (the Mother's signature mechanism), GPQA Diamond is expected to reach 60%+ at 1× inference cost (NEG-only Mode 1: 63.64% on the unmerged Mother).

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained("VIDraft/Darwin-9B-NEG-x-Negentropy-V8", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "VIDraft/Darwin-9B-NEG-x-Negentropy-V8",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

Lineage

Qwen/Qwen3.5-9B
   │
   ├──── Darwin V7 evolutionary merge ──→ Darwin-9B-Opus
   │                                          │
   │                                       + NEG-Head + NEG-Gate (V8)
   │                                          │
   │                                          ▼
   │                                  FINAL-Bench/Darwin-9B-NEG  (Mother)
   │
   ├──── Trace Inversion SFT ──→ Jackrong/Negentropy-9B  (Father)
   │
   └──── Darwin V8 Gen4 FFN blend (α=0.85) ──→ THIS MODEL

Citation

Trace Inversion: Zhang, Morris, Shmatikov. How to Steal Reasoning Without Reasoning Traces. arXiv:2603.07267 (2026).
NEG: VIDRAFT internal Darwin V8 architecture.
Darwin V8 Gen4 engine: VIDRAFT evolutionary merge framework.

Built on 2026-05-10 by VIDRAFT (seawolf2357).