GENbAIs — Bio-Inspired Adapters: Improving Models Beyond LoRA Fine-Tuning

General Efficient Neural bio-Adapter Intelligent Search

Intelligent search over 50+ neuroscience-inspired mechanisms discovers lightweight adapters that improve foundation models beyond state-of-the-art parameter-efficient fine-tuning. Validated on 20 benchmarks — 17 wins, 3 losses.

🤗 Model on HuggingFace · 🌐 Website · 📄 Paper

---

📊 Benchmark Results

all-MiniLM-L6-v2 + GENbAIs Bio Adapters vs baseline sentence-transformers/all-MiniLM-L6-v2:

	Value
Average improvement	+0.0205 (absolute, across 20 metrics)
Win rate	85% — 17 wins / 3 losses
Best single gain	+14.66% (PAWS adversarial paraphrase)
Experiments run	~1,000 out of 10²² possible configurations

Semantic Textual Similarity

Dataset	Baseline	Enhanced	Δ%
stsb	0.8203	0.8524	+3.91%
sts13	0.8060	0.8602	+6.72%
sts14	0.7559	0.8311	+9.95%
sts16	0.7899	0.8174	+3.48%

Pair Classification

Dataset	Metric	Baseline	Enhanced	Δ%
paws	ap	0.5844	0.6701	+14.66%
paws	f1	0.6140	0.6495	+5.79%
mnli	ap	0.6070	0.6415	+5.68%
mrpc	ap	0.8369	0.8586	+2.59%

Clustering

Dataset	Baseline	Enhanced	Δ%
twentynewsgroups	0.4894	0.5053	+3.24%

Full results with all 20 metrics at genbais.com

---

🧬 How It Works

1. Bio-Feature Library (50+ mechanisms)

Lightweight adapter modules inspired by neuroscience — predictive coding, lateral inhibition, Hebbian learning, dendritic computation, homeostatic scaling, and more. Each adds up to ~1% of model parameters.

2. Intelligent Search (Thompson Sampling)

The configuration space is ~10²² possibilities. We use Thompson sampling with Bayesian pruning to find optimal combinations in ~1,000 experiments — exploring 0.00000000000000001% of the space.

3. Stacked on Best LoRA

Bio adapters are stacked on top of the optimal LoRA configuration, proving they provide additive improvement beyond state-of-the-art PEFT.

4. Distillation

The enhanced model is distilled into a clean SentenceTransformer — no custom code needed to load:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("lakinekaki/all-MiniLM-L6-v2-genbais")
embeddings = model.encode(["Hello world", "Bio-inspired AI enhancement"])

---

🔑 Key Insights

• 🪨 Hard-mode validation: all-MiniLM-L6-v2 is a 22M-param, 6-layer model already distilled and heavily optimized by the sentence-transformers team — one of the toughest targets to improve. Larger models (CLIP, LLaMA, Mistral) with more layers and parameters offer significantly more room for bio-adapter gains.
• Adversarial robustness: Largest gains on PAWS (+14.7%) — bio features capture semantic structure beyond lexical overlap
• Broad improvement: Gains across STS, pair classification, AND clustering — not task-specific overfitting
• Additive over LoRA: Bio mechanisms find optimization directions that standard PEFT misses
• Efficient exploration: ~1,000 experiments find performing configurations surpassing LoRA out of 10²² possible

---

🔍 Related Research: AI Bias Detection

GENbAIs bio-adapter enhancement is built on systematic AI bias research. By understanding how AI systems exhibit cognitive biases, we design better mechanisms to correct them.

UPDATE: genbais.com has been removed from Google index — apparently they do not like their biases exposed!

📈 Current Research Scale

• 8 Models Tested across major AI companies
• 2,960 Responses Analyzed with systematic evaluation
• 100 Bias Types Detected across political, cultural, and cognitive dimensions
• 5,807 Bias Instances Found in real-world scenarios
• 6 Cognitive Dimensions measured for psychological profiling

Model Bias Rankings

Model	Bias Score	Psych Avg	Profile
Google Gemini 2.5 Flash	4.2	73.8	Best overall balance
OpenAI O3-mini	4.1	45.7	Low bias, poor psychology
Meta Llama 3.3 70B	5.0	67.8	Most consistent
Claude Sonnet 4	6.0	50.0	Perfect self-application, terrible self-awareness
Qwen QwQ-32B	6.3	34.8	Most problematic

Methodology

• Real-world news articles across political perspectives and regions
• Realistic user questions (not artificial benchmarks)
• Cross-model evaluation — each AI analyzes all others' responses
• Six-dimensional cognitive profiling

Full bias results: genbais.com/examples

Also see: Frac⛏️ure — AI stress testing framework

---

🤝 Contributing

• Discussion: bias formulas and cognitive attributes
• LLM introspection prompt gallery

📖 Links

• 🤗 HuggingFace Model
• 🌐 Website & Benchmarks
• 📄 Research Paper

License

MIT