8% performance boost with fine-tunning via biology inspired adapters (www.genbais.com)

A startup reports that biologically inspired adapter modules — applied as lightweight fine-tuning on existing large models rather than re‑training from scratch — can yield double‑digit gains in practice, claiming up to ~8–10% improvement after an “intelligent pruning” search of just 1,000 experiments (they say this explores only ~1e‑17% of the theoretical search space). The team also ran a systematic cross‑model bias and cognitive-profile analysis across popular LLM families, finding consistent ideological fingerprints, dozens of bias types that can be dynamically elicited, and that simple leaderboard rankings miss important tradeoffs. Key outcomes: Google Gemini showed the best balance (bias score 4.2, psych avg 73.8), Llama 3 was the most consistent, while Qwen scored high on bias and low on psychological capability; OpenAI O3‑mini appears paradoxical (low bias but poor cognitive metrics). Technically, the evaluation uses six psychological dimensions (Self‑Awareness, Objectivity, Detection, Self‑Application, Consistency, Bias‑Resistance) plus a bias score; they note some metrics (Self‑Application) need refinement. Implications for the AI/ML community include a practical path to boost V+L and LLM performance without costly retraining, a demonstration that adapter‑level biological mechanisms can materially alter model behavior, and the need for multidimensional evaluation when assessing bias and capability. Tools (bookmarklet, evaluation framework, API/dashboard) and broader data coverage are promised pending funding.