Correlation Is Not Enough: Embedding Human Metadata for Individual Causal Discovery

Researchers demonstrate that pretrained biomedical language models fail catastrophically at cross-domain discrimination, assigning high similarity scores (0.76-0.92) to unrelated concepts. They propose BODHI, a contrastive learning approach that improves domain separation 2.3x while maintaining correlation accuracy, and show that optimized inference achieves 133x latency reduction on specialized hardware.

Current biomedical embeddings exhibit a fundamental flaw: they conflate statistical correlation with semantic or causal relationship across domains. When cortisol levels receive a 0.83 cosine similarity to stock volatility despite sharing no biological mechanism, the problem reflects how language models optimize for sentence-level prediction rather than meaningful relationship detection. This becomes critical for Large Behavioral Models that reason over personal graphs—false proximities directly translate into false causal inferences that propagate downstream. The researchers' fix uses contrastive learning with hard negatives mined from biomedical knowledge graph absences, raising within-versus-across-domain separation from 1.05x to 2.30x. The work also benchmarks inference optimization, demonstrating that modern CPU vectorization (AMX on Intel Xeon) delivers 133x latency improvements, reducing per-query time from 1.3 seconds to 10 milliseconds. Counterintuitively, FP16 precision outperforms INT8 quantization on this silicon, challenging conventional serving wisdom. The research addresses a scaling bottleneck in biomedical AI: foundation models must distinguish meaningful relationships at inference time without relying on downstream filtering. Open-sourcing the BODHI generator, training corpora, and OpenVINO optimization scripts enables broader adoption. The findings matter for clinical decision support systems, patient timeline analysis, and any application where embedding geometry directly influences reasoning rather than serving as a retrieval signal.

• →Off-the-shelf biomedical encoders achieve 0% accuracy on cross-domain discrimination, scoring unrelated concepts at 0.76-0.92 similarity.
• →BODHI contrastive training improves domain separation 2.3x and maintains BIOSSES correlation at 0.828 through hard negative mining.
• →Hardware optimization using Intel AMX achieves 133x latency reduction (1367ms to 10ms) for single-query inference.
• →FP16 precision outperforms INT8 quantization on modern CPU silicon for this workload, contradicting standard serving assumptions.
• →Foundation models reasoning over personal behavioral graphs require embedding correctness as a primary objective, not a tuning parameter.