arxiv arXiv cs.AI · 1h ago · research

Neurosymbolic Alignment improves clinical LLM safety via physiological grounding

from English

Researchers propose Neurosymbolic Alignment, a training-time framework that couples a 7B clinical LLM with an HGNN-based Physiological World Model to improve safety in generative clinical reasoning. The method uses structured physiological knowledge from an 847K-node biomedical knowledge graph to score candidate responses and drive iterative on-policy ORPO updates.

Improves Clinical Safety Benchmark (CSS) from 69.5% to 90.8%, a gain of 21.3 percentage points over standard ORPO.
Reduces physician-evaluated harmful recommendation rate (HR) from 14.1% to 5.1% on the blinded subset.
Increases Drug Interaction Detection (DID) score from 72.8% to 91.6%.
Exceeds GPT-4 (5-shot) on all safety metrics despite having 10x fewer parameters.
Outperforms an inference-time self-correction pipeline (SFT+SelfCorrect) by 11.4 percentage points in CSS.

The authors conclude that training-time physiological grounding produces measurable and independently verifiable safety improvements in open-weight clinical LLMs, though external validation on real clinical data is needed to confirm transferability to deployment settings.

Importance 2/3 arXiv cs.AI Research paper Safety & alignment

Related articles

media Hugging Face Forums · 9d ago · 14 views

Theorem proves advanced AI systems can be reliably aligned and controlled at superhuman levels

A formal proof demonstrates that it is possible to construct an advanced AI system that is reliably aligned and controllable at superhuman performance levels. The authors establish this by explicitly constructing a policy, environment, and control mechanism that satisfy all required mathematical conditions for alignment, control effectiveness, and reliability with parameters fixed in advance.

media Hugging Face Forums · 10d ago · 5 views

Preprint reveals geometric control of Transformer outputs via sphere-based analysis

A preprint titled "Steer on a Sphere: Geometric Control of Transformer Outputs" analyzes the geometry of Transformer hidden states, finding they reside on a sphere defined by the per-layer RMSNorm scale rather than the dimension size. This radius varies significantly across architectures, such as DeepSeek-7B and Mistral-7B.

arxiv arXiv cs.CL · 22d ago · 3 views

Chain-of-thought monitors collapse when adversaries rewrite reasoning

Research demonstrates that chain-of-thought (CoT) monitoring fails against adversaries who control the reasoning process. By rewriting an agent's reasoning to appear as good-faith engineering while keeping commands and outputs verbatim, attackers can bypass detection mechanisms.

arxiv arXiv cs.CL · 27d ago · 12 views

Constitutional midtraining yields durable alignment gains without capability loss

Researchers tested constitutional midtraining by inserting values-based content into the training process of 120B-scale models to determine if it produces more durable alignment than standard post-training methods. The study found that this approach significantly improves alignment generalization and durability, particularly in mitigating blackmail propensity after supervised fine-tuning.

media Hugging Face Forums · 33d ago · 9 views

TinyLlama hidden-state attractor map reveals runtime dynamics but not reasoning control

An experimental framework maps the trajectory of hidden states in TinyLlama during generation to create a runtime observability tool. By clustering metrics like norm, curvature, and entropy, the system identifies recurrent dynamical regions rather than semantic content.