April 17: Model efficiency, safety gaps, and clinical AI headline a busy research day

Thursday's research output covered a wide range of machine learning topics, with particular density around model architecture, safety evaluation, and clinical applications. A handful of entries appear to cover the same underlying papers from different angles, noted below where relevant.

Architecture and efficiency. Several papers addressed how transformer and alternative architectures can be made more efficient. A cyclic channel-partitioning approach called the Three-Phase Transformer achieved measurable perplexity gains at low parameter cost (article). Separately, a two-stage distillation method was shown to transfer transformer performance into Mamba state-space models by routing through linearized attention as an intermediate representation (article). On the quantization side, researchers documented a three-phase failure pattern in INT4 post-training quantization that is tied to weight update dynamics rather than learning rate schedules, suggesting current compression timing assumptions may be flawed (article).

Safety and alignment. A new benchmark found that speech language models lose the safety, fairness, and privacy behaviors they exhibit in text when context is delivered through audio cues such as speaker identity or tone (article). A formal analysis of projection-based safe reinforcement learning showed that placing safety filters inside the policy versus at the environment level produces structurally different degradation modes (article). A new RLHF method, Rejection-Gated Policy Optimization, replaces importance-weighted sample reuse with learned differentiable gates, reducing variance during policy updates (article).

Clinical and applied AI. A study using three frontier models to score real hospital cases found that calibrated LLM panels can match expert clinician panels in diagnostic evaluation reliability (article). Two closely related entries described a retrieval-then-classify pipeline for assembling clinical code value sets that outperforms direct LLM generation (article, article). Transformer-based segmentation models, specifically SwinUNETR, outperformed convolutional networks on prostate MRI across varied annotation sources (article).

Foundations and other topics. Additional work addressed estimating the theoretical minimum classification error using soft labels without raw data (article), a formal framework for quantifying sample complexity in hypothesis verification (article), and a quantum kernel inference algorithm that eliminates training-set-size dependence via amplitude estimation (article). A queueing-theory model of cybersecurity found that symmetric AI automation in attack and defense can paradoxically raise exploit success rates, while a reinforcement-learning defense reduced active vulnerabilities substantially (article). Other entries covered machine learning analysis of antiviral drug binding to SARS-CoV-2 RNA (article), a generative augmented inference framework that treats LLM outputs as features rather than labels to reduce annotation costs (article), a comparison of foundation models and task-specific models for electricity price forecasting (article), hybrid physics-informed neural networks augmented with finite-difference regularization (article), and a modular neural system called THEIA that learns three-valued logic and generalizes to sequences far longer than its training data (article).