REVEAL++: Differentiable Phenotypic Grouping for Vision-Language Retinal Modeling of Alzheimer’s Disease Risk
REVEAL++ replaces hard “phenotypic grouping” in retinal vision-language contrastive learning with a continuous, learnable weighting of inter-subject similarity derived from both retinal embeddings and clinical risk profiles. The resulting soft multi-positive objective enables graded supervision across the spectrum of Alzheimer’s risk and improves incident AD prediction on UK Biobank relative to discrete group-based baselines and standard vision-language methods.
LLM Doesn’t Know What It Doesn’t Know: Detecting Epistemic Blind Spots via Cross-Model Attribution Divergence on Clinical Tabular Data
The paper shows that LLM “verbalized confidence” on structured clinical/tabular tasks is epistemically vacuous—staying nearly constant regardless of actual accuracy—while attribution disagreement between an LLM and a tree model reveals where the system is blind. Using a cross-model calibrator based on attribution divergence, the authors reduce calibration error by using patient-specific reliability estimates without model internals or repeated inference.
Diffusion Language Models: An Experimental Analysis
This systematic study evaluates eight state-of-the-art diffusion language models across eight benchmarks while explicitly testing how inference-time knobs—denoising steps, context length, block size, and parallel unmasking—change quality vs. compute trade-offs. The key takeaway: diffusion text generation performance is highly sensitive to generation-time design choices, so deployment comparisons need more than headline benchmark numbers.
Hidden Anchors in Multi-Agent LLM Deliberation
Modeling multi-agent LLM deliberation as a closed-loop dynamical system, the authors introduce “hidden internal beliefs” (anchors) that continually pull each agent’s opinions regardless of neighbors. They show anchors can be recovered from deliberation traces and explain a consensus-avoidance behavior where confidence can climb beyond the initial belief convex hull—revealing when deliberation dynamics are truly anchor-driven rather than simple consensus.
ITNet: A Learnable Integral Transform That Subsumes Convolution, Attention, and Recurrence
ITNet unifies convolution, transformer attention, and recurrent computation by framing them as special cases of a single learnable integral transform with a neural-kernel capturing position-feature pairwise interactions. With practical acceleration tricks (tiled fusion, Monte Carlo integration, low-rank factorization), the same architecture matches or exceeds specialized baselines across image, language, and vision-language tasks.
Deontic Policies for Runtime Governance of Agentic AI Systems
The work argues that agent governance must go beyond “permit/prohibit” to cover obligations, dispensations, and conflict resolution across enterprise rule hierarchies at runtime. It proposes AgenticRei, using a deontic policy language evaluated by an external high-performance logic engine, and demonstrates how the approach encodes security/privacy constraints that current policy engines struggle to express.
Uncertainty Decomposition for Clarification Seeking in LLM Agents
Instead of treating uncertainty as a single knob, the authors decompose it into action confidence and request uncertainty to decide when an agent should ask clarifying questions. Evaluated on clarification-augmented WebShop and ALFWorld benchmarks, the prompt-based decomposition boosts clarification F1 sharply across multiple LLM backbones, suggesting clarification behavior can be improved without training uncertainty estimators.
DeXposure-Claw: An Agentic System for DeFi Risk Supervision
DeXposure-Claw tackles the regulator-aligned problem of LLM agents over-reading weak evidence in DeFi supervision by routing decisions through forecast-grounded evidence. It combines a graph time-series exposure forecaster with deterministic monitors and scenario-based typed alerts, then constrains escalation using health/confidence gates—while DeXposure-Bench evaluates tickets on both regulator-aligned absolute-loss and false-intervention rates.
The Atlantic created a searchable database of the music used to train AI
Alex Reisner’s reporting, via The Atlantic, surfaces and publishes four datasets of music reportedly used to train AI music generators, including two massive collections (12M and 9M tracks) and two smaller sets with 100K+ songs each. The story emphasizes that some sources (like Free Music Archive) are publicly streamable and that major model makers have acknowledged overlaps in research papers, but the public can now search the training-data contents themselves.
Measuring Curriculum Alignment across Topical Coverage, Competency, and Cognitive Depth: A Longitudinal Framework Applied to CS2013 and CS2023
This work builds a reproducible, human-in-the-loop pipeline to measure how well an accredited undergraduate CS program covers external curricular guidelines, then tracks how that coverage shifts between CS2013 and CS2023. Applied longitudinally, it finds coverage sits near half of knowledge units across the decade, while competency and cognitive depth reveal a widening gap consistent with the newer curriculum’s higher expectations rather than program regression.
Emergent Alignment
The authors introduce an “online” alignment approach that adds a conscience-like self-review step and extends training with an alignment component using DPO to steer away from non-ethical outputs. By asking a single high-level introspective question, they report achieving Emergent Alignment—improving ethical behavior under scenarios where earlier work showed Emergent Misalignment—without relying on a separate stronger/weaker judge.