The Archive

Causal writability in video models reveals correct motion latently present even when model generates physically incorrect output, enabling low-dimensional corrective edits.

Directional decomposition analysis reveals parallel components in transformer representation evolution beyond residual paths, with space-dependent asymmetry in attention/MLP updates.

Discovery Foundation Models (DFMs) proposed as next frontier: systems that participate in creating new problems and representations rather than solving human-specified ones.

Mind2Dialogue framework simulates unobserved user mental states to address supervision gap in training human-aware LMs for long-term collaboration tasks.

This story appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. This weekend, Dario Amodei, CEO of Anthropic, posted an essay calling for a brake on the pace of development of LLMs. Amodei cites the looming dangers he sees from the technology, from its…

Verifiable-by-construction clinical QA generates fine-grained verbatim citations enabling clinicians to verify claims without external document lookup.

Federated learning approach aligns record-level differential privacy with personalization by releasing private client context once and confining adaptation to fixed coefficient space.

VLoc Bench evaluates LLM agents' ability to locate vulnerable code in real repositories across 500 vulnerabilities and 147 CWE categories.

HypoEvolve combines multi-agent LLMs with genetic algorithms to test how agent collaboration patterns affect scientific hypothesis discovery quality.

Recurrent GNNs with set-based aggregation admit characterization as Boolean reachability formulas with verifiable conditions from network weights.

Real-options economic model evaluates enterprise AI adoption timing decisions under technological uncertainty, pilots, and architectural obsolescence risk.

Safe meta-RL framework enforces safety constraints during task adaptation by reasoning over information space and belief uncertainty.

SlipSense combines piezoresistive tactile arrays and accelerometers for low-latency, cross-platform slip detection in dexterous robotic manipulation.

Discrete Beckmann Transport Models enable one-step language generation via time-independent flow with provable fixed-point convergence, bypassing distillation.

Review article unifies control theory, optimal transport, probabilistic inference, non-equilibrium thermodynamics, and machine learning via free-energy optimization.

Quenched Ensemble Sampling generalizes hard energy constraints to repulsive potentials for robust high-dimensional sampling at phase transitions.

Lightweight multimodal model for Alzheimer's classification from MRI uses anatomical grounding and leak-aware contrastive learning to avoid spurious features.

Inoculation Midtraining uses neologisms to isolate unsafe behavior during training, allowing models to learn undesirable properties in controlled contexts that don't generalize at inference.

Privacy-enhanced federated learning framework combines differential privacy, homomorphic encryption, and asynchronous aggregation for distributed model training on CIFAR-10 and Purchase-100.

Theoretical work on Gaussian cooling sampling proves thin-shell stability for logconcave distributions, improving complexity bounds from n^2.75 to n^2.5.

Apple’s long-delayed Siri overhaul is finally here with iOS 27, and it changes how useful the assistant feels day to day

OptiFlow learns multimodal one-step flow policies for offline RL by treating policy learning as value-weighted optimal transport allocation.

Thanks to the launch of iOS 27, Daydream's app now includes features that can turn saved outfit photos into shoppable results and search for products through Siri without opening the app.

Schema-aware split learning framework uses LLMs as semantic bridges for privacy-preserving mental health prediction across heterogeneous survey designs.

LongAgent is an agent-based system for autonomous feature selection and longitudinal outcome prediction in heterogeneous medical datasets.

Residual Full-Rate PR architecture interprets AddUNet as filter bank and enables task-directed representation learning while preserving perfect reconstruction.

K-Bench evaluates 33 base models from 14 providers on 200 multi-turn mental health vignettes with suicide, self-harm, and domestic violence scenarios—clinician-calibrated safety benchmark.

Learning to Coach trains a dedicated LLM to extract actionable experiential knowledge from actor model trajectories for improved guidance on same and cross-instance tasks.

Deliberative Polling Diagnostic Framework benchmarks whether LLM personas update beliefs like humans when exposed to new arguments—tests dynamic fidelity for opinion simulation.

FREDI framework optimizes early-exit CNN inference on wireless edge networks via dual-threshold confidence and proportional-fair resource allocation.