The Archive

PrismaCoder open-source multimodal model for code-to-visual generation using decomposed rubric synthesis to improve data quality and training pipelines.

Analysis of simple relational baselines versus GNNs and Transformers for temporal heterogeneous graph prediction over relational databases.

Self-supervised tokenizer via pointwise mutual information factorization avoids dimensional collapse compared to DINO and LeJEPA in generative representation learning.

Counterfactual Evidence Interaction Graphs enable multimodal agents to decide which visual evidence to acquire under cost constraints and frozen execution traces.

PolyNets couple neural polynomial dynamics with recurrent networks for training-free physics emulation without learning substitutes for known equations.

Multi-developer agentic systems outperform single-developer systems on GAIA benchmark by enabling collaborative frontier capability development.

VASPO integrates privileged value estimation with PPO for improved self-distillation in reasoning models, outperforming GRPO-based approaches.

Bayesian online bandit algorithm selects classifiers balancing predictive performance and fairness subgroup metrics with delayed label feedback.

Two-level semantic novelty framework prioritizes 99,527 disease trajectory motifs from MIMIC-IV EHRs using embeddings and clinical reliability.

VascuSR combines iterative diffusion with 3D vascular guidance for anatomically-plausible CT super-resolution avoiding spurious vessel hallucination.

AI design tools make proteins that break bonds that are stable for hundreds of years.

ReaDy leverages hybrid linear attention recurrent states for position-specific context in speculative decoding, accelerating LLM generation.

Graph-enhanced framework discovers 99,527 disease trajectory motifs from MIMIC-IV with leakage-controlled splits and time-aware Transformer endpoint forecasting.

HP-RLVR assigns token-level credit in single-rollout RL without group sampling by using trajectory hindsight, improving on GRPO.

Study quantifies patient privacy leakage from frozen pretrained MRI encoders using membership inference on OASIS-2 and MIRIAD datasets.

This is Optimizer, a weekly newsletter sent from Verge senior reviewer Victoria Song that dissects and discusses the latest gizmos and potions that swear they're going to change your life. Opt in for Optimizer here. For the past 100 days or so, my mornings have had a new rhythm. When I wake up, the first thing I have to do is feed my digital twin. It consumes my metabolic data, and its first "meal" of the day is my body composition data, sent straight from a smart scale. Once that's logged in the app, I scoot over to the kitchen to fill an Owala with lemon water to take my medications with. T...

Novel RL algorithm learns optimal information-gathering policies from single continuing trajectory without sample-efficiency penalty.

Decoupled Marginal Sharpening enables training-free LLM inference-time scaling via improved power-sampling without reward models.

Bias audit across 10 open multilingual LLMs in 8 languages shows stereotype score variance is measurement noise, not fairness signal.

AdaRSI framework selectively retrains vision-language-action model executors via on-demand recursive self-improvement for long-horizon embodied tasks.

SafeAdapt-Poker applies counterfactual risk calibration to budgeted adaptation in adversarial games with opponent behavior shifts.

ChronoCulture benchmark of 10,360 temporal pairs from 72 countries shows six LLMs fail to track cultural attitude evolution over time.

Selective Growing Q-Networks replaces uniform discretization with adaptive action grids for continuous-control Q-learning on DMControl tasks.

Pretrained tabular foundation model TabPFN generalizes to power grid steady-state analysis (power flow, contingency screening, voltage stability) without retraining.

semCER metric weights OCR character errors by severity via LLM judge, outperforming uniform Character Error Rate across English, Vietnamese, Chinese.

OpenAI claims a former employee took more than 100 files related to its product plans before joining Apple. In a document filed on Wednesday, OpenAI accuses the unnamed employee of sending this information to a personal account and then presenting it to Apple leadership, including the vice president of the company's operations in Asia. The filing, reported earlier by Reuters, is tied to a legal battle between Apple and OpenAI over trade secrets - but this time, OpenAI is doing the accusing. It alleges that the files in question contain images of at least one unreleased device, along with supp...

MolLean introduces a Lean-based verifiable language for chemical reasoning that grounds claims in explicit molecular state to prevent hallucinated structural properties.

Study shows LLM agents and scientific verifiers tacitly select single interpretations of underspecified claims, causing benchmark disagreements that hide interpretation variance.

OpenAI engineer faces 5 years and $250K fine as FBI lobs foreign spy charges.

COUNT method scales language model memory via n-gram embeddings stored in host RAM instead of GPU, enabling training-free adaptation without expensive GPU writes.