Machine Learning

S*: Hybrid test-time scaling for code generation, boosting both coverage and selection accuracy.

MLGYM: A new framework & benchmark to advance AI Research Agents

Logic-RL unlocks LLM reasoning via rule-based reinforcement learning, generalizing to math problems after training on logic puzzles.

PURE: LLM-driven user profile management boosts recommendation by harnessing user reviews for personalized insights while tackling token limits. PURE enhances LLMs for better recommendations.

Unveiling the surprising potential of noise: transferable knowledge in semi-supervised heterogeneous domain adaptation (SHDA).

AdaptiveStep: Divides reasoning steps automatically through model confidence, enhancing PRM training & performance.

S2R: Teaches LLMs to self-verify and self-correct, boosting reasoning with efficient reinforcement learning.

NExT-Mol: Combines 1D language models with 3D diffusion for molecule generation, achieving state-of-the-art performance and validity.

Eager updates drastically speed up training massive language models by cleverly overlapping communication and computation in DiLoCo, achieving near-optimal performance even with low bandwidth.

ThinkPO improves LLM reasoning by preferring longer CoT, boosting performance without new data.

Small language models struggle to learn complex reasoning from large models, but a novel ‘Mix Distillation’ method balances complexity for effective capability transfer.

High-quality, task-relevant pretraining data surpasses large-scale pretraining in atomic property prediction, achieving comparable performance at 1/24th the computational cost.

MIKASA, a new benchmark for memory-intensive reinforcement learning, provides a unified framework for evaluating memory capabilities in diverse scenarios, including complex robotic manipulation tasks.

AdaPTS effectively adapts pre-trained univariate time series models to probabilistic multivariate forecasting, improving accuracy and uncertainty quantification.

ProbeLog: Zero-shot model search directly from weights, boosting efficiency and accuracy!

Agency, a key concept in AI, is shown to be relative to the observer’s perspective (frame-dependent), challenging traditional binary definitions and necessitating a more nuanced approach for AI system…

AI agents now master complex tasks with improved Transformer World Models, achieving a new state-of-the-art in data-efficient reinforcement learning.

AceCoder uses automated test-case synthesis to create a large-scale dataset for training reward models, enabling effective reinforcement learning to significantly boost code generation model performan…

W2SD: A novel framework boosts diffusion model quality by using the difference between weak and strong models to refine sampling trajectories, achieving state-of-the-art performance.

Streaming DiLoCo achieves two orders of magnitude bandwidth reduction in billion-scale parameter LLM training by synchronizing parameter subsets sequentially, overlapping communication with computatio…