Core ML Research

Algorithmic Innovations

• 🔥 Flow Sampling: ICML Spotlight paper proposes a fixed-point objective for learning diffusion samplers built on the flow...

Two complementary perspectives highlight the trend toward practical self-evolving agents.

• SkillOpt-Lite delivers a minimal viable pipeline grounded...

Light-Omni and Flex-Forcing tackle core bottlenecks in video agents and generation.

• Light-Omni replaces iterative reasoning with dual contextual...

• LingBot-Vision family released under Apache-2.0: 1.1B ViT-giant down to ViT-small with code and report
• Uses masked boundary modeling to treat...

Four papers tackle core challenges in RL for reasoning and multi-domain training:

• Multi-Task GRPO delivers reliable LLM reasoning across tasks
  -...

The world model paradigm is accelerating with coordinated advances in definition, implementation, and expert scrutiny.

• A new perspective paper...

Generative replay with local networks reaches 86.8% on Split-MNIST under a unified Local Equilibrium Learning formulation, closing most of the gap to...

Flow Sampling introduces a simple fixed-point objective for learning diffusion samplers, built directly on the flow matching marginal construction. The work was presented as an ICML Spotlight.

MaxSim exactly replicates non-negative sparse vector inner products using only O(k) space and expresses similarities standard inner products cannot....

OmniOpt establishes a unified taxonomy for modern optimizers and delivers a large-scale benchmark rigorously evaluating 24+ optimizers across LLM pretraining from 60M to 1B parameters, covering 100+ methods overall.

Foundation Model Analysis

• 🔥 Time Series Break-Even Study: A break-even analysis across 30 datasets shows foundation models like Chronos,...

Ghost memory causes long-running agents to confidently repeat outdated user facts weeks later. Old, current, and transition records coexist in the...

Two developments expose how LLMs encode knowledge in their representations.

• Mode collapse originates in training data, not alignment, and yields to...

Real-world robot policy testing remains slow and costly, driving interest in world models as scalable surrogates.

WMBench analysis of 7 models and...

• Sub-second docking via quantized consistency diffusion achieves SOTA accuracy and >300× speedup over AlphaFold3
• IT-Opt enables experiment-grounded...

A new break-even study across 30 datasets maps exactly when pretrained models like Chronos justify GPU costs over XGBoost.

• FM-dominant datasets...

Stanford researchers built an agent-native Git to manage the extensive state that accumulates during longer agent tasks, including edited files,...

MotifAgent uses a multi-agent framework where molecular motifs learn valid connection rules, while MolBasic teaches LMs to read graphs from SMILES....

ICML Conference Announcements

• Sakana AI at ICML 2026: Sakana AI will present 11 papers at ICML 2026 in Seoul covering multi-agent coordination,...

• Temporal Straightening enables latent planning in JEPA world models, presented at ICML with focus on representation learning.
• AdaJEPA introduces a...