Agentic AI & Simulation

NVIDIA's new Nemotron-Labs-Diffusion family introduces a single tri-mode architecture that switches between autoregressive, diffusion, and...

SkillOpt-Lite strips agent skill optimization to three justified components that outperform complex pipelines.

• Core principles: file-system...

RynnWorld combines digital teleoperation and 4D embodied modeling to scale robot learning data without physical hardware constraints.

-...

SIEVE selects robot demonstrations by discovering reusable visuo-motor primitives and transition interfaces, then picking medoids within high-value...

A two-stage synthetic data method generates CommonSyn, the first large-scale dataset for diversified generative commonsense reasoning, letting...

Two complementary approaches are advancing agents beyond current limits:

• Harness engineering treats the surrounding system—workflow loops,...

Most agent benchmarks rely on idealized settings with pre-packaged tools and clean inputs, understating real deployment challenges. AgentGym2 reveals...

Two complementary advances address tool-use reliability in agents:

• Architectural choice between tools and subagents determines coordination...

Anthropic's J-space gives models a reportable internal workspace for tracking hidden reasoning and misbehavior like fabricated data or planted...

• OpenAI's prompt caching cut p95 latency 25% across Realtime voice models, fixing tail delays that break phone conversations.
• The new mini model...

Enterprise AI agents face a dual barrier: 95% fail to reach production and multi-agent systems outpace traditional governance. The fix lies in...

Synthetic data is shifting from research concepts like synthetic user populations to practical enterprise tools for covering rare or sensitive...

Formal Disco tackles the data scarcity barrier in verified programming by deploying LLM agents—initiators drawing from GitHub READMEs, fixers...

NanoQuant delivers the first practical post-training method for sub-1-bit LLM compression by framing quantization as low-rank binary factorization and...

New LLM Training and Verification Methods

• 🔥 dOPSD: Introduces on-policy self-distillation for diffusion language models.
• InternVLA-A1.5:...

• MEAN framework uses five-stage agent workflow (move-to-sense, cooperative sensing, semantic processing, channel-aware mobility, uplink) with BS...

Two papers outline a practical embodied pipeline:

• InternVLA-A1.5 keeps VLM semantics intact while adding latent foresight tokens supervised by...

dOPSD offers a novel on-policy self-distillation method for diffusion LLMs that derives teacher signals directly from the model's own denoising...

CareConnect introduces a multi-layered safety architecture that combines deterministic pre-LLM intent classification with schema-constrained tool...

Evaluating robot policies has been bottlenecked by expensive real-world rollouts. GigaWorld-1 tackles this with a data-driven roadmap built on...