AI Frontier Digest

Large language models can automatically optimize large-scale software systems by drawing on a catalog of 43 performance optimization patterns...

NVIDIA's midtraining stage raises SWE-bench Multilingual from 55.8 to 62.8 on Nemotron 3.5 Lightning while keeping the SFT stage identical.

Interactive agents benefit from emitting constrained structured actions directly rather than free-form text at each step.

• Generating free-form...

DeltaStreamer introduces a human-like agentic system for streaming video that maintains change-aware memory before queries arrive, then uses...

Routing Policy Optimization (RPO) turns attention routing into a stochastic exploration action at selected denoising levels, using a controlled...

• EdgeWise pairs a local SLM as primary solver with selective cloud-LLM calls only for critical steps
• A step-level RL method jointly optimizes task...

Endpoint-only distillation misses inconsistent local policy changes during RL training, as identical final shifts can arise from reinforcing or...

Instead of letting agents bloat their histories with iterative edits, periodically discard sessions and restart from the best candidate found so far....

• LangDiT redesigns Transformers for continuous diffusion LMs with feature normalization, noise conditioning, and confidence-gated self-conditioning,...

AI materials discovery currently optimizes only for functional performance, deferring manufacturing, environmental, use-phase, and end-of-life...

Data utility for LLMs is training-dependent, not intrinsic to the text: it varies with the target objective, model, and current state. Strategies are...

• Standard proxy evaluations can reverse dataset rankings with modest learning-rate changes, risking misdirected full-scale pretraining.
• Training...

• Super VL release: 120B open model delivers frontier-level multimodal agent accuracy at ~1/5 the size of Ultra, with 5x faster task completion.
  -...

• Prompt or frame passing alone cannot supply a shared spatial frame or time-indexed record for revisiting scenes or events.
• ATAS maintains explicit...

• A non-IT VP at Grand Isle Shipyard builds fleet and safety agents by describing tasks in plain language to Samsara Agent Studio
• One early agent...

A new survey proposes a two-axis framework to design memory systems that let multimodal agents learn continuously from heterogeneous data.

• Axis I...

Benchling's Science Foundries restructure scientific work by assigning agents responsibility for entire processes while scientists focus on high-level...

• HybridAgents dynamically switch between efficient latent reasoning and tool-integrated text actions via HybridLink and continuous memory
• Dedicated...

A large-scale study of ten self-distillation methods versus GRPO reveals why headline benchmark scores mislead agentic RL progress.

• Limited gains:...

Humanoids require tools to exceed physical limits, yet existing benchmarks ignore joint selection, manipulation, and locomotion.

HumanoidToolBench...