Agentic AI & Simulation

Panel enables agents to dynamically create custom research panes like viewers or browsers alongside core tools. Modules support typed workflows with...

Two contrasting paths emerge for harder agentic RL tasks:

• FlashREINFORCE cuts rollout costs in half versus GRPO while matching or exceeding...

Atria Dawn Preview shows deployable agentic capability for research and engineering without claiming superintelligence or recursive...

Agentic image restoration can be framed as a closed-loop control problem where agents inspect outputs, select restoration actions, and iteratively correct errors, per a controller-centred synthesis reviewing studies from 2018–2026.

BVB benchmarks whether agents truly understand videos by requiring them to reconstruct scenes and actions as executable Blender programs. The...

Proving core Transformer invariants from scratch in Lean shows what mechanical trustworthiness for LLM components requires. Verified properties could...

LARA-HPC pairs agentic planning with executable validation checks to automate atomistic and HPC workflows, tackling the reliability gap in current LLM-driven scientific systems.

• GRPO-style RL eliminates critics via multiple rollouts per prompt but struggles with long-horizon agentic tasks due to high compute and sync...

Reliable production agents require an explicit supervisory layer that owns permissions, state, recovery, and stop conditions rather than relying on...

A controlled multi-agent system explicitly separates intent classification, service selection, and response generation into auditable stages with...

Can learned context ranking reduce long-context attention cost without sacrificing the information agents need?

• SAS introduces simple attention...

Specialized inference engines enable open Qwen3-TTS and Qwen3-ASR to hit sub-50 ms latency at 10 RPS, topping closed models like 11Labs on accuracy...

Otis offers a minimal agent runtime that auto-recommends and downloads hardware-matched local models, then executes them via llama.cpp while...

Uniform sampling under GRPO delivers a 7.76-point accuracy gain on domain-balanced benchmarks and is not outperformed by any of eight...

Recurrent Looped Transformer (RLT) carries the full decoder state—final output plus layerwise SWA cache—across every token with no reset, creating a...

State-of-the-art language models reached 82.4% accuracy identifying detectives from anonymized narratives using investigative traits, showing how behavioral signals can drive structured story generation beyond generic text output.

Booting straight into a local LLM on Raspberry Pi without Linux shows how ultra-lightweight deployments can turn small devices into dedicated edge inference appliances.

Small tool-using agents map natural-language requests to action names and JSON arguments via a 4B-parameter Gemma model, yet suffer semantic collapse that targeted diagnostics can now identify and repair.

OpenAI's new Agents API keeps the harness—sessions, memory, recovery—while partners like Blaxel run the self-hosted execution sandboxes. This emerging...

CFD-copilot is a domain-specialized LLM framework that enables natural language-driven CFD simulation from setup to post-processing.