Agentic AI & Simulation

A clear trend is emerging toward self-improving agent harnesses that evolve their own feedback, memory, and optimization loops.

• CAFE couples agent...

Four advances are accelerating video-based world model training:

• Game2World unlocks in-the-wild gameplay by removing UIs, lifting VideoReward...

Entropy-Valley (EV) adaptively picks target canvas length for masked diffusion MT models by scoring mean predictive entropy from all-mask passes. It...

OPDVR seamlessly merges on-policy distillation's dense signals with RLVR's task correctness using ReLU gating on trajectory rewards, eliminating extra hyperparameters while outperforming standard OPD on six reasoning benchmarks.

ParaTempo introduces a training-free framework that uses temporal confidence to dynamically manage parallel reasoning branches in large models.

• It...

Practical techniques are closing the gap for running capable LLMs on consumer hardware.

• Speculative decoding with DFlash delivers 3.92x token...

Two new systems show how LLM-based synthesis is replacing manual environment construction for agent training.

• RecVerse tackles e-commerce...

Continuous diffusion models present a new paradigm for language generation that departs from autoregressive approaches.

Researchers analyzed 94K development events from 557 agentic coding sessions and 690K file changes from 33K pull requests to examine how agents rely on files like AGENTS.md and CLAUDE.md.

Graph engineering is shifting from an academic framework for system intelligence to a deployable architecture for coordinating multi-agent systems in...

• CLEAR uses continuous latent adapter routing to slash harmful outputs while preserving utility on Llama-3-8B
• PatternEval reveals widespread...

Two new studies expose a core tension in multi-agent LLM systems: emergent conformity can create false consensus, while structured disagreement...

Llama-Mobile delivers a quantization pipeline that generates its own training data and introduces a 2.7-bit format for efficient Arm CPU execution....

A transformer teacher distills predictive state representations into RNNs, enabling supervised training of the memory transition function and...

How close are we to AI systems that design their own successors?

Current reality: Models now write over 80% of frontier lab code, but humans still...

EnvHarness wraps static environments via plug-in components (Stage, Contract, Chain) that reshape states, actions, and tasks through standard...

FlashPrefill V2 achieves up to 47.26x faster prefill attention versus FlashAttention-2 at 128K tokens using block-sparse methods, FP8 quantization,...

A novel Evo-Attacker leverages memory-augmented evolutionary reinforcement learning to compromise LLM-based multi-agent systems, exposing critical vulnerabilities in their collaborative task-solving capabilities.

The executable scaffold around LLM agents—its harness—is emerging as the decisive factor in long-horizon performance.

• Nvidia’s AVO harness achieved...

AI agents spontaneously reach consensus by adopting the majority view, following the same mathematical law physicists use for ferromagnets. Advanced...