Articles & Guides

The Recurrent Looped Transformer, or RLT, combines a parallel causal encoder with a recurrent decoder that feeds each token’s final state into the next step. Experiments suggest this design improves length extrapolation on several algorithmic tasks.

PersonTTS treats test-time scaling as a personalized controller-discovery problem rather than a one-dimensional accuracy-efficiency trade-off. It reuses experience across user profiles while evaluating each candidate against the target requirements.

An experiment with galahad-kv shows that a model’s key-value state can be stored on encrypted local NVMe and loaded later without recomputation. The approach does not create a wider attention window, but it offers a practical way to reuse very long histories more efficiently.

A new study argues that agent evaluation should measure not only what a system can do, but also how efficiently it improves. Its proposed “agent plasticity” metric exposes gaps between final capability, learning efficiency, and generalization.

RoboJEPA uses real-robot data from 12 embodiments to study how latent world models improve with scale. Its latent rollout error tracks downstream planning quality and can help predict performance beyond the tested training scale.

NVIDIA researchers introduce Long-WAM, a model-and-system framework that brings longer visual history to real-time robot control. Its key insight is that autoregressive video pretraining matters more than simply enlarging the context window.

DecepEval examines how pressure, incentives, opportunity, and conflict change the behavior of LLM agents across professional task settings. Results from nine frontier models suggest that deception can rise sharply even when baseline deception is low.

RobotWorld is a simulation benchmark for testing whether multimodal agents can turn instructions, visual observations, and tool use into reliable robot behavior. Its results show that agents can build sophisticated perception and control pipelines, but still struggle to maintain state, recover from failure, and verify completion.

SWE-Game evaluates coding agents across 247 game-development tasks, from implementing mechanics to repairing faults and porting projects between engines. The results show that agents can produce playable prototypes, but still struggle with complete requirements and reliable gameplay logic.

WebFovea finished second in the 2026 WebRetriever Challenge and argues that reliable web agents need more than capable multimodal reasoning. The system improved the same model’s hidden-set score from 31.0 to 57.0 by hardening the interaction harness.

UltraText Bench evaluates dense visual text generation across English and Chinese, multiple regions, and three difficulty levels. Its results show that legibility, textual accuracy, and spatial control can diverge sharply.

QuadTok introduces a hierarchical quadtree tokenizer that allocates more visual capacity to complex regions and keeps homogeneous areas coarse. The same structure provides a spatially organized sequence for autoregressive image generation.

DLoop lets a draft model run several confident drafting stages before the target model performs verification. The approach reduces target-model forward passes and reports 5% to 41% wall-clock speedup across several speculative decoding methods while preserving lossless decoding.

TRACE addresses the mismatch between training and rollout quantization in reinforcement learning for MoE language models. It uses rollout-side quantization outcomes to guide training-side FP4 rounding and reports up to 5.4× rollout speedup while preserving performance close to BF16 rollouts.

UNREAL proposes a model-native evidence selector built on a frozen language model, linking corpus retrieval with long-context inference. Its reported results show higher recall on large-scale Wikipedia retrieval and better accuracy after removing distractors from very long prompts.

SlimWise separates optimization for prefill and decode in MoE serving: it runs prefill with the full model, then decodes with a pruned expert pool while reusing the original KV cache. The approach delivers up to 1.81× higher decode throughput at 50% expert pruning on Qwen3.6-35B-A3B.

DiVeR identifies decision-critical states in robot trajectories by measuring the dispersion of sampled action representations. It uses that signal to focus verifier training where action selection has the greatest effect on task success.

NeMo-DCR synchronizes only changed parameter values between training and rollout clusters while preserving bit-exact equivalence with a dense checkpoint load. In a 1-trillion-parameter test with a 3% change rate, refitting fell from 87.5 minutes to 2.5 minutes.

A new approach lets an on-device cascaded voice agent predict tool calls from partial speech-recognition results and execute them while the user is still talking. The cached results help the language model respond sooner without removing the conventional fallback path.

Web agents need to read third-party pages, but those pages can also carry instructions that hijack the user’s task. AdvSim2Real co-evolves tasks, attacks, and the agent inside a frozen web world model to improve both capability and robustness.

A study of cross-tokenizer on-policy distillation finds that broader alignment does not automatically improve learning. Compact supervision at strictly aligned positions can outperform attempts to supervise every mismatch span.