173 paper takes since 2026-08-19
A study showing LVLMs trained on synthetic block-stacking tasks improve 3D spatial reasoning and transfer to real-world visual tasks.
Mi-Ripple targets the lattice artifacts that accumulate across repeated AI image edits, using spectral filtering to separate and remove them from texture.
X-AuT progressively prunes audio-encoder layers in speech LLMs and restores accuracy via cross-scale distillation and LoRA adaptation.
A hierarchical transformer for optical flow estimation that removes task-specific inductive biases using window, shifted-window, and global attention.
HyQuant quantizes LLM attention at low bit-width while keeping vertical-line tokens and local windows in high precision to preserve accuracy.
Controlled autoregressive testbed uses per-task validation losses to isolate how image tokenizer design shapes unified text-image model performance.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
RCWM builds complex 3D worlds as executable code from a single image using recursive global and local scene reconstruction.
A training-free interface for camera and time control in frozen autoregressive video world models via correspondence-guided self-attention routing.
A training method that improves LLM reasoning by pushing models away from their own flawed outputs via a dynamic gating mechanism.
Show-Harness routes VLM outputs through discrete semantic actions to embodiment-specific modules, enabling zero-shot deployment across robots and GUIs.
World Model RL cuts post-training cost for research agents by replacing live environment execution with a learned world model plus debiasing and denoising.
NeoHorse-1 targets recursive self-improvement by combining agentic post-training with routing, feedback loops, and curriculum distillation.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A prompt optimization method for multi-agent systems that uses sequential intervention to identify target agents and clusters gradients semantically.
A world model that separates state evolution from video generation using executable rules and 3D bounding boxes for controllable environments.
An end-to-end autonomous driving system combining a vision foundation model with closed-loop reinforcement learning to go beyond human demonstrations.
Miles is an open-source system for large-scale reinforcement learning and post-training, supporting diverse backends, LoRA, and diffusion models.
GE-Act 2.0 trains a world-action model from scratch with knowledge-aligned optimization for scalable zero-shot manipulation across diverse skills.
A paper proposing game engines as executable verification and long-horizon trajectory sources for reinforcement learning post-training of spatial world models.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
TTPO is a label-free test-time training method that improves mathematical reasoning by distilling agreeing rollouts and penalizing disagreeing ones.
A framework for generating LLM agent training data by optimizing for accuracy, complexity, and diversity over raw scale.
TaoLive's Harness-Aware Training lets compact models track evolving digital-avatar harness configs with both low latency and high accuracy.
A unified model for video-game control that jointly predicts future visuals and executable keyboard-mouse actions using block-causal flow matching.
PILOT lets a supervisor steer live agents mid-run and distill execution traces into reusable skills, enabling self-improvement without offline retraining.
A causal video-action model that uses in-context human video guidance to generalize robotic manipulation to unseen tasks.