HumanTracker introduces a large-scale benchmark and a preference-aligned metric, HumanScore, designed to evaluate humanoid motion tracking beyond simple kinematic errors. By focusing on physical stability and contact realism, it addresses the disconnect between traditional pose-difference metrics and human-perceived quality.
The GAS framework introduces a novel training paradigm that utilizes visual generation as auxiliary supervision to enhance multimodal understanding. By employing a decoupled architecture, it achieves performance gains in spatial precision and visual retention without incurring any additional inference overhead.
The University of Southern Denmark has released Mimir v1, a 1-billion-parameter model built on the Hierarchical Reasoning Model architecture using strictly permissible data. It achieves state-of-the-art performance for Danish while remaining highly competitive in English benchmarks against larger models.
PACE-Bench introduces a rigorous evaluation framework for self-evolving agents, revealing significant failures when adapting code to dynamic physics environments. The benchmark demonstrates that current models struggle with structural mechanism redesign, highlighting a major gap between parameter inference and functional adaptation.
V-RAE shifts the paradigm of video latent generation by utilizing frozen foundation models rather than training reconstruction-heavy autoencoders from scratch. This approach improves generative quality and convergence speed by prioritizing semantic structure over pixel-perfect reconstruction.
HarnessEval-W introduces a hierarchical, agent-based framework for evaluating world models, replacing opaque scalar scores with verifiable evidence trees. By decomposing complex visual rollouts into specialized sub-problems, this pipeline enables fine-grained diagnostics of causality and physical consistency.
A comprehensive empirical study demonstrates that pixel-space diffusion models can outperform latent-space counterparts by utilizing a latent-to-pixel training strategy. This approach achieves significant inference speedups while maintaining high generative quality, challenging the current dominance of latent-space architectures.
Researchers have achieved a new upper bound for the matrix multiplication exponent, ω < 2.371177, by combining reformulated optimization techniques with AlphaEvolve. This advancement refines the long-standing combination loss analysis method, pushing the theoretical limits of computational complexity.
PixRestore introduces a pixel-space Diffusion Transformer for unified image restoration, bypassing the limitations of VAE-based latent diffusion models. By training from scratch and utilizing flow matching, the model achieves high-fidelity results with significantly reduced parameter counts and single-step inference.
Researchers have introduced aDSL, a domain-specific language designed to align LLM reasoning capabilities with 3D geometric constraints. By replacing absolute coordinate generation with relational operators and a multi-agent feedback loop, the system significantly improves the reliability of programmatic 3D asset generation.
GS-Voxel introduces a fitting-free framework that converts irregular 3D Gaussian Splatting reconstructions into structured, sparse voxels. This enables scalable, image-conditioned generation of large-scale 3D scenes without the overhead of per-scene optimization.
A new research paper introduces Chain-of-Experience (CoE), a framework allowing LLMs to learn from iterative feedback at test time. By accumulating experiential traces, models achieve higher accuracy and token efficiency compared to traditional zero-shot inference.
A recent study of frontier mixture-of-experts models reveals that standard accuracy benchmarks are insufficient for evaluating language adaptation. While supervised fine-tuning shifts reasoning to the target language, reinforcement learning is essential to correct systemic formatting and leakage defects.
4DAnyone introduces a novel framework for reconstructing 4D humans from casual monocular video by overcoming the bounded-attention-context limitations of current video diffusion models. By utilizing Reference Context Packing and Target Context Routing, the system enables high-fidelity 4D Gaussian Splatting without the structural drift common in previous methods.
Researchers have identified that retrieved memories often degrade LLM reasoning, introducing 'cognitive traps' that reduce performance compared to models without memory. The new MemTrapBench framework quantifies these failures and proposes AdaptiveMem as a mitigation strategy.
The open-weights community is rapidly adopting 'abliteration' techniques to strip safety-aligned models of their refusal behaviors. This shift is exemplified by the massive popularity of modified Qwen 3.8-27B variants, signaling a move toward unconstrained model access.
The new WithEveryone framework enables consistent, multi-identity image generation by decoupling layout planning from visual synthesis. By using explicit identity-layout grounding rather than embedding-based matching, it achieves significantly higher fidelity for groups of up to ten people.
The release of abliterated, uncensored variants of the Qwen 3.8-27B model marks a significant shift in how developers access high-performance, refusal-free LLMs. These GGUF-formatted models allow for local execution, bypassing standard alignment constraints through structural weight modification.