Timeflies jointly infers whether future observations exist and predicts their values, outperforming methods that assume future observation times are known.
REEF proposes relation tokens as graph foundation model units and uses hypernetworks to adapt aggregators and classifiers, outperforming existing methods in pre-training and transfer learning.
DiPO disentangles perplexity into exploration and exploitation subspaces to enable fine-grained trade-offs, improving LLM reasoning and function calling via stable perplexity-guided policy optimization.
xHC expands Transformer hyper-connections beyond four streams via sparse updates and temporal augmentation, improving scaling efficiency. It boosts 18B MoE downstream scores by 4.0 points over mHC with lower compute and reduced memory traffic via xHC-Flash.
MedMemoryBench introduces a streaming benchmark with synthetic long-horizon medical trajectories to evaluate agent memory, revealing severe bottlenecks in reasoning and noise resilience due to memory saturation.
Next Forcing uses multi-chunk prediction to accelerate convergence 2.3x, boost high-frame-rate accuracy 93.1%, and double inference speed for world models.
DiffScore evaluates text with masked diffusion models using bidirectional context to eliminate positional bias and decompose quality into fluency and faithfulness, outperforming autoregressive baselines.
TreeGraft combines small and large drafters with a scheduler to build shared draft trees, boosting speculative decoding by 15.1% over single-drafter methods.
Falcon-X maps heterogeneous time series variates into a unified latent prototype space using diff-attention and latent entity attention to enable cross-variate modeling and zero-shot structural transfer with strong forecasting results.
CausalMix frames data mixture optimization as causal inference to dynamically estimate optimal mixtures via conditional average treatment effects, improving LLM performance without retraining proxy models.