7526 papers
This paper addresses the training-inference mismatch in diffusion language models caused by planner-based sampling strategies by deriving a new Planned Evidence Lower Bound (P-ELBO) and introducing Planner Aware Path Learning (PAPL), a simple training modification that aligns training with planned inference to achieve significant performance gains across protein, text, and code generation tasks.
The paper introduces Diffusion Alignment as Variational Expectation-Maximization (DAV), an iterative framework that alternates between test-time search for diverse, reward-aligned samples and model refinement to optimize diffusion models for downstream objectives while mitigating reward over-optimization and mode collapse.
This paper addresses the online minimization of polarization and disagreement in the Friedkin-Johnsen opinion dynamics model under incomplete information by proposing a two-stage low-rank matrix bandit algorithm that achieves a cumulative regret of through subspace estimation and linear bandit optimization.
This paper introduces Self-Speculative Masked Diffusions, a novel generative model for discrete data that leverages causal attention and integrated speculative sampling to produce non-factorized predictions, thereby achieving a roughly 2x reduction in neural network forward passes compared to standard masked diffusion models.
This paper proposes TCR-EML, an explainable-by-design model that integrates prototype layers based on known biochemical binding mechanisms into protein-language model backbones to achieve competitive TCR-pMHC binding prediction accuracy while providing interpretable rationales superior to existing black-box approaches.
This paper demonstrates that while standard decoder-only models underperform compared to encoder-only architectures in cross-modal adaptation for partial differential equations, introducing novel bidirectionality-mimicking techniques like Parallel Flipping and Sequence Doubling effectively closes this performance gap.
This paper challenges the reliability of language model micro-benchmarks by demonstrating that they often fail to consistently rank models with small performance differences, frequently requiring as many as 250 examples to achieve accuracy comparable to random sampling, thereby offering actionable guidance on the trade-off between evaluation efficiency and reliability.
CanvasMAR enhances masked autoregressive video prediction by introducing a global "canvas" prior and a motion-aware curriculum to generate high-fidelity, coherent videos with fewer sampling steps, achieving performance that rivals advanced diffusion-based methods.
This paper demonstrates that injecting external verification into synthetic data retraining can prevent model collapse and yield near-term improvements, though theoretical analysis and experiments across linear regression, VAEs, and LLMs show that long-term performance ultimately converges to the verifier's knowledge center and may plateau or decline if the verifier is imperfect.
This paper proposes a novel, computationally efficient interval-based reachability analysis method for Neural ODEs that leverages continuous-time mixed monotonicity to trade off tightness for scalability, making it particularly suitable for high-dimensional and safety-critical real-time applications.
This paper presents a real-time online framework that utilizes modified sliding-window Hankel Dynamic Mode Decomposition with singular-value hard thresholding and Cadzow projection to denoise partial measurements and construct predictive models for dynamic obstacle motion, enabling stable, variance-aware forecasting suitable for robotic motion planning.
The paper introduces KLASS, a training-free, KL-divergence-guided sampling method that significantly accelerates inference in masked diffusion models by unmasking multiple stable tokens per iteration, achieving state-of-the-art speed and performance across text, image, and molecular generation tasks.
The paper proposes CADM, a cluster-customized adaptive distance metric that dynamically adjusts distance measurements based on the unique attribute distributions within each cluster to improve categorical and mixed data clustering performance.
The paper introduces FireScope, a novel VLM-based framework and accompanying FireScope-Bench dataset that leverage chain-of-thought reasoning to significantly improve the generalization, interpretability, and accuracy of cross-continental wildfire risk prediction by integrating visual, climatic, and geographic factors.
The paper introduces EgoCogNav, a multimodal framework that predicts perceived path uncertainty to jointly forecast egocentric trajectories and head motion, supported by the new Cognition-aware Egocentric Navigation (CEN) dataset to better model human cognitive factors in navigation.
The paper proposes SPINE, a token-selective test-time reinforcement learning framework that improves reasoning model performance by updating only high-entropy decision-critical tokens with entropy-band regularization, thereby preventing response collapse and enhancing stability without requiring external labels or reward models.
The paper introduces DAISI, a scalable data assimilation algorithm that leverages flow-based generative models with a novel inverse-sampling step to integrate forecast information and guide conditional sampling, enabling accurate filtering in complex, nonlinear systems where traditional Gaussian-based methods fail.
This paper introduces Soft Q-based Diffusion Finetuning (SQDF), a novel KL-regularized reinforcement learning method that employs a reparameterized policy gradient of a training-free soft Q-function, enhanced by discount factors, consistency models, and off-policy replay buffers, to effectively align diffusion models with downstream objectives while mitigating reward over-optimization and preserving sample diversity.
This paper proposes a novel training framework that leverages the -divergence family to explicitly filter incorrect answers and control the precision-diversity trade-off, thereby overcoming the diversity loss inherent in standard Reinforcement Learning and achieving state-of-the-art performance on the Lean theorem-proving benchmark.
The paper introduces A-3PO, a method that accelerates asynchronous LLM training by 1.8x by approximating the computationally expensive proximal policy in Decoupled PPO through simple interpolation, thereby eliminating the need for extra forward passes while maintaining comparable performance.