All papers

Introduces SWE-DeceptionBench, a benchmark measuring active and passive deceptive model behaviors across software development lifecycle tasks.

Proposes a model-agnostic method to integrate biological domain knowledge into machine learning models trained on weakly annotated bioimaging.

Presents a constraint acquisition tool that synthesizes size-independent constraint programs from input data and example solutions.

Introduces SelectArena, a benchmark designed to evaluate tabular feature selection methods across real-world domain datasets and contemporary models.

Evaluates how pretraining and test-time training computational budgets affect ARC benchmark performance across two distinct reasoning model architectures.

Proposes a dynamic curriculum for GRPO that re-estimates sample difficulty to progressively transition policy training toward harder problems.

Presents Outlier-Aware Attention Linearization, using quadratic approximations to retain outlier dimensions during efficient Transformer attention evaluation.

Introduces Loopless Variance-Reduced Policy Gradient, proving near-optimal sample complexity under weak gradient domination using first-order updates.

Employs a bilevel Stackelberg formulation to jointly optimize classification accuracy and confidence calibration without separate calibration datasets.

Combines load-adaptive hierarchical key-value caching with speculative decoding to accelerate long-context language model serving.

Presents a graph neural process framework to answer counterfactual queries directly from observational data without prior causal graph discovery.

Proposes an attention mechanism that switches from linear complexity during prefilling to full attention during token decoding.

Examines error localization and credit assignment failures in critique-conditioned on-policy self-distillation for mathematical reasoning.

Introduces a latent sociability prior to mitigate edge-independence issues in discrete graph diffusion models.

Proposes a calibration-free post-training quantization method that jointly optimizes weight and activation scaling for 4-bit model inference.

Replaces frequent state memory updates in state-space models by caching inputs and periodically checkpointing state during decoding.

Introduces a Ukrainian meme dataset and evaluates vision-language models on detecting persuasive and propaganda techniques.

Demonstrates that discrete graph diffusion models exhibit conditional memorization when reverse trajectories are conditioned on corrupted training graphs.

Describes an industrial framework for evaluating, ablating, and deprecating features in production recommender systems without performance regressions.

Measures output variability across repeated runs at temperature zero in semantic query engines backed by language models.

Trains vision-language models to perform active spatial reasoning through novel-view generation and simulated mentalization trajectories.

Applies vision-language models and structured annotation to parse and analyze historical comic books and manga at scale.

Investigates how changing tasks redistributes attention head influence and alters internal activations in open-weight large language models.

Evaluates regression objective functions and global early stopping limits across neural network backbones and established tabular datasets.

Introduces a model-agnostic, concept-level attribution method for black-box vision models by masking concepts located with vision-language models.