All papers

Proposes reflected proximal dynamics to combine exact target preservation with momentum-based acceleration in log-concave sampling.

Measures teacher entity familiarity to filter out uninformative search trajectories when distilling deep-search capabilities into smaller agents.

Presents an activation quantization method that groups tokens via pre-scan features to improve low-bit post-training quantization for Mamba super-resolution models.

Surveys the use of flow-matching generative modeling approaches within reinforcement learning across policy parameterization and foundation model alignment.

Extracts reusable local control signals from frontier model traces to guide smaller language model agents without extra target-model training.

Introduces an INT8 integer-only training algorithm using composable building blocks that removes floating-point arithmetic across all training operations.

Proposes an attention-free language model architecture separating drafting from multi-scale prefix revision to bypass sequence-level quadratic attention costs.

Develops a reasoning-based attack that pairs malicious actors with benign contexts to guide language models along harmful multi-step reasoning trajectories.

Presents a multi-agent framework that allocates test-time computation only to unresolved steps for video spatial reasoning tasks.

Combines next-token prediction transformers with deep clustering to discover interpretable behavioral patterns in open-ended game trajectories.

Analyzes activation steering interventions in language models by isolating the effects of covariance filtering from target gain and displacement.

Presents a breast cancer risk prediction framework that explicitly aligns longitudinal examinations across multiple mammography views.

Introduces a training-free Bayesian state-update framework to reduce error accumulation and representational drift in streaming 3D reconstruction models.

Proposes a world-centric rotary positional embedding using 3D Gaussian primitives to improve cross-view attention in multi-view vision tasks.

Introduces an adaptive multi-agent debate framework that uses a leader-supporter structure to trigger debate only when simple aggregation is insufficient.

Presents an architecture separating natural language interpretation from numerical time series foundation models for instruction-driven forecasting.

Introduces a margin-based analysis metric to evaluate class boundary representations in forward-forward learning algorithms.

Applies a supervised variational autoencoder to brain connectome data for sex classification in young adults.

Introduces internal predictive registry slots into diffusion transformers to explicitly track physical dynamics in world models.

Presents a loop-aware post-training quantization method to compress looped language models without severe accuracy loss.

Proposes a multi-route retrosynthesis planner that evaluates synthesizability and cost across chemical reaction options.

Applies a causal inference framework to evaluate the impact of robotic social navigation signals in corridor interactions.

Presents a loop-style parameter-sharing scaling strategy to improve recommendation rankers without increasing parameter counts.

Proposes an adaptive retrieval-augmented generation framework that dynamically expands clinical queries for medication recommendation.

Introduces a world-action model framework using predictive rehearsal and action perturbations to guide robotic policy learning.

Proposes a structured zeroth-order optimization framework using subspace estimation and spectral control for memory-efficient model fine-tuning.

Evaluates a two-stage domain adaptation method for multilingual patent translation and analyzes failure modes in low-resource languages.

Develops an explainability-guided Fourier domain generalization framework to improve sim-to-real reinforcement learning for autonomous driving.

Introduces a benchmark and baseline model for evaluating crystal structure determination from powder X-ray diffraction data.

Introduces a calibration-free latent merging framework to compress Mixture-of-Experts models and reduce memory overhead.

Proposes structural growth pathways to preserve pretrained weights during reinforcement learning and enable transfer across language models.

Presents speculative recursive self-improvement to draft successor agents before full evaluation and optimize verification budgets.

Analyzes how persona construction in generative social simulations causes network segregation and collective political polarization.

Introduces a physics-informed latent variable model combining dimensional analysis and differentiable physics heads for data-scarce multi-physics systems.

Uses analytic hierarchy processes and active learning to rank exploratory data patterns according to individual user preferences efficiently.

Co-evolves the user simulator, environment tools, and rewards during training to improve interactive multi-turn language model agents.

Monitors model explanation drift on evolving graphs to provide early warnings before predictive uncertainty guarantees fail.

Distills multi-agent discussions into a single student model by grouping redundant reasoning traces and revisiting unlearned examples.

Accelerates video diffusion models by using closed-loop calibrated caching instead of fixed heuristic thresholds across denoising steps.

Applies rate-distortion theory to structure ordered visual codes so text captions form the initial low-rate representations for images.

Uses a learned scene-coordinate localization module to maintain consistent 3D spatial context for image-goal robot navigation.

Improves reinforcement learning for reasoning models by sampling continuations from truncated traces to provide more accurate intermediate rewards.

Presents a deep learning framework that integrates heterogeneous surface EEG and intracranial EEG signals for multimodal seizure classification.

Classifies languages at the token level across 111 language-script combinations in mixed-language and code-switched texts.

Dynamically balances relevance and visual diversity per sample to compress visual tokens in vision-language models more effectively.

Proposes a generative adversarial network guided by differentiable segmentation models to edit image colors realistically and with precise control.

Generates prompt-adapted adversarial images that block unauthorized visual questions on vision-language models while still answering legitimate prompts correctly.

Evaluates different agent control granularities for post-route FPGA timing closure to balance execution reliability and frequency gains.

Incorporates human understandability scores directly into feature selection to make tabular machine learning explanations easier for users to interpret.

Combines physical models and state space neural networks in a closed loop to forecast long-term chemically reactive air quality.

Introduces a general visual memory planner that retrieves, creates, and updates visual evidence across shots for long-horizon video generation.

Improves long-context handling in Mamba state space models without retraining by using surprisal scores to guide state routing.

Reconstructs divergence-free 3D fluid velocity fields from sparse, irregular observations in a single forward pass using a set-to-field transformer.

Detects adversarial nodes on evolving graphs without fraud labels by measuring the consistency between node features and aggregated neighborhood contexts.

The paper analyzes the impact of teacher rationales before distilling medical reasoning capabilities into compact language models.

The paper introduces Trace-to-Architecture, a framework that diagnoses LLM agent execution failures to automatically revise agent harness components.

The paper introduces an architecture that incorporates token queue dynamics into wireless communications for foundation model inference services.

The paper proposes VAttnPO, using teacher visual attention to refine token-level policy optimization for multimodal reinforcement learning.

The paper analyzes pre-sink hidden-state geometry to detect and modulate attention sink formation in language models.

The paper proposes evaluating partial generation outputs to stop harmful model interactions early while avoiding unnecessary prompt rejections.