A spike activity pruning method for spiking neural network

This paper introduces Diffusion-Classifier Synergy (DCS), a novel framework where a diffusion model and an FSCIL classifier co-evolve through a dynamic, classifier-derived reward system to achieve state-of-the-art few-shot class-incremental learning.

We present a method to generate high-quality 3D shape composed of individual and complete parts from a single-view image.

We propose EfficientNav, enabling on-device memory-augmented object-goal navigation system for zero-shot in-door navigation.

A multimodal framework that leverages OpenStreetMap data to enhance remote sensing foundation models in both self-supervised pretraining and downstream tasks.

We propose a novel joint Velocity-Growth Flow Matching (VGFM) approach, driven by semi-relaxed optimal transport, for single-cell dynamics modeling.

We propose Reward Reasoning Models, which leverage additional test-time compute for complex queries where appropriate rewards are not immediately apparent.

We propose and analyze a new class of unbalanced weak optimal transport (OT) problems with total variation penalties.

An effective parameter-efficient model merging method for multimodal large language models from the perspective of direction robustness in low-rank space

SATURN enables scalable, verifiable, and curriculum-controlled reinforcement learning that enhances the reasoning capability of LLMs, and generalizes to math and programming tasks.

We propose SEEK, a novel watermarking method for large language models that simultaneously improves robustness against both scrubbing and spoofing attacks, achieving a Pareto-optimal balance superior to existing approaches.

We ddvance vision-language action models via introducing comprehensive knowledge prediction.

This paper presents a lightweight MCMC-based approach for constrained decoding in LLMs

This paper introduces DoctorRAG, a medical AI that mimics doctor-like reasoning by combining textbook knowledge with insights from similar patient cases, using "textual gradients" to refine its answers for improved accuracy and relevance.

We provide a computational efficient algorithm to achieve $O(H)$ deployment cost with polynomial sample complexity.

We propose CryptoMoE, the first framework to enable private, accurate, and efficient inference for MoE-based LLMs.

We provide the first theoretical sample-efficiency results of conditional diffusion models via transfer learning

We propose EAGLE-3, observing that it can benefit from scaling up the data.

We introduce MMOral, the first large-scale multimodal instruction dataset and benchmark tailored for panoramic X-ray interpretation. MMOral-Bench is a comprehensive evaluation suite covering five key diagnostic dimensions in dentistry.

This work investigates whether MLLMs can be used as unified evaluator for AI-generated videos and proposes a benchmark to assess this ability.

DrivAerStar introduces a validated dataset of 12,000 industrial-grade CFD simulations for automotive aerodynamics, enabling machine learning models to achieve engineering accuracy while drastically reducing computational costs.

We embed personalized LLM Watermarks responsively into generated text with Sparse Autoencoders, with only inference-time sampling on black-box access LLMs and no extra training, and achieved high accuracy while preserving text quality.

We compare RL methods DPO & GRPO for image generation, showing their strengths & how rewards affect generalization. Explores scaling for better CoT-based synthesis.

This paper presents a dynamic framework that enables reliable execution under violated conditional independence assumptions.

KARMA is a multi-agent LLM framework that automatically update and expand knowledge graphs from scientific papers. By using nine specialized agents, KARMA successfully add new entities with high verified accuracy and reduce inconsistencies.

A training-free style-aligned image generation method for Flux.

The paper provides a novel approach to ICL example selection/ordering through the extension of DPO.

How can we leverage the existing LoRA weights to adapt to the latest model version with less effort?

MEgoHand is the starting point for generating high-quality motion sequences of hand-object interactions, conditioned on egocentric RGB images, textual instructions, and given initial MANO hand parameters.

A novel alignment framework that integrates generative reward models with multi-modal RLHF.

We propose a training-free visual token pruning method CDPruner for MLLM inference acceleration by maximizing the conditional diversity of retained tokens.

A foundation model specifically designed for medical time series forecasting

We introduce a Stochastic Interpolants-based data assimilation framework for stochastic dynamic systems.

We present 3D Gaussian Splatting to Event Generation (GS2E), a large-scale event synthetic dataset designed for high-fidelity event vision tasks from real-world sparse multi-view RGB images.

We propose Pseudo-Zeroth-Order (PseuZO), a framework that estimates Jacobian matrices via model output differentiation and applies EMA for variance reduction, with theoretical convergence guarantees and empirical validation.

We rethink LLMs from the perspective of recommender systems and propose Language System, which exploits the model’s output distribution more efficiently and effectively.

We develop a new framework that combines autoregressive modeling and diffusion for video data.

Propose a fault-tolerant algorithm with minimal memory and computation overhead.

This work improves Monte Carlo Tree Search for symbolic regression through an extreme bandit strategy and evolution-inspired state-jumping actions.

We address the multiscale singularity in Euclidean diffusion models for manifold-constrained data and propose two novel methods to improve sampling accuracy on complex geometries.

We propose Fuz-RL, a fuzzy logic-based framework that enhances robustness of safe reinforcement learning by measuring and balancing uncertainties in a non-additive way, achieving better safety and performance under uncertainty.

Learning to forecast in the transformed domain improves forecasting performance.

We propose Taccel, a high-performance GPU-based physical simulator for robots with vision-based tactile sensor.

QOM uses quantized autoencoders to model opponent types, updates beliefs in real-time via Bayesian inference, and enables efficient belief-aware planning.

Var-RUOT is a new framework improving RUOT for recovering dynamics from high-dimensional snapshots. It incorporates first-order optimality conditions, achieving better performance on single-cell datasets.

We introduce the Martingale Score, an unsupervised metric from Bayesian statistics, to show that reasoning in LLMs often leads to belief entrenchment rather than truth-seeking, and shows this score predicts ground-truth accuracy.

A novel structural generation-augmented retrieval method for motion-text retrieval.

CoT-Diff tightly couples MLLM reasoning with diffusion to enable step-by-step 3D layout planning and dynamic spatial control within a single generation process.

Learning the individualized treatment effects under the assumption of rank preservation.

This paper presents IPAD, a novel AI detection framework that identifies predicted prompts that could have generated the input text. IPAD outperforms the baselines by 9.73% on in-distribution data and 12.65% on OOD data.

PartCrafter is a structured 3D generative model that jointly generates multiple parts and objects from a single RGB image in one shot.

To tackle the complex yet highly practical challenges, we propose a Theory-Driven Label-Specific Representation (TDLSR) framework.

The first global synthetic dataset for physics-ML seismic wavefield modeling and full-waveform inversion

We propose an explainable and extendable framework to enhance deepfake detection via multimodal large-language models.

We propose AutoToM, an automated agent modeling method for scalable, robust, and interpretable mental inference.

We present a dataset of 210,000 triplets (content, style, stylized images) and an end - to - end stylization framework tailored for it, ensuring efficient style transfer.

We propose Magical, an asymmetric LoRA architecture to meet the requirement for semantic fidelity and diverse lay-style generation in medical lay language generation task.

We propose a noval fine tuning framework that uses reinforcement learning and adapter compression to selectively shield model layers in TEEs, which achieves strong model protection with 2–3× efficiency gains and no accuracy loss.

URDF-Anything is an end-to-end MLLM-based framework that jointly optimizes geometric segmentation and kinematic prediction for articulated object reconstruction.

HFTP aligns syntactic processing representations in both large language models (LLMs) and the human brain

Safe RLHF-V, the multimodal safety alignment framework.

We design orthogonalization to eliminate redundancy between different-order Common Neighbors and normalization to mitigate over-smoothing for link prediction..

We introduce SyncHuman, a full-body human reconstruction model using synchronized 2D and 3D generative model. Given a single image of a clothed person, our method generates detailed geometry and lifelike 3D human appearances across diverse poses.

We propose Fira, a plug-and-play memory-efficient training framework of LLMs to enable full-rank training consistently under the low-rank constraint.

This paper tackles the problem of zero-shot inference on text-attributed graphs, and proposes a novel method that uses bundles to query LLMs and supervise GNNs.

Spiking Neural Networks for Neural Radiance Fields

This paper propose a large-scale, high-quality editing dataset, accompanied by a comprehensive benchmark, an advanced editing model, and an effective edit evaluator.

We introduce PHYBench, an original and challenging benchmark for evaluating reasoning capabilities of large language models (LLMs) by leveraging physics problems.

The First Ultra-Scale Dataset and Benchmark for Subject-to-Video Generation

HawkBench is a human-labeled, multi-domain benchmark with 1,600 samples for evaluating RAG systems on diverse queries, revealing limits in generalizability and the need for adaptive strategies.

We introduce Anywhere3D-Bench, a holistic 3D visual grounding benchmark consisting of 2.8K referring expression-3D bounding box pairs spanning four different grounding levels: area, space, object, and part.

We present CellVerse, a unified language-centric single-cell analysis benchmark within the realm of LLMs for cell biology.

MedAgentBoard finds multi-agent useful for some medical tasks, but single LLMs or conventional methods are often better, especially for VQA/EHR. Task-specific AI choice is key.

A progressive instruction evolution framework that iteratively unlock LVLMs' reasoning ability

A storage-efficient solution for online Free-viewpoint video recosntruction using 3D Gaussian Splatting.

We unify more than 10 existing one-step diffusion distillation approaches, with a new SoTA 1-step generation on CIFAR10 and ImageNet64x64 benchmarks.

Structure-first, training-free premise selection for Lean4: CSE trees with WL screening and TED reranking that beat semantic baselines on Mathlib4.

We propose Uni-MuMER, full fine-tuning a VLM for handwritten math expression recognition via structured spatial reasoning (Tree-CoT), error-driven learning (EDL), and symbol counting (SC), achieving state-of-the-art results.

We propose an algorithm to recover the correct generation order from textual data and solve reasoning and planning problems adaptively without order annotations.

We propose a progressive consistency distillation framework that enhances the efficiency of MLLMs by significantly reducing computational cost while preserving strong performance.

This study introduces a novel causality-driven robust optimization approach that selectively updates model components sensitive to causal reasoning, enhancing model causality while preserving valuable pretrained knowledge to mitigate overfitting.

Masked Diffusion Models can generate low-perplexity text efficiently, but can not handle tasks requiring high accuracy efficiently.

We analyze optimal budget allocation strategies for iterative bootstrapping with synthetic data to maximize model performance.

A novel framework for spike camera reconstruction

We introduce Risk-aware Direct Preference Optimization (Ra-DPO), a novel approach that incorporates risk-awareness by employing a class of nested risk measures.

GRACE achieves significant gains on performance and efficiency in prompt optimization by strategically introducing information loss through a dual mechanism of refinement and compression.

We propose Q-Insight, a reinforcement learning-based IQA approach using limited annotations, achieving superior performance and enhanced visual reasoning capabilities.

SBSN is a scalable framework for link sign prediction in signed bipartite graphs, combining efficient subgraph sampling, directed message passing, and node feature distillation to achieve high accuracy and fast inference with low memory cost.

This paper proposes a novel framework that generates better contrastive pairs for contrastive learning by integrating LLM-based semantic retrieval with a learnable sample synthesizer.

ChemCoTBench bridges complex chemical reasoning with arithmetic-inspired step-by-step workflows, enabling LLMs to systematically tackle real-world tasks like molecular optimization and reaction prediction.

A Reasoning Benchmark on Visual Perception

A novel real-world benchmark for long dependency understanding.

A human preference dataset for multi-turn interleaved multimodal understanding and generatin tasks

ViSpec accelerates vision-language model inference by integrating vision-aware speculative decoding with compressed image tokens and global feature injection, achieving up to 3.22× speedup.

We propose a co-evolving reinforcement learning method that jointly optimizes the coder and unit tester without relying on ground-truth code supervision.

We propose UnCLe, a deep learning framework that uses temporal perturbations to uncover dynamic, time-varying causal graphs from observational time series.

We propose Spatial Adversarial Alignment (SAA), which leverages a witness model to fine-tune the surrogate model via spatial-aware and adversarial-aware alignment to generate highly transferable adversarial examples.

We propose the infinite-Gaussian process model, enabling efficient Bayesian optimization under complex, non-stationary, and heavy-tailed reward landscapes.

We introduce the concept of semantic orientation, representing the object orientation condition on open vocabulary language.

A novel approach; Very practical

We introduce a Functional Scaling Law that predicts full SGD loss dynamics under arbitrary learning rate schedules.

An efficient data generation framework for dexterous manipulation.

We introduce CytoBridge, a new deep learning approach that models unbalanced stochastic cellular dynamics and interactions from snapshot data.

Long-term feed-forward 3D point tracking in persistent 3D point maps.

Design-Based Bandits Under Network Interference: Trade-Off Between Regret and Statistical Inference

We present a unified audio-visual framework to uncover how humans and AI respond to modality conflicts and bias.

The paper designs a training objective to suppress the learning of redundant features in time series models.

An efficient PbRL method that mitigates overfitting and overestimation via dual regularization, enhancing feedback efficiency in both online and offline settings

OLinear is a linear-based time series forecasting model that outperforms state-of-the-art Transformer-based methods with significantly higher efficiency.

We present a novel approach that addresses the low-rank bottleneck in LoRA by integrating nonlinear mappings with compressed rank, achieving an optimal balance between parameter efficiency and model performance.

A choice-theoretic loss that distills knowledge by teaching the student the teacher’s class ranking under the Plackett–Luce model, keeping the true label as the top choice and unifying distillation without an additional cross-entropy term.

We propose Adaptive Classifier-Free Guidance (A-CFG), which dynamically re-masks low-confidence tokens for a more targeted unconditional input in iterative models.