A GUI-Critic for Pre-Operative Error Diagnosis method with Data Collection Pipeline with Reasoning Bootstrapping and Reinforcement Learning

SpecEM is a plug-and-play LLM ensemble that boosts strong models and enables fast, segment-level collaboration. It outperforms prior methods across six tasks and five LLMs.

We propose ETT, an end-to-end tokenizer tuning approach that enables joint optimization between vision tokenization and target autoregressive tasks.

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

Our meta-analysis reveals vision models' underappreciated potential for graph understanding, offering insights for developing more effective graph learning systems based on global structural perception.

A Multi-Task Transfer Learning method via tactic-level and individual skill knowledge.

Transformers fall short of surpassing the performance ceiling of GNNs, but introducing appropriate constraints can effectively enhance the generalization of GNNs.

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

LayerNavigator is a low-overhead method that scores each LLM layer's steerability to guide multi-layer activation steering, significantly outperforming baselines while offering clear interpretability.

We introduce provable and practical watermarking approaches for data poisoning attacks.

We propose GOOD, a training-free framework that leverages off-the-shelf classifiers to guide diffusion models for generating diverse, informative OOD samples—improving outlier exposure without requiring external datasets or embedding alignment.

RayFusion is a superior camera-only collaborative perception framework based on ray occupancy information fusion.

We present LangSplatV2, which achieves high-dimensional feature splatting at 476.2 FPS and 3D open-vocabulary text querying at 384.6 FPS for high-resolution images.

We propose CLAdapter to refine and adapt the rich representations learned from large-scale data for transfer across diverse, data-limited scientific downstream domains.

To mitigate temporal redundancy for SNNs, we propose a module named Mutual Information-based Temporal Redundancy Quantification and Reduction (MI-TRQR), constructing energy-efficient SNNs.

We introduce NeuComBack, a novel benchmark for Neural Compilation, and propose an automatic prompt learning method that significantly improves both correctness and optimization of LLM-generated assembly.

Transformers can learn self-verifying reflection without language, and reinforcement learning enhances performance through shallow statistical patterns.

Backdooring MoE-based LLMs by exploiting inherent vulnerabilities in the expert routing mechanism.

Use the memorization theorem to analyze the power of chain of thought.

A parameter-free regularized Newton-type algorithm that simultaneously achieves optimal global complexity and quadratic local convergence in nonconvex optimization.

A SoTA sequence parallelism for linear attention with a brand new collective communication.

GoRA improves LoRA by adaptively assigning ranks and initializing weights using gradient info, achieving better performance while staying efficient. It outperforms LoRA and full fine-tuning on GLUE (+5.88) and GSM8k (+5.13) benchmarks.

We propose SEA, a simple inference-time alignment method that reformulates alignment as an iterative optimization procedure on an energy function over logits in the continuous space defined by the optimal RLHF policy for deep and effective alignment.

Training LLMs to combine reasoning with external knowledge retrieval via RL without any supervised data on reasoning steps.

A defense method against RAG corpus attacks that combines semantic of document.

We propose a novel framework employs a MoE-based multi-agent world model for learning and planning in multi-task scenarios.

We introduce GUI Exploration Lab, a flexible simulator for GUI agent navigation. Experiments show a staged SFT + RL approach (especially multi-turn RL) significantly boosts navigation and exploration capabilities.

Efficient Long Context Fine-tuning through Dynamic Data Scheduling

We propose Riemannian Consistency Model (RCM) as an extension of the original consistency model for non-Euclidean domains.

Proposed PMQ-VE, a new quantization method for video enhancement that achieves high performance with low-bit models.

A Dynamic and Scalable Reasoning Framework for Solving RPMs

This work proposes SegGraph, a novel propagation method based on SAM segment graphs that leverages 2D foundation model features and geometric relationships for few-shot 3D part segmentation.

This work presents a lattice Boltzmann model, a novel physics-inspired framework for real-time online pixel tracking and object tracking.

This paper presents InstructSAM, a training-free framework that counts, detects, and segments objects in remote sensing imagery based on tailored instructions.

This paper introduces MigGPT, a novel framework that leverages large language models and a code fingerprint structure to automate the migration and maintenance of out-of-tree Linux kernel patches.

We propose a sparse autoencoder that maps the semantics of vision and language representations into a unified concept set.

We present SolidGeo, the first benchmark focused on solid geometry, revealing the poor performance of current MLMs on solid geometry and analyzing the inference flaws of current models.

We present IR-OptSet, a public LLVM IR dataset tailored for optimization-sensitive LLM training, significantly improving compiler code generation performance.

More accurate RNA inverse design model

A versatile data mixture ratio optimization framework for LLM training that enjoy both theoretical and practical advantages.

This work achieves proactive intervention against LLM harmful content through real-time monitoring of streaming and individual tokens.

Give VLM the ability to call tools to retrieve and filter information

We introduce a noval framework for red-teaming black-box T2I systems, termed Rule-based Preference modeling Guided Red-Teaming (RPG-RT).

We propose AlphaDecay, a per-module weight decay method guided by heavy-tailedness, improving large language model performance.

REArtGS

This study introduces VADB, the largest video aesthetic database with 10,490 videos, and VADB-Net, a novel framework that outperforms existing models in video aesthetic assessment.

A novel algorithm that estimates fine-grained, token-level advantages in reinforcement learning without introducing additional models.

We propose Safe-Sora, the first framework that integrates graphical watermarks directly into the video generation process.

This work improves CLIP’s visual detail capturing ability by inverting the unCLIP generative model, which we find suitable for achieving this goal.

A novel approach; Very practical

This work presents DAIL, a semantic aligned reinforcement learning method that improves task discrimination and generalization in language-conditioned problems.

We provides the first systematic analysis of large reasoning model as translation evaluator and offer insights in improvements.

We propose LM-EEC, a SAM2-based dual-memory segmentation framework that improves egocentric-exocentric object correspondence by adaptive fusion and efficient memory compression.

MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition

We propose a Human Motion-Vision-Language Model (HMVLM) based on MoE LoRA for diverse human-centric downstream tasks. A novel "zero expert" is introduced to mitigate catastrophic forgetting during the instruction-tuning.

Exploring and Mitigating hallucination in LMMs towards scene text spotting and understanding

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks.

ProteinConformers is a novel large benchmark dataset enabling better evaluation of protein structure generators by capturing realistic conformational diversity and plausibility and proposing new assessment metrics.