π― What it does: This paper proposes a proximal operator-based generative model framework called PGM, which uses the Moreau score instead of the traditional time-dependent posterior score to achieve efficient sampling for Bayesian inverse problems.
Lin Zheng (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
CodeCompressionComputational EfficiencyRepresentation LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextSequential
π― What it does: Propose a Proxy Compression training scheme, which mixes the compressed sequences generated by an external compressor with the original UTF-8 byte sequences during training, allowing a single language model to learn both representations simultaneously; during inference, only the original byte sequences are used, completely eliminating the dependence on the compressor.
PRPO: Paragraph-level Policy Optimization for Vision-Language Deepfake Detection
Tuan Nguyen (Qatar Computing Research Institute, Hamad Bin Khalifa University), Issa Khalil (Qatar Computing Research Institute, Hamad Bin Khalifa University)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
π― What it does: Proposed a multimodal large language model based on paragraph-level reinforcement learning, which can provide interpretable and visually consistent reasoning and decision-making when detecting deepfake images.
PSBench: Editing Image via GUI Agents in Photoshop
Yinuo Zhang (Harbin Institute of Technology), Dianbo Sui (Harbin Institute of Technology)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringVision-Language-Action ModelImageTextMultimodalityBenchmark
π― What it does: Proposed the PSBench benchmark for evaluating GUI agents in image editing tasks within Adobe Photoshop, containing 600 manually annotated tasks covering core functions such as layer operations, canvas manipulation, and filters, and providing fine-grained evaluation metrics and the Non-Destructive Editing Consistency (NDEC) metric.
π― What it does: The paper proposes a point cloud data augmentation framework called PSMix based on spherical harmonic transformation, which can perform rotation-aware hierarchical mixing in the frequency domain to generate structurally consistent and diverse augmented samples.
π― What it does: Propose the PULSE framework, which utilizes physics-driven phase-anchored decomposition and generative phase routing to address the phase forgetting problem under the non-stationarity of time series.
π― What it does: For diffusion models with extremely low number of steps, a method called Quantization-Aware Scheduling (Q-Sched) is proposed, which significantly improves the image quality of the compressed model by introducing learnable pre-adjustment coefficients in the sampler, while keeping the model weight quantization unchanged.
Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization
Haocheng Xi (University of California, Berkeley), Kurt Keutzer (University of California, Berkeley)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelContrastive LearningVideo
π― What it does: Proposes Quant VideoGen (QVG), an untrained KV-Cache quantization framework that significantly reduces the memory consumption of autoregressive video diffusion models while maintaining video quality;
Magda Dubois (UK AI Security Institute), Cozmin Ududec (UK AI Security Institute)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextBenchmark
π― What it does: Propose a statistical framework based on Bayesian generalized linear models to quantify and correct various systematic biases in LLM evaluators during assessment tasks (such as self-bias, length bias, writing style bias, etc.)
Quantifying Frontier LLM Capabilities for Container Sandbox Escape
Rahul Marchand (University of Oxford), Harry Coppock (UK AI Security Institute)
CodeSafty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes the SANDBOXESCAPEBENCH benchmark for security evaluation of LLMs' escape capabilities within container sandboxes.
Quantile-Free Uncertainty Quantification in Graph Neural Networks
Soyoung Park (Chungnam National University), Sungsu Lim (Chungnam National University)
CodeGraph Neural NetworkGraph
π― What it does: Proposed a graph neural network framework called QpiGNN, which does not require quantile inputs, to achieve confidence interval estimation for node regression tasks;
Quantum Robust Inner Minimization for Reinforcement Learning with Quadratic Speed-Up in Query Complexity
Hyun Kyu Lee (Ulsan National Institute of Science and Technology), Sung Whan Yoon (Ulsan National Institute of Science and Technology)
CodeReinforcement LearningTabularTime Series
π― What it does: Developed and evaluated a quantum-accelerated robust reinforcement learning method, QRIM, which achieves fast search over the minimization (worst-case) within the uncertainty set through quantum access to the environment, thereby reducing the query complexity from O(|U|) to O(β|U|) while keeping the external RL pipeline unchanged.
π― What it does: Propose a query-driven asymmetric time-frequency Transformer framework, TF-Restormer, for speech restoration (including noise reduction, echo removal, bandwidth expansion, etc.) under mismatched input and output sampling rates.
R$^3$L: Reasoning 3D Layouts from Relative Spatial Relations
Zhifeng Gu (Hong Kong Polytechnic University), Bing WANG
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringDiffusion modelScore-based ModelTextMultimodalityPoint CloudMeshRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Leverage multi-modal large language models (MLLM) to reason about relative spatial relationships in 3D layouts and convert them into executable layouts; improve the reliability and consistency of reasoning through three major techniques; further optimize the layout to ensure physical feasibility.
RACER: Risk-Aware Calibrated Efficient Routing for Large Language Models
Sai Hao (Southern University of Science and Technology), Bingyi Jing (Chinese University of Hong Kong Shenzhen)
CodeFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose a post-training, risk-aware, calibration-efficient routing method called RACER, which transforms traditional single-model routing into confidence-based ensemble model routing, and improves the final answer quality through aggregation.
Randomized Advantage Transformation (RAT): Computing Natural Policy Gradients via Direct Backpropagation
Mingfei Sun (University of Manchester)
CodeOptimizationReinforcement LearningContrastive LearningTabularTime Series
π― What it does: Propose Randomized Advantage Transformation (RAT), which utilizes the Woodbury formula to transform the Tikhonov-regularized natural policy gradient into an advantage function transformation, and directly estimates this gradient via backpropagation on mini-batch samples using the stochastic block Kaczmarz iteration.
Rationality Measurement and Theory for Reinforcement Learning Agents
Kejiang Qian (University of Edinburgh), Fengxiang He (University of Edinburgh)
CodeReinforcement Learning
π― What it does: Proposes a rationality measure and theoretical framework for reinforcement learning agents, defining perfectly rational actions and quantifying rationality risk.
RBCBF: Decoding Time Safety Alignment via Risk Guided Rollback and Barrier Control
Tianxiang Chen (National University of Singapore), Kaidi Yang (National University of Singapore)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelReinforcement LearningText
π― What it does: Proposes a rollback-based inference-time safe alignment framework called RBCBF, which utilizes risk aggregation and control barrier functions to locate and perform distribution-level correction on the generation process;
CodeComputational EfficiencyKnowledge DistillationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsAuto EncoderContrastive LearningTextRetrieval-Augmented Generation
π― What it does: Proposed RealtimeTool, a real-time LLM function calling framework that compresses low-entropy information through special tokens and achieves parallel decoding of function names and parameters.
Reasoning about Reasoning: BAPO Bounds on Chain-of-Thought Token Complexity in LLMs
Kiran Tomlinson (Microsoft Research), Jennifer Neville (Microsoft Research)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextGraphChain-of-Thought
CodeFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsContrastive LearningTextChain-of-Thought
π― What it does: The study investigates the differences in reasoning between abstract forms (FL) and natural language forms (NL) of the same logical task in large language models, and proposes an abstract alignment intervention achieved through translation training;
RECAST: Model Reconstruction via Counterfactual-Aware Wasserstein Geometry under Limited Data
Xuan Zhao (Forschungszentrum JΓΌlich), Ira Assent (Forschungszentrum JΓΌlich)
CodeClassificationExplainability and InterpretabilityData-Centric LearningContrastive LearningTabularBenchmark
π― What it does: Propose a black-box model reconstruction method called RECAST under limited data and one-time counterfactual explanations, which constructs Wasserstein centroid prototypes to approximate class distributions and generate approximate surrogates.
Benjamin Howard (Institute for Defense Analyses, Center for Communications Research), Keith Frankston (Institute for Defense Analyses, Center for Communications Research)
CodeOptimizationConvolutional Neural NetworkTransformerReinforcement LearningMixture of ExpertsTabularSequential
π― What it does: A recursive AlphaZero-style MCTS algorithm called RMCTS was studied, which achieves high parallelism and significant speed improvement during single-threaded search by recursively computing and optimizing the posterior policy for each node.
RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates
Ali Asad (Queen's University), Xiaodan Zhu (Queen's University)
CodeSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This work proposes the RedDebate framework, which utilizes multi-agent LLMs to engage in debates within red team adversarial scenarios, and achieves fully automated security discovery and improvement through an evaluator and feedback generator; the framework integrates multiple memory modules to support continuous learning and defense.
RedVisor: Reasoning-Aware Prompt Injection Defense via Zero-Copy KV Cache Reuse
Mingrui Liu (Nanyang Technological University), Kwok-Yan Lam (Nanyang Technological University)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose a unified framework called RedVisor, which utilizes a lightweight top adapter and two-stage inference to achieve both detection and prevention of Prompt Injection in LLMs, while retaining the original model's performance during the generation phase and offering strong interpretability.
π― What it does: A framework named RTC is proposed, which utilizes Sinkhorn divergence to perform distributed learning of Meta-values, and achieves task optimization rebalancing through importance weighting via a recursive Bellman error model.
Reflective Hamiltonian Monte Carlo: Mixing Analysis and Application to Sampling on Stiefel Manifold
Kwangmin Lee (Chonnam National University), Sewon Park (Sookmyung Women's University)
CodeOptimizationComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackScore-based ModelContrastive LearningGraphTabularTime SeriesSequentialPhysics RelatedStochastic Differential Equation
π― What it does: Proposed and theoretically proved the non-asymptotic mixing time of the reflective Hamiltonian Monte Carlo (ReHMC) algorithm on bounded domains, and constructed a 'convex container plus sampling' framework to handle arbitrary bounded supports, which was subsequently applied to sampling on the Steifel manifold.
π― What it does: Propose the Reflex framework to achieve streaming inference to support real-time Vision-Language-Action (VLA) control; by partitioning attention context, AdaRMSNorm normalization, asynchronous pipeline, and future state prediction, achieve 50 Hz continuous inference;
Regret-Based Federated Causal Discovery with Unknown Interventions
Federico Baldo (Sorbonne Universite), Charles K. Assaad (Sorbonne Universite)
CodeFederated LearningSafty and PrivacyExplainability and InterpretabilityDiffusion modelContrastive LearningGraphTabular
π― What it does: In a federated learning environment, an algorithm named I-PERI for causal structure learning is designed, which can recover the causal graph on the server side even when unknown interventions exist on the clients and the intervention targets are not shared.
Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Training
Song Lai (Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences), Fei Zhu (Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Compared the forgetting performance of supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RFT) in continual training, and proposed the rollout-based instance filtering algorithm RIF-RFT;
Reinforcement Learning for Reachability: Guaranteeing Asymptotic Optimality
Amogh Palasamudram (Georgia Institute of Technology), Krishnendu Chatterjee (Institute of Science and Technology)
CodeReinforcement LearningTabularBenchmark
π― What it does: Propose a phased model-based reinforcement learning algorithm specifically designed for reachability goals in unknown MDPs, achieving asymptotically optimal policies by progressively estimating the minimum transition probability and using PAC learning.
Reinforcing Real-world Service Agents: Balancing Utility and Cost in Task-oriented Dialogue
Ning Gao (Meituan), Chaozheng Wang (Meituan)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringGenerative Adversarial NetworkTextRetrieval-Augmented Generation
π― What it does: Proposed the InteractCS-RL framework, which redefines task-oriented dialogue as a multi-granularity reinforcement learning process, and achieves high-fidelity online training through a user-centric interaction framework.
π― What it does: Proposes a solution to the class imbalance problem in entity classification for relational databases, designing a relation-based minority class synthetic oversampling GNN (Rel-MOSS), while introducing a relational gate controller and a relation-guided minority synthesizer;
Relational In-Context Learning via Synthetic Pre-training with Structural Prior
Yanbo Wang (Peking University), Muhan Zhang (Peking University)
CodeClassificationData SynthesisComputational EfficiencyMeta LearningTransformerPrompt EngineeringMixture of ExpertsGenerative Adversarial NetworkContrastive LearningTabularBenchmark
π― What it does: Built a relation database foundation model RDBβPFN pre-trained only on synthetic data, and achieved zero-gradient context learning.
Relative Entropy Estimation in Function Space: Theory and Applications to Trajectory Inference
CHAO WANG, Pietro Michiardi (EURECOM)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTabularTime SeriesBiomedical DataBenchmarkStochastic Differential Equation
π― What it does: Proposed a general framework for estimating KL divergence in function space, and applied it to the evaluation of trajectory inference
π― What it does: Propose a lightweight candidate retrieval method (RBE) based on query/item relevance vectors, which approximates any complex similarity model by representing each query or item as a relevance vector with respect to a set of supporting items/queries, and directly using these representations in candidate retrieval.
Haobin Li (Sichuan University), Xi Peng (Sichuan University)
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose a reliable image thinking framework, RTWI, which eliminates noise thinking in multi-modal chain-of-thought reasoning by utilizing text-centric reliability estimation, two-stage filtering, and reliable voting.
RELO: Reinforcement Learning to Localize for Visual Object Tracking
Xin Chen (City University of Hong Kong), Kede Ma (City University of Hong Kong)
CodeObject TrackingReinforcement Learning from Human FeedbackRecurrent Neural NetworkTransformerReinforcement LearningAuto EncoderContrastive LearningOptical FlowImageVideo
π― What it does: This paper proposes a visual object tracking method called RELO based on reinforcement learning, which directly models target localization as a Markov decision process on feature maps, replacing traditional handcrafted spatial priors (center heatmap or corner distribution), and achieves localization through reward-driven policy learning;
CodeComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningMixture of ExpertsText
π― What it does: Proposes the ReMoE framework, which fine-tunes the router of the MoE model to enhance expert reuse in memory-constrained scenarios, reducing the number of times experts are read from storage.
π― What it does: The paper proposes the Boosted Direct Output (BDO) framework, combining the AR structure with the DO method, using recursive sub-predictions and achieving implicit prediction combination within the network, while employing EMA to smooth the training process to address the validation-test generalization gap;
Replay Failures as Successes: Sample-Efficient Reinforcement Learning for Instruction Following
Kongcheng Zhang (Zhejiang University), Mingli Song (Zhejiang University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: Propose the HiR framework, which achieves efficient RL training using only binary rewards by employing a hindsight replay approach of 'select-then-rewrite' that rewrites failure samples into successful ones.
Huayang Li (Nara Institute of Science and Technology), Richard Sproat (Sakana AI)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
π― What it does: Proposed a context-aware repositioning mechanism called REPO, which allows LLMs to assign context-related continuous positions to each token through a differentiable module within self-attention, thereby reducing the external burden on the attention layer.
CodeData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
π― What it does: Propose the REPRO method, which uses reinforcement learning to train a small language model to perform high-quality restatements of web data, thereby improving the quality of pre-training data.
π― What it does: This paper proposes the SpineFL framework, achieving federated learning on heterogeneous devices through backbone sharing and activation-guided submodel extraction.
ReTabSyn: Realistic Tabular Data Synthesis via Reinforcement Learning
Xiaofeng Lin (University of California, Los Angeles), Guang Cheng (University of California, Los Angeles)
CodeData SynthesisOptimizationSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelContrastive LearningTabularBenchmark
π― What it does: Propose a table data generation framework based on reinforcement learning called ReTabSyn, which directly optimizes the generator's conditional distribution using unsupervised preference comparison, enabling the generated data to maintain higher downstream model performance in scenarios with low sample sizes, extreme imbalance, and distribution drift.
Rethinking Efficient Graph Coarsening via a Non-Selfishness Principle
Xu Bai (Shanghai Jiao Tong University), Meng Jin (Shanghai Jiao Tong University)
CodeClassificationComputational EfficiencyRepresentation LearningGraph Neural NetworkLarge Language ModelContrastive LearningTextGraph
π― What it does: Proposed a non-selfish graph coarsening method called NOPE and its fast version NOPE*, based on the neighborhood interference index, achieving linear memory and near-linear time text-attribute graph coarsening.
π― What it does: Construct a complete natural accuracyβvalidation accuracy Pareto frontier through multi-objective hyperparameter search, systematically evaluate IBP-based certified training methods, and rediscover and quantify their performance boundaries.
π― What it does: Propose a loss reweighting method based on the inverse problem perspective of the neural collapse (NC) theory, aiming to make the average loss of each class equal, thereby alleviating the imbalance caused by long-tailed distribution;
Rethinking Pretraining Data Detection for LLMs: From Local to Global
Chenye Ke (University of Science and Technology of China), Qi Liu (University of Science and Technology of China)
CodeAnomaly DetectionData-Centric LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: Propose a pre-trained data detection framework called AECA from a global sequence perspective, which utilizes self-information calibration and convolutional filtering to capture fluctuation signals in probability sequences;
π― What it does: This paper systematically analyzes the design space of RL in diffusion models and finds that the likelihood estimation of ELBO based on the final generated samples is key to improving training efficiency and performance.
Rethinking the Trust Region in LLM Reinforcement Learning
Penghui Qi (Sea AI Lab), Wee Sun Lee (National University of Singapore)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
π― What it does: This paper rethinks the trust region mechanism of PPO in reinforcement learning for large-scale language models, proposing Divergence Proximal Policy Optimization (DPPO). It replaces the traditional ratio clipping with directly estimating the TV or KL divergence of the policy distribution (using binary and Top-K approximations), thereby more reasonably constraining policy updates.
π― What it does: This paper proposes a guiding method called IGG for the scale-recursive autoregressive (SwAR) image generation model, aiming to address the issues of scattered guidance signals and semantic mismatch in traditional classifier-free guidance (CFG) within SwAR.
REVIS: Sparse Latent Steering to Mitigate Object Hallucination in Large Vision-Language Models
Jialin Wu (Ant Group), Zhou Yang (Ant Group)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Proposes REVIS, a training-free sparse latent modulation framework that uses orthogonal projection to separate pure visual information vectors from deep activations of vision-language models, and dynamically injects this vector into the automatically located optimal layer to alleviate object hallucination;
π― What it does: The ZoVH framework is proposed by reformulating the zeroth-order Hessian estimation as the Hessian of single-step policy optimization, designing a low-variance Hessian estimator with average baseline optimality and reusable queries, and further constructing curvature-aware zeroth-order optimizer for the invertible Hessian and its gradient product.
π― What it does: Proposed a visual Transformer with discrete rotation-reflection group equivariance (REViT), achieving equivariance without position encoding through convolutional self-attention;
Reward Redistribution for CVaR MDPs using a Bellman Operator on L-infinity
Aneri Muni (Universite De Montreal), Erick Delage (Mila - Quebec Ai Institute)
CodeReinforcement LearningTabular
π― What it does: Study the Bellman operator for static CVaR MDP, propose a reward redistribution method, and provide value iteration and Q-learning algorithms.
RGMem: Renormalization Groupβinspired Memory Evolution for Language Agents
Ao Tian (Beihang University), Yanfang Liu (Beihang University)
CodeRecommendation SystemExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
π― What it does: Propose the RGMem framework, modeling dialogue memory as a multi-scale self-evolving system, achieving long-term personalization through hierarchical coarsening, threshold updating, and separation of slow/fast variables.
Riemannian Optimization for Fair Spectral Clustering
Minh Phu Vuong (Texas State University), Chul-Ho Lee (Texas State University)
CodeOptimizationGraph
π― What it does: Propose a fair spectral clustering algorithm based on Riemannian geometry optimization, R-FairSC, which avoids expensive feature decomposition and addresses the scalability issue in fair clustering.
CodeOptimizationGraph Neural NetworkTransformerReinforcement LearningMixture of ExpertsContrastive LearningWorld ModelGraphTabularTime SeriesSequential
π― What it does: Proposed and implemented RiskZero, a model-agnostic risk-aware decision-making framework based on MuZero, which can learn distributed returns over complete trajectories and perform risk-sensitive planning in unknown environmental dynamics.
RL with Learnable Textual Feedback: A Bilevel Approach
Utsav Singh (University of Central Florida), Amrit Singh Bedi (University of Central Florida)
CodeOptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsText
π― What it does: This paper proposes a dual-layer natural language actor-critic framework (Bi-NAC), which enables the actor (LLM) to self-improve after receiving feedback by embedding learnable textual feedback into the reinforcement learning loop;
RLAnything: Forge Environment, Policy, and Reward Model in Completely Dynamic RL System
Yinjie Wang (University of Chicago), Ling Yang (Princeton University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningWorld ModelTextMultimodality
π― What it does: Designed and implemented RLAnything, a reinforcement learning framework that dynamically optimizes the environment, policy, and reward model in a closed-loop.
RLSF-V: Mitigating Hallucinations in MLLMs via Fuzzy Semantic Self-Feedback
Changhao He (Sichuan University), Peng Hu (Sichuan University)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelTextMultimodalityBenchmark
π― What it does: Propose the RLSF-V framework, which reduces hallucinations in multi-modal large language models by evaluating the fuzzy semantics of internally generated logit-based preference data.
RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments
Zhiyuan Zeng (University of Washington), Hannaneh Hajishirzi (University of Washington)
CodeComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
π― What it does: This paper proposes RLVE (Reinforcement Learning with Adaptive Verifiable Environments), a reinforcement learning framework that dynamically adjusts the difficulty distribution using verifiable environments, aiming to enhance the reasoning capabilities of large language models.
Kianoosh Ashouritaklimi (University of Oxford), Francois Caron
CodeAnomaly DetectionOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningRobotic IntelligenceReinforcement Learning from Human FeedbackScore-based ModelContrastive LearningGaussian SplattingImageTabularTime SeriesSequentialBiomedical DataBenchmarkStochastic Differential EquationOrdinary Differential Equation
π― What it does: Propose the RoBAS (Robust Bayes-Assisted Shrinkage) framework, which constructs Bayesian-assisted inconsistency scores in the residual layer. This framework can leverage accurate prior information to generate efficient prediction intervals, and automatically degenerate into robust DTA scores when the prior conflicts with the data.
Robust Causal Discovery in Real-World Time Series with Power-Laws
Matteo Tusoni (Sapienza University of Rome), Novella Bartolini (Sapienza University of Rome)
CodeTime Series
π― What it does: We propose a causal discovery method called PLaCy, which restores causal structures in real-world time series data with non-stationary and noisy characteristics by fitting power-law to the time series spectrum and applying Granger causality tests on frequency domain features.
π― What it does: Proposed two mechanisms, Rooted Absorbed Prefix Trajectory Balance (RapTB) and Submodular Replay (SubM), to address the mode collapse, prefix collapse, and length bias issues of GFlowNet in terminable prefix trees (such as LLM generation tasks).
Rule2DRC: Benchmarking LLM Agents for DRC Script Synthesis with Execution-Guided Test Generation
Jinuk Kim (Seoul National University), Hyun Oh Song (Seoul National University)
CodeAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextTabularBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed a large-scale benchmark called Rule2DRC for evaluating the ability of LLMs to translate natural language rules into executable design rule check (DRC) scripts, and introduced SplitTester, a testing tool that leverages execution feedback to improve script selection.
RVAS: Referring Video Active Exploration and Segmentation
Hengrui Hu (Fudan University), Henghui Ding (Fudan University)
CodeSegmentationRobotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelVideoTextMultimodalityRetrieval-Augmented Generation
π― What it does: Propose the Referring Video Active Exploration and Segmentation (RVAS) task, which requires the agent to actively explore and segment target objects in real-time video streams based on natural language expressions.
S-Quant: Rethinking Weight Quantization with Seed-Based Generation
Mingzi Wang (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
CodeCompressionOptimizationComputational EfficiencyTransformerLarge Language ModelAuto EncoderText
π― What it does: Propose S-Quant, a post-training quantization method that only compresses weights, utilizing LFSR seed to dynamically generate basis matrices and reconstruct weight blocks through linear combinations.
π― What it does: Developed a unified diabetes simulation platform and a safety evaluation benchmark under distribution shift, and proposed a predictive shielding mechanism based on BAβNODE.
π― What it does: This paper studies the impact of backdoors and data poisoning attacks on models, and establishes a theoretical framework based on kernel ridge regression to explain the geometric mechanism of the attacks;
SAGE: A Dataflow-Native Framework for Modular, Controllable, and Transparent LLM-Augmented Reasoning
Jun Liu (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: A full-stack dataflow system called SAGE has been built for the inference pipeline of large language models (LLMs), which can compile multi-stage inference workflows (retrieval, embedding, vector search, state memory, context refinement, generation, etc.) into a physically executable graph that can be distributed. It achieves controllable optimization of tail latency and resource contention through explicit queue backpressure, pluggable scheduling/placement strategies, and a unified observation interface.
SALAAD: Sparse And Low-Rank Adaptation via ADMM for Large Language Model Inference
Hao Ma (ETH Zurich), Michael Muehlebach (Max Planck Institute for Intelligent Systems)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
π― What it does: During the pre-training of large language models, a framework named SALAAD is proposed, which dynamically introduces sparse + low-rank structures during training via ADMM, enabling the model to achieve elastic adjustable capacity with a single training session.
π― What it does: Propose a two-stage framework called SAMT, which generates structured, detail-rich 3D avatar meshes from a single facial image and synthesizes texture consistent with the viewpoint.
SARSteer: Safeguarding Large Audio Language Models via Safe-Ablated Refusal Steering
Weilin Lin (Hong Kong University of Science and Technology), Li Liu (Hong Kong University of Science and Technology)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextAudio
π― What it does: Proposes a reasoning-time security defense framework called SARSteer, aimed at enhancing the rejection behavior of large audio-language models (LALMs) and reducing false rejections.
SaTeen: Learning Structural Alignment for Continual Test-Time Adaptation
Chang Liu (Northwestern Polytechnical University), Yupei Zhang (Northwestern Polytechnical University)
CodeDomain AdaptationContrastive LearningImage
π― What it does: Propose the SaTeen method, which achieves continual test-time adaptation by aligning the internal structure of samples and the inter-sample structure during testing.
Scalable Reinforcement Learning via Adaptive Batch Scaling
Jongchan Park (Hyundai Motor Company)
CodeReinforcement LearningImageVideoTabular
π― What it does: Propose Adaptive Batch Scaling (ABS), which dynamically adjusts the batch size in RL training by measuring behavioral diversity, to balance early rapid adaptation and later fine convergence.
π― What it does: Proposes an scalable topology-preserving graph compression method called STPGC, which achieves complete preservation of topological features (connected components, cycles, holes) through graph strong compression, graph edge compression, and neighborhood coning.
π― What it does: Propose an unsupervised cross-domain person re-identification framework called SCALE, which addresses the scale inconsistency between the source and target domains, and improves the reliability of pseudo labels through bidirectional clustering regularization.
Scaling Continual Learning to 300+ Tasks with Bi-Level Routing Mixture-of-Experts
Meng Lou (University Of Hong Kong), Yizhou Yu (University Of Hong Kong)
CodeClassificationComputational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningTransformerPrompt EngineeringMixture of ExpertsContrastive LearningImageBenchmark
π― What it does: Propose an scalable continual learning framework called CaRE, which can achieve class-incremental learning on long sequences with over 300 tasks, and dynamically retrieves and aggregates expert knowledge at each network layer based on a pre-trained model.
π― What it does: Proposed the StP method, a size-consistent prior scaling approach, to address the inconsistency in denoising dynamics caused by molecular size in 3D molecular diffusion models.
Scaling Transformers for End-to-End Discrete Audio Tokenization
Yitian Gong (Fudan University), Xipeng Qiu (Fudan University)
CodeRecognitionGenerationCompressionTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderGenerative Adversarial NetworkContrastive LearningTextMultimodalityAudio
π― What it does: Propose a full end-to-end Transformer-based Audio Codec (TAC), achieving high-fidelity discrete audio tokenization through a unified causal Transformer encoder, quantizer, and decoder, and applying this tokenizer to autoregressive text-to-speech (TTS) and automatic speech recognition (ASR) tasks.
Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation
Heng Qu (Wuhan University), Jian Luan (Xiaomi)
CodeAutonomous DrivingOptimizationSafty and PrivacyComputational EfficiencyTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelGaussian SplattingImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Constructed a large-scale Chinese mobile GUI task dataset called HyperTrack, developed a unified evaluation framework named GUIEvalKit, and systematically studied the training scale of VLM agents, the comparison between reinforcement learning and supervised learning, semi-online evaluation methods, and the impact of reasoning on decision diversity and stability.
scDataset: Scalable Data Loading for Deep Learning on Large-Scale Single-Cell Omics
Davide D'Ascenzo (University of Milan), Sebastiano Cultrera di Montesano (Broad Institute of MIT and Harvard)
CodeComputational EfficiencyData-Centric LearningTabularBiomedical Data
π― What it does: This paper proposes scDataset, an IterableDataset for PyTorch, designed to efficiently load single-cell data directly from disk, enabling deep learning training on large-scale single-cell omics data.
π― What it does: Proposes ScenePilot, a controllable boundary-driven scene generation framework guided by two signals: physical feasibility and AV risk, which generates high-risk scenarios that can cause deployed autonomous driving stacks to fail while ensuring physical feasibility.
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Proposed the TIE scheduling algorithm, which models the output length of LLMs using a log-t distribution and combines CVaR with Tail-Inflated Expectation to achieve non-preemptive LLM inference scheduling.
Scientific logicality enriched methodology for LLM reasoning: A practice in physics
Zhaoxin Yu (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences), Wenji Mao (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmarkPhysics Related
π― What it does: Designed and implemented a logic evaluation and training method for large language models, constructing a physics problem set and the first multidimensional logic benchmark, PHYSLOGIC;
SciPredict: Can LLMs Predict the Outcomes of Scientific Experiments in Natural Sciences?
Udari Madhushani Sehwag (Scale AI), Bing Liu (Scale AI)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkPhysics RelatedRetrieval-Augmented Generation
π― What it does: This study proposes and constructs the SciPredict benchmark to evaluate the ability of large language models (LLMs) in predicting the results of natural science experiments;
ScoreMatchingRiesz: Score Matching for Debiased Machine Learning and Policy Path Estimation
Masahiro Kato (Mizuho-DL Financial Technology Co Ltd)
CodeOptimizationFederated LearningExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackDiffusion modelScore-based ModelTabularTime SeriesBenchmark
π― What it does: Proposed a Riesz representative estimator framework based on score matching, named ScoreMatchingRiesz, to assist in estimating causal and structural parameters in debiased learning.
Scout Before You Attend: Sketch-and-Walk Sparse Attention for Efficient LLM Inference
Hoang Anh Duy Le (Rice University), Anshumali Shrivastava (Rice University)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose a training-agnostic, unified sparse attention method called Sketch&Walk, which can significantly reduce computational and memory costs during the prefill and decoding phases in long-context reasoning of large language models (LLMs).
SCOUT: Cyclic Causal Discovery Under Soft Interventions with Unknown Targets
Alpar Turkoglu (Georgia Institute of Technology), Faramarz Fekri (Georgia Institute of Technology)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkFlow-based ModelContrastive LearningGraphTabularBiomedical Data
π― What it does: Propose the SCOUT framework, which can learn nonlinear directed cyclic causal graphs from soft intervention data and simultaneously infer intervention targets without knowing the intervention objectives.