ICLR 2025 Papers — Page 30
International Conference on Learning Representations · 3704 papers
Scalable Influence and Fact Tracing for Large Language Model Pretraining
Tyler A. Chang (Google DeepMind), Ian Tenney (Google DeepMind)
RetrievalOptimizationTransformerLarge Language ModelText
🎯 What it does: A gradient-based influence method named TrackStar is proposed to efficiently retrieve training samples that significantly impact model predictions during the pre-training phase of large language models (8B parameters) and to perform 'fact tracking' for factual reasoning tasks.
Scalable Mechanistic Neural Networks
Jiale Chen (Institute of Science and Technology Austria), Francesco Locatello (Institute of Science and Technology Austria)
Time SeriesSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes the Scalable Mechanistic Neural Network (S-MNN), which improves the original Mechanistic Neural Network (MNN) to efficiently model long time series while balancing prediction and equation discovery.
Scalable Universal T-Cell Receptor Embeddings from Adaptive Immune Repertoires
Paidamoyo Chapfuwa (Microsoft Research), Julia Greissl (Microsoft Research)
ClassificationRepresentation LearningDrug DiscoverySupervised Fine-TuningBiomedical Data
🎯 What it does: Using the co-occurrence statistics of T cell receptors (TCR), combined with the GloVe algorithm and Johnson-Lindenstrauss random projection for initialization, we learn low-dimensional dense TCR embeddings and aggregate them into individual immune repertoire embeddings for predicting HLA types and various diseases.
Scale-Aware Contrastive Reverse Distillation for Unsupervised Medical Anomaly Detection
Chunlei Li (MedAI Technology), Lichao Mou (MedAI Technology)
Anomaly DetectionConvolutional Neural NetworkContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A method for unsupervised medical anomaly detection based on scale-aware contrastive reverse distillation is proposed.
Scale-aware Recognition in Satellite Images under Resource Constraints
Shreelekha Revankar (Cornell University), Kavita Bala (Cornell University)
RecognitionRetrievalOptimizationKnowledge DistillationConvolutional Neural NetworkTransformerLarge Language ModelImage
🎯 What it does: This paper proposes a recognition scheme for conceptual scales in satellite images, utilizing low-resolution (LR) images to achieve high-precision retrieval under a limited budget of high-resolution (HR) images, addressing the trade-off between scale and cost.
Scale-Free Graph-Language Models
Jianglin Lu (Northeastern University), Yun Fu (Northeastern University)
ClassificationGraph Neural NetworkSupervised Fine-TuningTextGraph
🎯 What it does: A unified scale-invariant graph-language model (SFGL) is proposed, which simultaneously completes graph structure generation and text embedding in semi-supervised text classification.
Scaling and evaluating sparse autoencoders
Leo Gao (OpenAI), Jeffrey Wu (OpenAI)
OptimizationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This study explores the scaling and sparsity expansion laws of autoencoders by training large-scale sparse autoencoders on the residual flows of language models (such as GPT-2 small and the GPT-4 series), and proposes several new metrics for evaluating the quality of sparse features.
Scaling Autonomous Agents via Automatic Reward Modeling And Planning
Zhenfang Chen (MIT IBM Watson AI Lab), Chuang Gan (UMass Amherst and MIT IBM Watson AI Lab)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelText
🎯 What it does: The ARMAP framework is proposed, which utilizes LLM to automatically generate reward models and combines these reward models with LLM agents and various planning algorithms (Best-of-N, Reflexion, MCTS) to enhance the performance of LLM in multi-step decision-making tasks.
Scaling Diffusion Language Models via Adaptation from Autoregressive Models
Shansan Gong (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Adapt existing autoregressive large models (such as GPT-2 and LLaMA-2) using techniques like attention mask annealing and output displacement to train diffusion language models DiffuGPT and DiffuLLaMA with 127M, 355M, and 7B parameters.
Scaling FP8 training to trillion-token LLMs
Maxim Fishman (Intel), Daniel Soudry (Technion)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Trained a 7B parameter LLM on the 2 trillion token Red Pajama dataset and addressed instability issues during the FP8 training process.
Scaling In-the-Wild Training for Diffusion-based Illumination Harmonization and Editing by Imposing Consistent Light Transport
Lvmin Zhang (Stanford University), Maneesh Agrawala (Stanford University)
Image TranslationImage HarmonizationDiffusion modelImage
🎯 What it does: Proposes a constraint for enforcing consistent light transmission in the training of diffusion models, achieving stable training of large-scale unsupervised lighting editing models.
Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation
Linda He (Harvard University), Ce Zhang (University of Chicago)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A hierarchical synthetic data generation process is proposed to automatically construct instruction tuning data with context lengths exceeding 100K and even 1M tokens, gradually expanding the context window using RoPE;
Scaling Large Language Model-based Multi-Agent Collaboration
Chen Qian (Tsinghua University), Maosong Sun (Tsinghua University)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes the MACNET framework, which organizes multiple agents through a directed acyclic graph (DAG) to collaboratively reason and complete complex tasks, achieving cooperation among over a thousand agents.
Scaling Laws for Adversarial Attacks on Language Model Activations and Tokens
Stanislav Fort (Independent Researcher)
Adversarial AttackLarge Language ModelText
🎯 What it does: This study explores a class of adversarial attacks targeting the activations of language models and derives the upper scaling law of their susceptibility to attacks. By manipulating a relatively small subset of model activations, it demonstrates the ability to precisely control a large number of subsequent tokens.
Scaling Laws for Downstream Task Performance in Machine Translation
Berivan Isik (Google Research), Sanmi Koyejo (Stanford University)
TransformerSupervised Fine-TuningText
🎯 What it does: This study investigates the impact of pre-training scale on downstream performance in machine translation and proposes a logarithmic scaling law for translation quality.
Scaling Laws for Precision
Tanishq Kumar (Harvard University), Aditi Raghunathan (Carnegie Mellon University)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the impact of low-precision training and inference on the loss of language models and presents a unified precision scaling law.
Scaling LLM Test-Time Compute Optimally Can be More Effective than Scaling Parameters for Reasoning
Charlie Victor Snell (University of California Berkeley), Aviral Kumar (Carnegie Mellon University)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies how to improve the inference performance of large language models by dynamically allocating computational resources (search + stepwise correction) during the inference phase, and proposes a 'computationally optimal' strategy based on problem difficulty.
Scaling Long Context Training Data by Long-Distance Referrals
Yonghao Zhuang (Carnegie Mellon University), Hao Zhang (University of California San Diego)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study addresses the issue of scarce training data for long contexts and proposes the LongPack data pipeline, which generates high-quality long texts by stitching together web hyperlinks to enhance the long-context learning effectiveness of large models.
Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining
Jie Cheng (Chinese Academy of Sciences), Yisheng Lv
TransformerReinforcement LearningWorld ModelImage
🎯 What it does: In multi-task offline reinforcement learning, a single large-scale model JOWA is trained, utilizing a world model of image observations and joint optimization of Q-values to achieve a general agent for multiple Atari games.
Scaling Optimal LR Across Token Horizons
Johan Bjorck (Nvidia), Xia Song (Microsoft)
OptimizationHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: This paper explores the relationship between learning rate (LR) and the number of training tokens (token horizon) in large language model training through large-scale experimental systems. It finds that LR decreases with the token horizon and provides a power-law scaling rule that can be used for hyperparameter transfer across token horizons.
Scaling Speech-Text Pre-training with Synthetic Interleaved Data
Aohan Zeng (Tsinghua University), Jie Tang (Tsinghua University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: This paper proposes a pre-training method for speech language models (SpeechLM) that utilizes synthesized interleaved speech-text data to massively expand the text corpus.
Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study
Shawn Tan (Massachusetts Institute of Technology IBM Watson Artificial Intelligence Laboratory), Yikang Shen (Massachusetts Institute of Technology IBM Watson Artificial Intelligence Laboratory)
RetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelTextSequential
🎯 What it does: Proposed and implemented an attention mechanism based on the stick-breaking process, replacing the traditional softmax + RoPE;
Scaling Transformers for Low-Bitrate High-Quality Speech Coding
Julian D Parker, Xubo Liu (Stability AI)
CompressionTransformerAuto EncoderAudio
🎯 What it does: Designed and trained a Transformer-based audio autoencoder (TAAE) with a tunable discrete bottleneck using Finite Scalar Quantization (FSQ) for extremely low bitrate (400-700 bps) encoding of 16kHz speech.
Scaling up Masked Diffusion Models on Text
Shen Nie (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelText
🎯 What it does: This study investigates the scalability and effectiveness of Masked Diffusion Models (MDM) in text generation and language understanding, and proposes an unsupervised Classifier-Free Guidance (CFG) to enhance conditional reasoning performance.
Scaling up the Banded Matrix Factorization Mechanism for Large Scale Differentially Private ML
Ryan McKenna (Google Research)
OptimizationSafty and Privacy
🎯 What it does: A method for extending the DP-BANDMF mechanism is proposed to address scalability limitations in large-scale training scenarios, allowing for the handling of an arbitrary number of model parameters and training iterations.
Scaling Wearable Foundation Models
Girish Narayanswamy (Google Research), Daniel McDuff (Google Research)
RecognitionGenerationTransformerAuto EncoderMultimodalityTime Series
🎯 What it does: This paper constructs a foundational model for multimodal wearable sensors (LSM), performing self-supervised pre-training on over 40 million hours of wearable sensor data from 165,000 individuals, and systematically studies the scaling laws of computation, data, and model size. It also evaluates the model's performance on generative tasks (random filling, temporal interpolation, sensor filling, temporal extrapolation) and discriminative tasks (motion detection, activity recognition).
SCBench: A KV Cache-Centric Analysis of Long-Context Methods
YUCHENG LI, Lili Qiu (Microsoft Corporation)
TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the SCBench benchmark for evaluating the KV cache lifecycle performance of long-context LLMs in multi-turn, cross-request shared context scenarios;
Schur's Positive-Definite Network: Deep Learning in the SPD cone with structure
Can Pouliquen (ENS de Lyon), Titouan Vayer (ENS de Lyon)
Tabular
🎯 What it does: A neural network module named SpodNet is proposed, which ensures that the output matrix is strictly symmetric positive definite (SPD) during training and can embed additional structural constraints (such as element-wise sparsity), achieving joint learning of SPD and sparsity.
ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery
Ziru Chen (Ohio State University), Huan Sun (Ohio State University)
Large Language ModelTextBenchmark
🎯 What it does: A benchmark called ScienceAgentBench is proposed to evaluate the automation capabilities of language models in data-driven scientific discovery.
SciLitLLM: How to Adapt LLMs for Scientific Literature Understanding
Sihang Li (University of Science and Technology of China), Hengxing Cai (DP Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A specialized LLM for scientific literature understanding, SciLitLLM, has been constructed, and a complete training pipeline combining Continuous Pre-Training (CPT) and Supervised Fine-Tuning (SFT) has been proposed.
ScImage: How good are multimodal large language models at scientific text-to-image generation?
Leixin Zhang (University of Twente), Zhixue Zhao (University of Sheffield)
GenerationTransformerLarge Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: The ScImage benchmark is proposed to evaluate the capabilities of multimodal large language models in generating images (or code to images) from scientific texts, and systematic experiments were conducted on the performance of seven mainstream models on this benchmark.
SCOPE: A Self-supervised Framework for Improving Faithfulness in Conditional Text Generation
Song Duong (Criteo AI Lab), Patrick Gallinari (Sorbonne Université)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By generating untrustworthy samples through self-supervision and performing preference fine-tuning based on this, the fidelity of large language models in conditional text generation tasks is improved.
Score Forgetting Distillation: A Swift, Data-Free Method for Machine Unlearning in Diffusion Models
Tianqi Chen (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
GenerationKnowledge DistillationDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes Score Forgetting Distillation (SFD), a data-free, score-based machine unlearning method for forgetting specified categories or concepts in diffusion models and achieving first-order sampling.
Score-based free-form architectures for high-dimensional Fokker-Planck equations
Feng Liu (Beihang University), Xiao Zhang (Beihang University)
Score-based ModelPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper proposes a Fokker-Planck Neural Network (FPNN) based on fractional PDE loss, which addresses high-dimensional steady-state FP equations through a two-stage separation of normalization and fitting.
Score-based Self-supervised MRI Denoising
Jiachen Tu (University of Illinois), Fan Lam (University of Illinois)
RestorationScore-based ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A self-supervised MRI denoising framework C2S based on score matching is proposed, which can directly learn denoising from noisy images.
Scrutinize What We Ignore: Reining In Task Representation Shift Of Context-Based Offline Meta Reinforcement Learning
Hai Zhang (Tongji University), Lanqing Li (Zhejiang Lab)
Meta LearningReinforcement LearningTabular
🎯 What it does: This paper studies the issue of task representation drift in Context-Based Offline Meta Reinforcement Learning (COMRL) and proposes a theoretical framework for achieving monotonic performance improvement by controlling the updates of the context encoder.
SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning
Yichen Wu (City University of Hong Kong), Ying Wei (Zhejiang University)
ClassificationKnowledge DistillationTransformerImage
🎯 What it does: A scalable and decoupled low-rank adaptation method, SD-LoRA, is proposed for class-incremental learning, which is completely sample-free and can be optimized end-to-end.
SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection
Han Shen (Rensselaer Polytechnic Institute), Tianyi Chen (Rensselaer Polytechnic Institute)
OptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The SEAL framework is proposed, which uses a two-layer optimization to learn a data selector that automatically filters safe and high-quality samples during the fine-tuning process to enhance the model's safety and performance.
Searching for Optimal Solutions with LLMs via Bayesian Optimization
Dhruv Agarwal (University of Massachusetts Amherst), Rashmi Gangadharaiah (Amazon Web Services AI Labs)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A search framework called BOPRO has been designed and implemented, which combines Bayesian optimization with large language models to automatically adjust search strategies in uncertain environments.
SEBRA : Debiasing through Self-Guided Bias Ranking
Adarsh Kappiyath (University of Texas at Austin), Lu Yin (University of Surrey)
ClassificationData-Centric LearningContrastive LearningImageText
🎯 What it does: An unsupervised self-guided bias ranking framework called Sebra is proposed, which can perform fine-grained ranking of sample spuriosity without the need for manual annotations, and utilizes the ranking results for contrastive learning to achieve debiasing.
SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents
Zhuoshi Pan (Tsinghua University), Jianfeng Gao (Microsoft Corporation)
GenerationRetrievalCompressionTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a retrieval-augmented generation framework called SECOM, based on dialogue segmentation and compression denoising, to address the issues of memory retrieval and contextual coherence in long dialogues.
Second Order Bounds for Contextual Bandits with Function Approximation
Aldo Pacchiano (Boston University)
Reinforcement Learning
🎯 What it does: This paper presents second-order bounds for context-dependent bandits with function approximation and develops new algorithms to meet these bounds.
Second-Order Fine-Tuning without Pain for LLMs: A Hessian Informed Zeroth-Order Optimizer
Yanjun Zhao (Xi'an Jiaotong University), Ivor Tsang (Nanyang Technological University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: For full parameter or PEFT fine-tuning of large language models, a zero-order optimizer HiZOO utilizing diagonal Hessian information is proposed, along with theoretical analysis and experimental validation.
Second-Order Min-Max Optimization with Lazy Hessians
Lesi Chen (Tsinghua University), Jingzhao Zhang (Tsinghua University)
OptimizationTabular
🎯 What it does: Two second-order optimal methods, LEN (Lazy Extra Newton) and LEN-restart, are proposed for solving min-max optimization problems that are convex-concave and strongly convex-strongly concave, utilizing 'lazy Hessian' to reuse old Hessians, significantly reducing computational costs.
SecureGS: Boosting the Security and Fidelity of 3D Gaussian Splatting Steganography
Xuanyu Zhang (Peking University), Jian Zhang (Peking University)
CompressionSafty and PrivacyGaussian SplattingPoint Cloud
🎯 What it does: A 3D Gaussian Splatting steganography framework named SecureGS is proposed, which can losslessly embed and recover 3D objects, images, or bits while maintaining the original scene's visualization quality for authorized users.
See It from My Perspective: How Language Affects Cultural Bias in Image Understanding
Amith Ananthram (Columbia University), Kathleen McKeown (University of North Carolina Chapel Hill)
RecognitionObject DetectionTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Evaluate and explore the cultural bias of visual language models (VLM) due to language, comparing the differences in image understanding between Western and East Asian cultures, and examining the impact of multilingual VLM training on bias recognition.
See What You Are Told: Visual Attention Sink in Large Multimodal Models
Seil Kang (Yonsei University), Seong Jae Hwang (Yonsei University)
TransformerVision Language ModelMultimodality
🎯 What it does: This study investigates the phenomenon of visual attention sink in large-scale multimodal models (LMM) and proposes the Visual Attention Redistribution (VAR) method to reallocate attention, thereby enhancing the model's focus on images.
SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators
Rasoul Shafipour (Apple), Saman Naderiparizi (Apple)
CompressionComputational EfficiencyLarge Language ModelText
🎯 What it does: This paper proposes SeedLM, a post-training compression method for LLM weights based on pseudo-random generator seed compression.
Seeing Eye to AI: Human Alignment via Gaze-Based Response Rewards for Large Language Models
Ángela López-Cardona (Telefónica Scientific Research), Ioannis Arapakis (Telefónica Scientific Research)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the GazeReward framework, which integrates implicit feedback generated by eye tracking (ET) into the reward model (RM) to enhance the human alignment effects of large language models (LLMs).
SegLLM: Multi-round Reasoning Segmentation with Large Language Models
XuDong Wang (University of California Berkeley), Trevor Darrell (University of California Berkeley)
SegmentationTransformerLarge Language ModelImageTextBenchmark
🎯 What it does: We constructed SegLLM, a multi-turn interactive reasoning segmentation model that can re-inject previously generated segmentation masks into the input stream of the LLM in each round of dialogue and generate new segmentation results through a mask-aware decoder.
Segment Any 3D Object with Language
Seungjun Lee (National University of Singapore), Gim Hee Lee (National University of Singapore)
Object DetectionSegmentationTransformerVision Language ModelMultimodalityPoint Cloud
🎯 What it does: A framework named SOLE is proposed, which achieves open vocabulary 3D instance segmentation based on free-form language instructions.
Select before Act: Spatially Decoupled Action Repetition for Continuous Control
Buqing Nie (Shanghai Jiao Tong University), Yue Gao (Shanghai Jiao Tong University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A spatially decoupled action repetition framework (SDAR) is proposed, which can independently make act-or-repeat decisions for each action dimension in continuous control tasks.
SelectFormer in Data Markets: Privacy-Preserving and Efficient Data Selection for Transformers with Multi-Party Computation
Xu Ouyang (University of Virginia), Yangfeng Ji (University of Virginia)
Safty and PrivacyComputational EfficiencyTransformerImageText
🎯 What it does: A privacy data selection process based on multi-party secure computation (MPC) has been designed, which can select the most valuable data for model owners while ensuring that model and data ownership is not disclosed.
Selective Aggregation for Low-Rank Adaptation in Federated Learning
Pengxin Guo (University of Hong Kong), Liangqiong Qu (University of Hong Kong)
Federated LearningTransformerLarge Language ModelText
🎯 What it does: Addressing the low-rank adaptation problem of LoRA in federated learning, the roles of matrices A and B were analyzed from both theoretical and experimental perspectives. It was found that matrix A learns general knowledge while matrix B learns client-specific knowledge. Based on this, FedSA-LoRA was proposed and extended to rsLoRA and VeRA, forming FedSA-rsLoRA and FedSA-VeRA.
Selective Attention Improves Transformer
Yaniv Leviathan (Google Research), Yossi Matias (Google Research)
TransformerLarge Language ModelText
🎯 What it does: A mechanism called Selective Attention is proposed, aimed at reducing the focus on unnecessary elements in the attention mechanism, thereby improving the performance of language modeling and downstream tasks.
Selective Induction Heads: How Transformers Select Causal Structures in Context
Francesco D'Angelo (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
TransformerSequential
🎯 What it does: A novel synthetic task based on interleaved Markov chains is proposed to study the ability of Transformers to select causal structures in context, and a three-layer attention network is constructed to implement Selective Induction Heads, which adaptively select the correct lags and predict the next token.
Selective Label Enhancement Learning for Test-Time Adaptation
Yihao Hu (Southeast University), Ning Xu (Southeast University)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the PASLE framework for Online Testing Time Adaptation (OTTA) on unlabeled test data, which classifies samples into deterministic (one-hot) and uncertain (candidate set) categories through selective label enhancement, and dynamically refines the candidate labels of uncertain samples during the training process.
Selective Task Group Updates for Multi-Task Optimization
Wooseong Jeong (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
OptimizationImage
🎯 What it does: An optimization algorithm is proposed for dynamically partitioning task groups and updating them sequentially in multi-task learning.
Selective Unlearning via Representation Erasure Using Domain Adversarial Training
Nazanin Mohammadi Sepahvand (McGill University), Gintare Karolina Dziugaite (Google DeepMind)
Domain AdaptationRepresentation LearningGenerative Adversarial NetworkImage
🎯 What it does: The SURE algorithm is proposed, which utilizes domain adversarial training to perform 'forgetting' operations in the representation space;
Self-Attention-Based Contextual Modulation Improves Neural System Identification
Isaac Lin (Carnegie Mellon University), Tai Sing Lee (Carnegie Mellon University)
Convolutional Neural NetworkImage
🎯 What it does: This study investigates the impact of incorporating self-attention layers into convolutional neural networks on predicting the responses of primate V1 neurons, and introduces a peak tuning metric to measure the model's ability to capture the strongest responses of neurons.
Self-Boosting Large Language Models with Synthetic Preference Data
Qingxiu Dong (Peking University), Furu Wei (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a self-enhancement paradigm called SynPO, which utilizes a small amount of SFT seed data to enable the LLM to generate a large number of synthetic prompts and corresponding pairs of good and bad responses. Through an iterative 'improver' model, the LLM self-improves its generated responses, ultimately achieving self-alignment and performance enhancement without relying on manual annotations or strong teacher models.
Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models
Ce Zhang (Carnegie Mellon University), Yaqi Xie (Carnegie Mellon University)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: A training-free self-correcting decoding method DeGF is proposed, which utilizes visual feedback from a text-to-image generation model to recursively correct hallucinations during the decoding process of LVLM-generated responses.
SELF-EVOLVED REWARD LEARNING FOR LLMS
Chenghua Huang (Fudan University), Qi Zhang (Microsoft)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A self-evolved reward learning model (SER) is constructed, which uses a small amount of manually labeled data to train a reward model that self-labels unlabeled data and iteratively filters high-confidence samples to further enhance the reward model. This reward model is then used to guide the PPO reinforcement learning of LLM.
Self-Evolving Multi-Agent Collaboration Networks for Software Development
Yue Hu (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
OptimizationAI Code AssistantLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: An adaptive EvoMAC multi-agent collaboration network is proposed, which can self-evolve during the testing phase through a single task iteration, achieving automatic code generation from function level to software level. At the same time, a demand-oriented software development benchmark, rSDE-Bench, is constructed, providing diverse requirements and automated unit testing evaluation.
Self-Improvement in Language Models: The Sharpening Mechanism
Audrey Huang (University of Illinois Urbana-Champaign), Akshay Krishnamurthy (Microsoft Research)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes a 'Sharpening' mechanism, studying how language models can self-improve their generation process through an internal verifier without external supervision, providing a theoretical framework, algorithm analysis, and experimental validation.
Self-Improving Robust Preference Optimization
Eugene Choi (Cohere), Mohammad Gheshlaghi Azar (Cohere)
OptimizationTransformerSupervised Fine-TuningText
🎯 What it does: A self-improving robust preference optimization (SRPO) framework is proposed for training language models from human preference data.
Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models
Fushuo Huo (Hong Kong Polytechnic University), Peilin Zhao (Tencent)
GenerationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes a Self-Introspective Decoding (SID) mechanism that utilizes self-assessment of the importance of visual tokens through a pre-trained large visual-language model. It retains only the least important visual tokens in the early decoding layers to generate visual-textual associations and suppresses hallucinations through contrastive decoding.
Self-MoE: Towards Compositional Large Language Models with Self-Specialized Experts
Junmo Kang (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Transform a single large language model (LLM) into a composable system composed of self-generated lightweight specialized experts (MiXSE), and dynamically call the appropriate experts through self-optimizing routing to enhance multi-domain capabilities.
Self-Normalized Resets for Plasticity in Continual Learning
Vivek Farias, Adam Daniel Jozefiak
OptimizationText
🎯 What it does: This paper proposes an adaptive threshold-based self-normalizing reset algorithm SNR to address the plasticity loss problem in continual learning.
Self-Play Preference Optimization for Language Model Alignment
Yue Wu (University of California), Quanquan Gu (University of California)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A self-play preference optimization (SPPO) framework is designed and implemented to fine-tune large language models for better alignment with human preferences without external supervision.
Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models
Guanting Dong (Alibaba Inc), Jingren Zhou (Alibaba Inc)
Data SynthesisOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes AUTOIF, a method for instruction-following data synthesis without manual annotation by automatically generating verifiable instructions and corresponding execution code;
Self-supervised contrastive learning performs non-linear system identification
Rodrigo González Laiz (Institute of Computational Biology), Steffen Schneider (Institute of Computational Biology)
Anomaly DetectionRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningTime SeriesSequential
🎯 What it does: Proposes the Dynamics Contrastive Learning (DCL) framework, utilizing self-supervised contrastive learning for systematic identification of time-series data;
Self-Supervised Diffusion Models for Electron-Aware Molecular Representation Learning
Gyoung S. Na (Korea Research Institute of Chemical Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)
Representation LearningDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A self-supervised diffusion model DELID is proposed, which learns electron-perceptive molecular representations through molecular substructure decomposition and electronic layer information retrieved from databases, thereby improving molecular property prediction performance without the need for expensive quantum mechanical calculations.
Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement
Chenxu Wu (University of Science and Technology of China), S Kevin Zhou
RestorationDiffusion modelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A fully self-supervised MRI denoising framework called Di-Fusion is proposed, which utilizes post-diffusion steps and adaptive sampling to achieve denoising and iterative refinement of dMRI.
Self-supervised Monocular Depth Estimation Robust to Reflective Surface Leveraged by Triplet Mining
Wonhyeok Choi (Daegu Gyeongbuk Institute of Science and Technology), Sunghoon Im (Stanford University)
Depth EstimationKnowledge DistillationContrastive LearningImage
🎯 What it does: A self-supervised monocular depth estimation training strategy is proposed, combining triplet learning with pixel-level reflection area mining and reflection-aware knowledge distillation, thereby improving the depth prediction accuracy for reflective surfaces.
Self-Updatable Large Language Models by Integrating Context into Model Parameters
Yu Wang (University of California San Diego), Julian McAuley (University of California San Diego)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A self-updating large language model called SELF-PARAM is proposed, which achieves this without additional parameters by minimizing KL divergence. It can directly embed contextual knowledge into model parameters, enabling rapid memory and long-term retention.
SelKD: Selective Knowledge Distillation via Optimal Transport Perspective
Liangliang Shi (Shanghai Jiao Tong University), Junchi Yan (Shanghai Artificial Intelligence Laboratory)
ClassificationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A Selective Knowledge Distillation (SelKD) framework is proposed, allowing the student model to learn knowledge from specified sub-tasks of the teacher model, and achieving flexible selection and matching of knowledge through Inverse Optimal Transport (IOT).
Semantic Aware Representation Learning for Lifelong Learning
Fahad Sarfraz (Eindhoven University of Technology), Bahram Zonooz (Eindhoven University of Technology)
Representation LearningContrastive LearningImage
🎯 What it does: This paper proposes Semantic-Aware Representation Learning (SARL), which utilizes sparse activation and object prototype learning to capture semantic similarity across tasks, guiding representation learning to reduce interference and enhance lifelong learning effectiveness.
Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations
Litu Rout (Google), Wen-Sheng Chu (Google)
RestorationGenerationRectified FlowImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a method for zero-shot image inversion and editing using a Rectified Flow model (Flux) that leverages controlled ODE/SDE, capable of mapping images to noise and reconstructing noise to images without additional training, optimization, or the use of complex attention modules.
Semantic Loss Guided Data Efficient Supervised Fine Tuning for Safe Responses in LLMs
Yuxiao Lu (Singapore Management University), Pradeep Varakantham (Singapore Management University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: During the SFT phase, a very small amount of unsafe (harmful) responses is used to fine-tune the LLM for safety, so that it no longer generates harmful answers when faced with toxic prompts.
Semantic Temporal Abstraction via Vision-Language Model Guidance for Efficient Reinforcement Learning
Tian-Shuo Liu (Nanjing University), Yang Yu (Nanjing University)
TransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A time abstraction method called VanTA is proposed, which is guided by a pre-trained visual-language model (VLM) to automatically extract semantic, task-related discrete skills from offline data, and employs a two-layer strategy (high-level skill selection and low-level action execution based on skills) for reinforcement learning.
Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors
Weixuan Wang (University of Edinburgh), Wei Peng (RMIT University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A dynamic activation intervention method based on input semantic adaptation (SADI) is proposed, which adjusts the internal activations of LLMs in real-time during the inference phase to achieve behavior alignment.
Semantix: An Energy-guided Sampler for Semantic Style Transfer
Huiang He (South China University of Technology), Tat-Jen Cham (Nanyang Technological University)
Image TranslationGenerationDiffusion modelImageVideoStochastic Differential Equation
🎯 What it does: This paper presents Semantix, an energy-guided sampler that achieves style and appearance transfer in images and videos through semantic correspondence.
SEMDICE: Off-policy State Entropy Maximization via Stationary Distribution Correction Estimation
Jongmin Lee (Yonsei University), Pieter Abbeel (University of California Berkeley)
Reinforcement Learning
🎯 What it does: A new offline unsupervised reinforcement learning algorithm called SEMDICE is proposed to learn a policy that maximizes state entropy from any offline dataset.
Semi-Parametric Retrieval via Binary Bag-of-Tokens Index
Jiawei Zhou (Hong Kong University of Science and Technology), Lei Chen (Hong Kong University of Science and Technology)
RetrievalContrastive LearningTextBenchmark
🎯 What it does: A semi-parametric retrieval framework SIDR is proposed, which supports both parameterized indexing based on neural embeddings and non-parameterized indexing based on Bag-of-Tokens;
Semi-Supervised CLIP Adaptation by Enforcing Semantic and Trapezoidal Consistency
Kai Gan (Southeast University), Tong Wei (Southeast University)
ClassificationRetrievalDomain AdaptationTransformerContrastive LearningImage
🎯 What it does: A semi-supervised CLIP adaptation method called SEMICLIP is proposed, which utilizes unlabeled images through semantic concept mining and trapezoidal consistency.
Semi-Supervised Vision-Centric 3D Occupancy World Model for Autonomous Driving
Xiang Li (Tsinghua University), yilun chen
Autonomous DrivingWorld ModelPoint CloudBenchmark
🎯 What it does: This paper presents PreWorld, a semi-supervised visual-centric 3D occupancy world model, which achieves unified modeling of 3D occupancy prediction, 4D prediction, and motion planning through a two-stage training process.
Semialgebraic Neural Networks: From roots to representations
S David Mis, Maarten V. de Hoop (Rice University)
Ordinary Differential Equation
🎯 What it does: This paper proposes Semialgebraic Neural Networks (SANN), a network architecture capable of accurately representing all bounded semialgebraic functions and solving them through numerical ODEs.
Sensitivity Verification for Additive Decision Tree Ensembles
Arhaan Ahmad (Indian Institute of Technology Bombay), S. Akshay (Indian Institute of Technology Bombay)
ClassificationOptimizationTabularBenchmark
🎯 What it does: This paper proposes a sensitivity validation framework for additive decision tree ensemble models and implements a tool called SENSPB based on pseudo-Boolean constraints to test the sensitivity of a given feature set.
Sensitivity-Constrained Fourier Neural Operators for Forward and Inverse Problems in Parametric Differential Equations
Abdolmehdi Behroozi (Penn State University), Daniel Kifer (Penn State University)
Physics RelatedOrdinary Differential Equation
🎯 What it does: A new sensitivity-constrained Fourier neural operator (SC-FNO) is proposed to address forward and inverse problems in parameter differential equations, particularly the challenges in sensitivity computation and concept drift.
Sensor-Invariant Tactile Representation
Harsh Gupta (University of Illinois), Wenzhen Yuan (University of Illinois)
ClassificationPose EstimationDomain AdaptationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This study investigates the cross-sensor transfer problem of visual-tactile sensors and proposes the Sensor-Invariant Tactile Representation (SITR), achieving zero-shot transfer.
SEPARATE: A Simple Low-rank Projection for Gradient Compression in Modern Large-scale Model Training Process
Hanzhen Zhao (Peking University), Zhouchen Lin (Peking University)
CompressionOptimizationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A low-rank projection gradient compression method called SEPARATE is proposed for large-scale LLM training.
Separation Power of Equivariant Neural Networks
Marco Pacini (University of Trento), Gabriele Santin (University of Venice)
Convolutional Neural Network
🎯 What it does: The theoretical analysis of the separation capability of equivariant neural networks (such as convolutional networks, permutation-invariant networks, circular CNNs, etc.) is conducted, providing a complete characterization that can distinguish inputs and exploring the impact of hyperparameters and network structure on separation capability.
SePer: Measure Retrieval Utility Through The Lens Of Semantic Perplexity Reduction
Lu Dai (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
RetrievalOptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: An automatic evaluation metric based on Semantic Perplexity (SePer) and its difference (ΔSePer) is proposed to measure the actual knowledge gain from retrieval increments in Retrieval-Augmented Generation (RAG). It estimates the belief distribution changes of large language models (LLMs) through sampling and semantic clustering.
Seq-VCR: Preventing Collapse in Intermediate Transformer Representations for Enhanced Reasoning
Md Rifat Arefin (Mila), Christopher Pal (New York University)
TransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposes the Seq-VCR regularization method, which combines pauses (tokens) to enhance the diversity of intermediate representations in Transformers, addressing the representation collapse problem in multi-step reasoning tasks.
Sequential Controlled Langevin Diffusions
Junhua Chen (University of Cambridge), Anima Anandkumar (California Institute of Technology)
Diffusion modelStochastic Differential Equation
🎯 What it does: A sampler named Sequential Controlled Langevin Diffusion (SCLD) has been developed, which combines continuous-time SMC with diffusion sampling to achieve an end-to-end trainable sampling method.
Sequential Stochastic Combinatorial Optimization Using Hierarchal Reinforcement Learning
Xinsong Feng (University of California Los Angeles), Haipeng Chen (William & Mary)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A hierarchical reinforcement learning framework named WS-option is proposed to solve the Sequential Stochastic Combinatorial Optimization (SSCO) problem, particularly for instances of adaptive influence maximization and path planning.
SeRA: Self-Reviewing and Alignment of LLMs using Implicit Reward Margins
Jongwoo Ko (Korea Advanced Institute of Science and Technology), Aram Galstyan (Amazon)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A self-review and alignment method based on implicit reward margins, SeRA, has been developed, which utilizes the reward information from the policy model itself for offline sample selection and self-generated preference pairs, thereby improving the alignment effects of direct alignment algorithms (such as DPO, IPO, SLiC-HF).
Severing Spurious Correlations with Data Pruning
Varun Mulchandani (North Carolina State University), Jung-Eun Kim (North Carolina State University)
Domain AdaptationData-Centric LearningImage
🎯 What it does: A training data pruning method is proposed that does not require domain knowledge or labeled information. By removing a small number of hard-to-learn samples that have a strong impact on model learning, the method weakens the model's dependence on spurious correlations and enhances the model's robustness in out-of-distribution scenarios.
SFESS: Score Function Estimators for $k$-Subset Sampling
Klas Wijk (KTH Royal Institute of Technology), Hossein Azizpour (KTH Royal Institute of Technology)
Score-based ModelImage
🎯 What it does: A score function-based k-subset sampling estimator (SFESS) is proposed, achieving unbiased gradients and the generation of k-hot samples.