ICLR 2025 Papers — Page 33
International Conference on Learning Representations · 3704 papers
SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints
Miruna Cretu (University of Cambridge), Pietro Lio
GenerationDrug DiscoveryGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: A generative model named SynFlowNet is proposed, which constructs molecules using chemical reaction templates and purchasable materials, and recursively generates target molecules in the reaction space through GFlowNet.
SynQ: Accurate Zero-shot Quantization by Synthesis-aware Fine-tuning
Minjun Kim (Seoul National University), U Kang (Seoul National University)
ClassificationData SynthesisOptimizationKnowledge DistillationTransformerImage
🎯 What it does: A zero-shot quantization method called SYNQ is proposed without any real data, which fine-tunes the quantization model and significantly improves its accuracy by applying low-pass filtering on synthetic data, class activation map alignment, and soft label strategies.
Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo
João Loula (Massachusetts Institute of Technology), Timothy J. O'Donnell (McGill)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Using the Sequential Monte Carlo (SMC) method to implement syntactic and semantic constraint generation in large language models allows for the flexible incorporation of different constraint signals during inference and dynamic adjustment of computational resources.
Synthesizing Programmatic Reinforcement Learning Policies with Large Language Model Guided Search
Max Liu (National Taiwan University), Shao-Hua Sun (National Taiwan University)
Large Language ModelReinforcement Learning
🎯 What it does: A programmatic reinforcement learning framework called LLM-GS is proposed, which utilizes large language models (LLM) to generate initial programs and further optimize them through Scheduled Hill Climbing, significantly improving sample efficiency.
Synthesizing Realistic fMRI: A Physiological Dynamics-Driven Hierarchical Diffusion Model for Efficient fMRI Acquisition
Yufan Hu (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
GenerationData SynthesisRecurrent Neural NetworkDiffusion modelTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This study proposes a diffusion model based on hypergraph hierarchical structure and multifractal dynamics for generating realistic functional magnetic resonance imaging (fMRI) time series, aiming to reduce scanning time and improve data quality.
Synthetic continued pretraining
Zitong Yang (Stanford University), Tatsunori Hashimoto (Stanford University)
Data SynthesisRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This study investigates a method to transform small-scale proprietary corpora into large-scale diversified synthetic corpora through Synthetic Continued Pretraining, and continues pretraining on this data to achieve efficient learning of small domain knowledge by pretrained language models.
Synthio: Augmenting Small-Scale Audio Classification Datasets with Synthetic Data
Sreyan Ghosh (NVIDIA), Dinesh Manocha (NVIDIA)
ClassificationData SynthesisLarge Language ModelDiffusion modelAudio
🎯 What it does: Synthio enhances data by generating synthetic audio using a text-to-audio diffusion model on a small-scale audio classification dataset.
SysBench: Can LLMs Follow System Message?
Yanzhao Qin (Peking University), Bin CUI
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A SysBench benchmark was constructed to systematically evaluate the ability of large language models to follow system messages in multi-turn dialogues.
SysCaps: Language Interfaces for Simulation Surrogates of Complex Systems
Patrick Emami (National Renewable Energy Lab), Truc Nguyen (National Renewable Energy Lab)
Recurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringMultimodalityTime Series
🎯 What it does: This paper proposes SysCaps—a lightweight multimodal agent model that combines system properties described in natural language with temporal inputs, using LLM to automatically generate these language descriptions for training simulation agents.
System 1.x: Learning to Balance Fast and Slow Planning with Language Models
Swarnadeep Saha (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a system called System1.x Planner that can dynamically switch between fast and slow planning modes in long-term planning tasks, significantly reducing search costs while maintaining accuracy.
Systematic Outliers in Large Language Models
Yongqi An (Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences), Jinqiao Wang (Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences)
TransformerLarge Language ModelText
🎯 What it does: A systematic analysis of activation, weight, and attention outliers in LLMs is conducted, demonstrating that they originate from softmax and act as context-aware scaling factors.
Systematic Relational Reasoning With Epistemic Graph Neural Networks
Irtaza Khalid (Cardiff University), Steven Schockaert (Cardiff University)
Graph Neural NetworkContrastive LearningGraph
🎯 What it does: The Epistemic GNN (EpiGNN) model is proposed for systematic relationship reasoning, capable of learning reasoning rules during training and generalizing to longer reasoning chains during testing.
Systems with Switching Causal Relations: A Meta-Causal Perspective
Moritz Willig (Technical University of Darmstadt), Kristian Kersting (Eindhoven University of Technology)
Tabular
🎯 What it does: This paper proposes and empirically validates the concept of 'Meta-Causal Models (MCM)' to capture the structural switching of causal graphs in different contexts or states. By extending traditional Structural Causal Models (SCM) through meta-causal states, an algorithm is provided for inferring meta-causal states from observational data. The EM+LO-RANSAC method is used to estimate the number of different mechanisms in a two-variable linear model, and the visualization and interpretation of meta-causal states are demonstrated in an idealized stress-fatigue system. Finally, MCM is compared with classical SCM, and the performance of the algorithm is evaluated on synthetic data.
T-JEPA: Augmentation-Free Self-Supervised Learning for Tabular Data
Hugo Thimonier (Emobot), Bich-Liên DOAN
ClassificationAnomaly DetectionRepresentation LearningTransformerTabular
🎯 What it does: A novel augmentation-free self-supervised pre-training method T-JEPA is proposed for feature representation learning of tabular data, enhancing model performance in downstream classification/regression tasks.
T-Stitch: Accelerating Sampling in Pre-Trained Diffusion Models with Trajectory Stitching
Zizheng Pan (Monash University), Anima Anandkumar (California Institute of Technology)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: A 'Trajectory Stitching' technique is proposed, which uses a pre-trained smaller diffusion model to replace a larger model in the early sampling phase, thereby accelerating the generation process.
T2V-Turbo-v2: Enhancing Video Model Post-Training through Data, Reward, and Conditional Guidance Design
Jiachen Li (University of California Santa Barbara), William Yang Wang (University of California Santa Barbara)
GenerationData SynthesisOptimizationKnowledge DistillationTransformerReinforcement LearningVideoTextOrdinary Differential Equation
🎯 What it does: In the post-training phase of the pre-trained text-to-video model, T2V-Turbo-v2 was introduced by integrating high-quality data, reward model feedback, and conditional guidance into the consistency distillation process, enhancing video generation quality and text alignment.
T2V2: A Unified Non-Autoregressive Model for Speech Recognition and Synthesis via Multitask Learning
Nabarun Goswami (University of Tokyo), Tatsuya Harada (RIKEN)
RecognitionGenerationTransformerAudio
🎯 What it does: A unified non-autoregressive model T2V2 is proposed, capable of simultaneously performing automatic speech recognition (ASR) and text-to-speech (TTS) tasks;
TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation
Juntong Shi (University of Southern California), Jure Leskovec (Stanford University)
GenerationData SynthesisTransformerDiffusion modelTabular
🎯 What it does: Designed and implemented a joint hybrid-type diffusion model TABDIFF for generating high-quality tabular data, supporting seamless integration of continuous and discrete features;
TabM: Advancing tabular deep learning with parameter-efficient ensembling
Yury Gorishniy (Yandex), Artem Babenko (Yandex)
ClassificationComputational EfficiencySupervised Fine-TuningTabular
🎯 What it does: This paper studies and proposes a TabM model based on parameter-efficient ensemble for supervised learning on tabular data, and systematically evaluates its performance and efficiency on 46 public datasets.
TabReD: Analyzing Pitfalls and Filling the Gaps in Tabular Deep Learning Benchmarks
Ivan Rubachev (Yandex), Artem Babenko (Yandex)
Data-Centric LearningTransformerTabularBenchmark
🎯 What it does: This paper proposes the TabReD benchmark, which constructs and evaluates eight industry-level datasets for tabular data in real industrial scenarios, particularly focusing on time drift and feature richness.
TabWak: A Watermark for Tabular Diffusion Models
Chaoyi Zhu (TU Delft), Lydia Y. Chen (TU Delft)
Data SynthesisAdversarial AttackDiffusion modelTabular
🎯 What it does: Designed and implemented the TabWak watermarking scheme, which embeds invisible watermarks during the sampling phase of table diffusion models, supporting row-level detection while maintaining data quality.
Tackling Data Corruption in Offline Reinforcement Learning via Sequence Modeling
Jiawei Xu (Chinese University of Hong Kong), Lei Han (Tencent Robotics X)
TransformerReinforcement LearningSequential
🎯 What it does: This study investigates the robustness of offline reinforcement learning when data is corrupted by random or adversarial noise, and proposes a robust sequence modeling algorithm called RDT.
TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models
Makoto Shing (Sakana AI), Takuya Akiba (Sakana AI)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes Temporally Adaptive Interpolated Distillation (TAID), a framework for dynamically interpolating the teacher and student distributions during knowledge distillation from large language models to smaller models.
Tailoring Mixup to Data for Calibration
Quentin Bouniot (Télécom Paris), Florence d'Alché-Buc (Télécom Paris)
ClassificationDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A variant of Mixup based on similarity kernel is proposed—Similarity Kernel Mixup, which dynamically adjusts the linear interpolation coefficient to reduce the deviation of mixed samples from the original class manifold, thereby enhancing the model's predictive performance and confidence calibration.
Talking Turns: Benchmarking Audio Foundation Models on Turn-Taking Dynamics
Siddhant Arora (Carnegie Mellon University), Shinji Watanabe (Apple)
Supervised Fine-TuningBenchmarkAudio
🎯 What it does: Evaluate the turn-taking ability of audio foundation models (FM) in natural conversations and propose a novel evaluation protocol based on a supervised turn prediction model; simultaneously conduct user studies and benchmarking on existing full-duplex and cascaded speech dialogue systems.
Taming Overconfidence in LLMs: Reward Calibration in RLHF
Jixuan Leng (Carnegie Mellon University), Jiaxin Huang (Washington University in St. Louis)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: To address the issue of overconfidence in LLMs trained with RLHF, two calibration methods without additional golden labels are proposed—PPO-M (calibrated reward model) and PPO-C (calibrated reward computation), which are seamlessly integrated into the existing PPO framework.
Taming Transformer Without Using Learning Rate Warmup
Xianbiao Qi (Intellifusion Inc.), Rong Xiao
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: A new optimization strategy (AdamW 2) is proposed, which dynamically adjusts the learning rate to suppress spectral energy concentration during the training process of Transformers, achieving stable training without the use of learning rate warm-up.
Tamper-Resistant Safeguards for Open-Weight LLMs
Rishub Tamirisa (University of Illinois Urbana-Champaign), Mantas Mazeika (Center for AI Safety)
Adversarial AttackMeta LearningLarge Language ModelTextBiomedical Data
🎯 What it does: The research focuses on the anti-tampering security mechanism for large language models (LLMs) with open weights, proposing the TAR method to ensure that the model remains secure even after its weights are modified, preventing easy removal of denial and knowledge constraints.
TANGO: Co-Speech Gesture Video Reenactment with Hierarchical Audio Motion Embedding and Diffusion Interpolation
Haiyang Liu (University of Tokyo), Takafumi Taketomi (CyberAgent)
GenerationRetrievalTransformerDiffusion modelContrastive LearningOptical FlowVideoMultimodalityAudio
🎯 What it does: Proposes the TANGO framework, which utilizes retrieval + diffusion generators to create high-quality body posture videos synchronized with target speech from a single speaker's video.
Targeted Attack Improves Protection against Unauthorized Diffusion Customization
Boyang Zheng (Shanghai Jiao Tong University), Xiaoyu Wu (Shanghai Jiao Tong University)
GenerationAdversarial AttackDiffusion modelImageStochastic Differential Equation
🎯 What it does: A targeted attack-based protection method for diffusion models, ACE and its variants, is proposed to prevent unauthorized customization of diffusion models.
TASAR: Transfer-based Attack on Skeletal Action Recognition
Yunfeng Diao (Hefei University of Technology), He Wang (UCL Centre for Artificial Intelligence, Department of Computer Science, University College London)
RecognitionAdversarial AttackGraph Neural NetworkTransformerVideoBenchmark
🎯 What it does: A transfer-based adversarial attack method for skeleton action recognition, TASAR, is proposed, and the first large-scale S-HAR robustness benchmark, RobustBenchHAR, is constructed.
Task Descriptors Help Transformers Learn Linear Models In-Context
Ruomin Huang (Duke University), Rong Ge (Duke University)
OptimizationTransformerTabular
🎯 What it does: This study investigates the contextual learning ability of Transformer in linear regression problems with task descriptors and provides a theoretical proof that gradient flow training converges to a global optimal solution. It verifies how a layer of Linear Self-Attention Network (LSA) utilizes task descriptors to standardize inputs and improve predictions.
Task-Adaptive Pretrained Language Models via Clustered-Importance Sampling
David Grangier (Apple), Pierre Ablin (Apple)
TransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: A method is proposed to adapt general pre-training data for task-specific purposes using Clustering-based Importance Sampling (CRISP) in the context of scarce specialized data, thereby training a small specialized language model.
TaskGalaxy: Scaling Multi-modal Instruction Fine-tuning with Tens of Thousands Vision Task Types
Jiankang Chen (Kuaishou Technology), Di ZHANG
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: TaskGalaxy is proposed, a large-scale multimodal instruction fine-tuning dataset that includes 19,227 hierarchical task types and 413,648 samples, aimed at addressing the issue of insufficient task diversity in existing datasets.
TAU-106K: A New Dataset for Comprehensive Understanding of Traffic Accident
Yixuan Zhou (Alibaba Cloud Computing), Heng Tao Shen (Tongji University)
RecognitionObject DetectionAutonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageVideoMultimodality
🎯 What it does: This paper constructs a large-scale multimodal traffic accident dataset TAU-106K, and based on this, trains a specialized multimodal large language model TABot to achieve tasks such as traffic accident identification, description, spatiotemporal localization, and spatial localization.
TC-MoE: Augmenting Mixture of Experts with Ternary Expert Choice
Shen Yan (Peking University), Zhouchen Lin (Peking University)
Mixture of ExpertsText
🎯 What it does: By expanding the expert space by multiplying each expert by the ternary set {-1, 0, 1}, TC-MoE is constructed to improve expert activation efficiency without changing the routing mechanism.
TD-Paint: Faster Diffusion Inpainting Through Time-Aware Pixel Conditioning
Tsiry Mayet (INSA Rouen Normandie), Clement Chatelain
RestorationDiffusion modelImage
🎯 What it does: Proposes TD-Paint, a method for image inpainting that accelerates diffusion models through pixel-level temporal conditioning;
TDDBench: A Benchmark for Training data detection
Zhihao Zhu (University of Science and Technology of China), Defu Lian (University of Science and Technology of China)
Computational EfficiencyData-Centric LearningConvolutional Neural NetworkTransformerLarge Language ModelImageTextMultimodalityTabularBenchmark
🎯 What it does: A large-scale training data detection benchmark named TDDBench has been constructed to systematically evaluate the performance and resource consumption of 21 TDD algorithms across 13 cross-modal datasets (images, tables, text) and 41 target models (including LLMs).
Teaching Human Behavior Improves Content Understanding Abilities Of VLMs
Somesh Kumar Singh, Balaji Krishnamurthy
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoMultimodality
🎯 What it does: Incorporating user behavior (likes, comments, playback curves) as supervisory signals when training visual-language models enhances their content understanding capabilities.
Teaching LLMs How to Learn with Contextual Fine-Tuning
Younwoo Choi (University of Toronto), Rahul Krishnan
OptimizationKnowledge DistillationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataFinance Related
🎯 What it does: This study proposes a new method called Contextual Fine-Tuning (CFT), aimed at rapidly fine-tuning large language models (LLMs) by mimicking human learning strategies through prompts, to enhance their understanding and reasoning abilities in specific domains.
TEASER: Token Enhanced Spatial Modeling for Expressions Reconstruction
Yunfei Liu (International Digital Economy Academy), Yu Li (International Digital Economy Academy)
GenerationPose EstimationConvolutional Neural NetworkGenerative Adversarial NetworkImageVideo
🎯 What it does: A TEASER method is proposed, utilizing a combination of mixed explicit 3DMM parameters and implicit multi-scale appearance tokens, along with a token-guided neural renderer, to achieve high-precision 3D facial expression reconstruction from a single image.
TeaserGen: Generating Teasers for Long Documentaries
Weihan Xu (Duke University), Hao-Wen Dong (University of Michigan)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoTextMultimodalityAudio
🎯 What it does: This paper proposes a two-stage documentary trailer generation system: first, a pre-trained large language model (GPT-4o) generates engaging trailer narration based on the documentary's voiceover, and then visual segments corresponding to the narration are selected using a language-visual matching method, ultimately generating a complete trailer video.
Tell me about yourself: LLMs are aware of their learned behaviors
Jan Betley (Truthful AI), Owain Evans (Truthful AI)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates whether LLMs can accurately articulate their implicit behavioral strategies without contextual prompts after fine-tuning, and examines the impact of different roles and triggers;
TempMe: Video Temporal Token Merging for Efficient Text-Video Retrieval
Leqi Shen (Tsinghua University), Guiguang Ding (Tsinghua University)
RetrievalCompressionComputational EfficiencyTransformerContrastive LearningVideoText
🎯 What it does: A text-video retrieval method named TempMe is proposed, which freezes the model based on the pre-trained CLIP and only trains the LoRA parameters. It utilizes a stepwise multi-granularity spatiotemporal token merging (ImgMe and ClipMe) to compress video tokens, thereby reducing computational complexity.
Temporal Difference Learning: Why It Can Be Fast and How It Will Be Faster
Patrick Schnell (Technical University of Munich), Nils Thuerey (Technical University of Munich)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper theoretically analyzes the eigenvalues and rotational properties of non-gradient temporal difference (TD) learning iterative operators, revealing the reasons why TD converges faster than gradient descent (GD), and proposes an update rule GDS-TDM that combines the sign of GD with the magnitude of TD, balancing convergence and speed.
Temporal Flexibility in Spiking Neural Networks: Towards Generalization Across Time Steps and Deployment Friendliness
Kangrui Du (University of Electronic Science and Technology of China), Shi Gu (University of Electronic Science and Technology of China)
Spiking Neural NetworkImage
🎯 What it does: Proposed and implemented Mixed Temporal Training (MTT), allowing for the random allocation of different time steps during training, enabling SNNs to adapt to various time step configurations during inference.
Temporal Heterogeneous Graph Generation with Privacy, Utility, and Efficiency
Xinyu He (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
GenerationSafty and PrivacyGraph Neural NetworkTransformerGenerative Adversarial NetworkGraphTime Series
🎯 What it does: A temporal heterogeneous graph generation framework named THEPUFF is proposed, which provides differential privacy protection while maintaining data usability.
Temporal Reasoning Transfer from Text to Video
Lei Li (University of Hong Kong), Qi Liu (University of Hong Kong)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoText
🎯 What it does: This paper diagnoses the bottleneck of video large language models (Video LLM) in temporal reasoning and proposes a text-only temporal reasoning transfer framework (T3) that enhances the temporal understanding ability of LLMs by generating diverse temporal reasoning question-answer data from image-text pairs, significantly improving the performance of Video LLMs on video temporal reasoning tasks.
TEOChat: A Large Vision-Language Assistant for Temporal Earth Observation Data
Jeremy Andrew Irvin (Stanford University), Stefano Ermon (Stanford University)
ClassificationRecognitionSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTime SeriesSequential
🎯 What it does: This paper proposes and implements the first visual language assistant TEOChat capable of processing time series Earth observation images. It constructs a spatiotemporal instruction tuning dataset TEOChatlas, containing 554,071 instruction-image-response triplets based on four mainstream EO datasets, to train TEOChat for various spatial and temporal reasoning tasks.
Test of Time: A Benchmark for Evaluating LLMs on Temporal Reasoning
Bahare Fatemi (Google Research), Bryan Perozzi (Google)
Large Language ModelTextGraphBenchmark
🎯 What it does: This paper proposes a novel temporal reasoning benchmark called Test of Time (ToT), which is divided into two main sub-tasks: semantic reasoning and arithmetic reasoning. It employs controllable graph structure generation and templated questioning to ensure that the evaluation is not influenced by the model's prior knowledge.
Test-Time Adaptation for Combating Missing Modalities in Egocentric Videos
Merey Ramazanova (Center of Excellence in Generative AI), Motasem Alfarra (Center of Excellence in Generative AI)
RecognitionDomain AdaptationKnowledge DistillationContrastive LearningVideoMultimodality
🎯 What it does: To address the multimodal missing problem in perspective videos, a method called MiDl is proposed, which performs online self-supervised adaptation of the pre-trained model only during testing;
Test-time Adaptation for Cross-modal Retrieval with Query Shift
Haobin Li (Sichuan University), Mouxing Yang (Sichuan University)
RetrievalDomain AdaptationContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes an online testing adaptive method called TCR for Query Shift, aimed at improving the performance of cross-modal retrieval models in scenarios with inconsistent real distributions.
Test-time Adaptation for Image Compression with Distribution Regularization
Kecheng Chen (City University of Hong Kong), Haoliang Li (City University of Hong Kong)
CompressionDomain AdaptationImageMagnetic Resonance Imaging
🎯 What it does: A high-level latent variable fine-tuning method for image compression during cross-domain testing is proposed, which improves joint probability approximation through distribution regularization and Bayesian approximation, enhancing rate-distortion performance.
Test-time Adaptation for Regression by Subspace Alignment
Kazuki Adachi (NTT Corporation), Tomoki Hamagami (Yokohama National University)
Domain AdaptationImageTabular
🎯 What it does: This paper proposes a Test-Time Adaptation (TTA) method for regression models called SSA (Significant-Subspace Alignment), which fine-tunes a regression network pre-trained on the source domain using only unlabeled target data under an unknown target distribution.
Test-time Alignment of Diffusion Models without Reward Over-optimization
Sunwoo Kim (Seoul National University), Dongmin Park (KRAFTON)
GenerationOptimizationReinforcement LearningDiffusion modelImage
🎯 What it does: A testing method based on Sequential Monte Carlo (SMC) sampling called DAS is proposed to align diffusion models with arbitrary reward functions, optimizing rewards while maintaining diversity and cross-reward generalization, thus avoiding over-optimization.
Test-Time Ensemble via Linear Mode Connectivity: A Path to Better Adaptation
Byungjai Kim (Samsung Electronics), Eunho Yang (Korea Advanced Institute of Science and Technology)
Domain AdaptationKnowledge DistillationImage
🎯 What it does: A test-time ensemble method based on linear pattern connectivity (TTE) is proposed, which enhances model representation and prevents collapse through adaptive weighted averaging and dropout.
TestGenEval: A Real World Unit Test Generation and Test Completion Benchmark
Kush Jain (Carnegie Mellon University), Baptiste Roziere (Meta AI)
GenerationAI Code AssistantLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Designed and released the TESTGENEVAL benchmark to evaluate the unit test generation and test completion capabilities of large language models in real-world Python projects, providing execution metrics such as coverage and mutation scores;
TetSphere Splatting: Representing High-Quality Geometry with Lagrangian Volumetric Meshes
Minghao Guo (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
GenerationOptimizationGaussian SplattingMesh
🎯 What it does: This paper proposes TetSphere splatting, a Lagrangian geometric representation method that utilizes deformable tetrahedral spheres to generate high-quality 3D shapes.
Text-to-Image Rectified Flow as Plug-and-Play Priors
Xiaofeng Yang (Nanyang Technological University), Guosheng Lin (Nanyang Technological University)
GenerationData SynthesisDiffusion modelFlow-based ModelRectified FlowImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes the use of the pre-trained Rectified Flow model as a plug-and-play prior for text-to-3D generation, image inversion, and editing;
Text2PDE: Latent Diffusion Models for Accessible Physics Simulation
Anthony Zhou (Carnegie Mellon University), Amir Barati Farimani (Carnegie Mellon University)
GenerationCompressionExplainability and InterpretabilityTransformerDiffusion modelAuto EncoderMeshPhysics Related
🎯 What it does: A physical simulation framework based on latent diffusion models has been designed and implemented, capable of compressing, encoding, and generating complete spatiotemporal solutions for PDE data on arbitrary grids (regular or irregular).
Text4Seg: Reimagining Image Segmentation as Text Generation
Mengcheng Lan (Nanyang Technological University), Wayne Zhang (SenseTime Research)
SegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: Reformulating the image segmentation problem as a text generation task, a decoder-free framework called Text4Seg is proposed, which implements segmentation using semantic descriptors and row-level run-length encoding (R-RLE).
TexTailor: Customized Text-aligned Texturing via Effective Resampling
Suin Lee (KAIST), Daeshik Kim
GenerationData SynthesisDiffusion modelMesh
🎯 What it does: This work proposes the TexTailor method, which can generate consistent and view-seamless textures for 3D mesh models based on text descriptions.
TFG-Flow: Training-free Guidance in Multimodal Generative Flow
Haowei Lin (Peking University), Jianzhu Ma (Tsinghua University)
GenerationDrug DiscoveryFlow-based ModelMultimodality
🎯 What it does: This paper proposes TFG-Flow, a training-free guidance method for multi-modal flow models that allows generated molecules to meet specified properties without additional training.
TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics
Lu Yi (Renmin University of China), Zengfeng Huang (Fudan University)
Recommendation SystemGraph Neural NetworkGraphTime SeriesSequentialBenchmark
🎯 What it does: A new TGB-Seq benchmark is proposed, which includes 8 temporal graph datasets with low redundancy and complex sequence dynamics, and systematically evaluates the performance of existing temporal GNNs on this benchmark.
The "Law'' of the Unconscious Contrastive Learner: Probabilistic Alignment of Unpaired Modalities
Yongwei Che (Princeton University), Benjamin Eysenbach (Princeton University)
Representation LearningReinforcement LearningContrastive LearningMultimodalityAudio
🎯 What it does: A theoretical framework based on Bayesian inference is proposed for probabilistic reasoning between unseen cross-modal (A↔C) during training, providing a closed-form condition for direct comparison ('Law of Unconscious Contrast Learners') and a Monte Carlo LogSumExp approximation method when this law is not satisfied.
The 3D-PC: a benchmark for visual perspective taking in humans and machines
Drew Linsley (Brown University), Thomas Serre (University of California-Irvine)
RecognitionDepth EstimationPrompt EngineeringGaussian SplattingImageVideoBenchmark
🎯 What it does: A 3D-PC (visual perspective orientation challenge) based on 3D Gaussian Splatting was constructed, and this challenge was used to simultaneously evaluate the performance of humans and 327 types of DNNs on three 3D perception tasks (depth ordering, VPT-Basic, VPT-Strategy).
The adaptive complexity of parallelized log-concave sampling
Huanjian Zhou (University of Tokyo), Masashi Sugiyama (RIKEN AIP)
🎯 What it does: The study investigates the adaptive (parallel) complexity of log-concave sampling algorithms in high dimensions, providing lower bounds for both unconstrained and box-constrained sampling.
The AdEMAMix Optimizer: Better, Faster, Older
Matteo Pagliardini (École Polytechnique Fédérale de Lausanne), David Grangier (Apple)
OptimizationTransformerImageText
🎯 What it does: This paper proposes a new optimizer called AdEMAMix, which combines fast and slow exponential moving averages (EMA) to more effectively utilize historical gradients.
The Belief State Transformer
Edward S. Hu (Microsoft Research), John Langford (Microsoft Research)
GenerationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a Belief State Transformer that simultaneously utilizes prefix and suffix information, capable of predicting the next word and the previous word under both forward and backward encoding, achieving more effective goal-conditioned generation.
The Breakdown of Gaussian Universality in Classification of High-dimensional Linear Factor Mixtures
Xiaoyi MAI, Zhenyu Liao (Huazhong University of Science and Technology)
ClassificationOptimizationTabular
🎯 What it does: Conduct a precise empirical risk minimization (ridge regularization) performance analysis for classification problems under high-dimensional linear factor mixed models (LFMM), revealing the mechanism of Gaussian universality failure in high dimensions.
The Case for Cleaner Biosignals: High-fidelity Neural Compressor Enables Transfer from Cleaner iEEG to Noisier EEG
Francesco S. Carzaniga (IBM Research), Abbas Rahimi (Bern University)
CompressionAuto EncoderTime SeriesBiomedical Data
🎯 What it does: A high-fidelity deep learning compressor named BrainCodec is proposed for lossless compression of EEG and iEEG signals.
The Complexity of Two-Team Polymatrix Games with Independent Adversaries
Alexandros Hollender (University of Oxford), Sai Ganesh Nagarajan (Zuse Institute Berlin)
Optimization
🎯 What it does: This study investigates the scenario of two teams in zero-sum polymatrix games where players can be divided into two teams with no internal coordination (i.e., independent opponents), proving that finding (approximate) Nash equilibria in this model is computationally difficult.
The Computational Complexity of Circuit Discovery for Inner Interpretability
Federico Adolfi (Max Planck Society), Todd Wareham (Memorial University of Newfoundland)
Explainability and InterpretabilityComputational Efficiency
🎯 What it does: This paper systematically studies the problem of circuit discovery in the internal interpretability of neural networks from a theoretical perspective, constructing a conceptual framework and classifying the complexity of various queries.
The Computational Complexity of Positive Non-Clashing Teaching in Graphs
Robert Ganian (TU Wien), Mathis Rocton (TU Wien)
Graph
🎯 What it does: This study investigates the computational complexity of calculating the forward non-conflicting teaching dimension in graphs, covering both strict and non-strict cases, and proves various hardness and solvability results.
The Crucial Role of Samplers in Online Direct Preference Optimization
Ruizhe Shi (Tsinghua University), Simon Shaolei Du
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: Analyze the convergence rate of DPO under different samplers, prove that the mixed sampler achieves quadratic convergence, and design practical samplers based on theory, followed by validation in language model alignment tasks.
The Crystal Ball Hypothesis in diffusion models: Anticipating object positions from initial noise
Yuanhao Ban (University of California), Minhao Cheng (Pennsylvania State University)
Object DetectionGenerationDiffusion modelImage
🎯 What it does: The study found that there are 'trigger patches' in the initial Gaussian noise of diffusion models, which can determine the position of objects in the generated images, and trained a detector that can predict these positions before image generation.
The Directionality of Optimization Trajectories in Neural Networks
Sidak Pal Singh (ETH Zurich), Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
OptimizationConvolutional Neural NetworkTransformerLarge Language ModelImage
🎯 What it does: This study investigates the directional characteristics of optimization trajectories during the training process of neural networks, proposing the Trajectory Map and related quantitative metrics, and conducting experimental analysis across various tasks.
The Effectiveness of Curvature-Based Rewiring and the Role of Hyperparameters in GNNs Revisited
Floriano Tori (Vrije Universiteit Brussel), Vincent Ginis (Harvard University)
OptimizationHyperparameter SearchGraph Neural NetworkGraphBenchmark
🎯 What it does: Evaluated the effectiveness of the discrete curvature-based graph reconnection method on real-world benchmark datasets and analyzed the impact of hyperparameter search on performance improvement.
The Foundations of Tokenization: Statistical and Computational Concerns
Juan Luis Gastaldi (ETH Zurich), Ryan Cotterell (ETH Zurich)
🎯 What it does: This paper proposes a unified formal framework for defining and analyzing tokenizers by constructing the category of stochastic maps, and provides the necessary and sufficient conditions for the tokenizer to maintain the consistency of language model estimation.
The Geometry of Categorical and Hierarchical Concepts in Large Language Models
Kiho Park (University of Chicago), Victor Veitch (University of Chicago)
Large Language ModelText
🎯 What it does: This paper clarifies how semantic hierarchies in large language models are encoded through orthogonal geometric structures by extending binary concepts from directions to vectors and modeling multi-dimensional concepts as polyhedra.
The Hidden Cost of Waiting for Accurate Predictions
Ali Shirali (University of California), Rediet Abebe (Max Planck Institute for Intelligent Systems)
OptimizationReinforcement LearningTabular
🎯 What it does: The study investigates the impact of collecting more observations on ranking quality and intervention timing in prediction-driven resource allocation, providing theoretical explanations and optimal allocation algorithms.
The Hyperfitting Phenomenon: Sharpening and Stabilizing LLMs for Open-Ended Text Generation
Fredrik Carlsson (RISE Research Institutes of Sweden), Joakim Nivre (Uppsala University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: This paper explores the performance of pre-trained large language models through hyperfitting on extremely small datasets and observes their performance in open-ended text generation.
The impact of allocation strategies in subset learning on the expressive power of neural networks
Ofir Schlisselberg (Tel Aviv University), Ran Darshan (Tel Aviv University)
Knowledge DistillationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This study investigates how to allocate limited learnable weights in neural networks to maximize the network's expressive power, and introduces matching probability (MP) as a metric.
The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws
Tian Jin (Massachusetts Institute of Technology), Gintare Karolina Dziugaite (Google)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the effects of Sparse Pre-Training in large language models and proposes a unified scaling law for dense/sparse training.
The KoLMogorov Test: Compression by Code Generation
Ori Yoran (Tel Aviv University), Taco Cohen (Meta AI)
CompressionLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequential
🎯 What it does: The KOLMOGOROV-TEST (KT) is proposed, transforming the compression task into a code generation task to evaluate the reasoning and search capabilities of LLMs.
The Labyrinth of Links: Navigating the Associative Maze of Multi-modal LLMs
Hong Li (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Jiao Tong University)
Large Language ModelPrompt EngineeringMixture of ExpertsTextMultimodalityBenchmark
🎯 What it does: This study constructs an unannotated benchmark for relational tasks to evaluate the capabilities of multimodal large language models (MLLMs) in relational reasoning.
The Last Iterate Advantage: Empirical Auditing and Principled Heuristic Analysis of Differentially Private SGD
Milad Nasr (Google DeepMind), Andreas Terzis (Google DeepMind)
OptimizationSafty and PrivacyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: A heuristic privacy analysis method for DP-SGD that only discloses the last iterate is proposed, and its effectiveness in practical deep learning training is validated.
The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling
Andre Cornman (Tatta Bio), Yunha Hwang (Tatta Bio)
TransformerLarge Language ModelMultimodality
🎯 What it does: An open mixed-modal meta-genomic corpus (OMG) consisting of 3.1 Tbp and 3.3 billion protein-coding sequences has been created and made publicly available, and the first mixed-modal gene language model gLM2 has been trained based on this.
The Optimization Landscape of SGD Across the Feature Learning Strength
Alexander Atanasov (Harvard University), Cengiz Pehlevan (Harvard University)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: Under the μP parameterization of wide networks, the interaction between the output scaling factor γ (feature learning intensity) and the learning rate η is systematically explored, a γ–η phase diagram is drawn, the optimization dynamics under different ranges of γ are analyzed, and it is verified that training with large γ can achieve or exceed the performance of γ=1.
The Pitfalls of Memorization: When Memorization Hurts Generalization
Reza Bayat (Mila), Pascal Vincent (Mila)
ClassificationDomain AdaptationImageText
🎯 What it does: This paper studies the interaction between 'memorization' and 'false association' that occurs during the learning process of neural networks, proving that the combination of the two can lead to model generalization failure under distribution shift. It proposes a novel training method called Memorization-Aware Training (MAT), which uses the predicted probabilities of retained samples to adaptively shift logits, thereby guiding the model to learn robust features across distributions.
The Power of LLM-Generated Synthetic Data for Stance Detection in Online Political Discussions
Stefan Sylvius Wagner (Heinrich Heine University Düsseldorf), Stefan Harmeling (Technical University Dortmund)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper combines synthetic reviews generated by offline LLM with traditional BERT model fine-tuning to enhance the stance detection performance in online political discussions.
The Ramanujan Library - Automated Discovery on the Hypergraph of Integer Relations
Itay Beit Halachmi, Ido Kaminer (Technion - Israel Institute of Technology)
Large Language Model
🎯 What it does: A 'Ramanujan Library' structured as a hypergraph has been constructed, and 75 previously unknown constant relationships have been discovered in this library using an automated search for integer relations method.
The Rise and Down of Babel Tower: Investigating the Evolution Process of Multilingual Code Large Language Model
Jiawei Chen (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences), Le Sun (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Conducted experimental research on the evolution of multilingual capabilities in code LLM during the pre-training process, proposing and validating the Babel Tower hypothesis.
THE ROBUSTNESS OF DIFFERENTIABLE CAUSAL DISCOVERY IN MISSPECIFIED SCENARIOS
Huiyang Yi (Southeast University), Wenwu Yu (Southeast University)
GraphTabularBenchmark
🎯 What it does: This paper systematically evaluates the robustness of 12 mainstream causal discovery algorithms under 8 scenarios of model assumption violations and provides a large-scale experimental benchmark.
The Same but Different: Structural Similarities and Differences in Multilingual Language Modeling
Ruochen Zhang (Brown University), Ellie Pavlick (Brown University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper explores whether large language models share internal circuits in cross-linguistic contexts through a mechanistic interpretability analysis of two tasks (indirect object recognition and tense generation) in English and Chinese.
The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities
Zhaofeng Wu (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)
TransformerLarge Language ModelTextMultimodalityAudio
🎯 What it does: This paper investigates whether language models share a semantic representation space under different languages and multimodal inputs, and experimentally verifies this hypothesis.
The Superposition of Diffusion Models Using the Itô Density Estimator
Marta Skreta (University of Toronto), Kirill Neklyudov (Mila - Quebec AI Institute)
GenerationData SynthesisProtein Structure PredictionDiffusion modelImageMultimodalityBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Without retraining, new multimodal or multi-condition samples are generated by 'superposition' of existing pre-trained diffusion models during the inference phase;
The Unreasonable Ineffectiveness of the Deeper Layers
Andrey Gromov (Meta), Dan Roberts
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Through layer pruning and lightweight fine-tuning, the study of LLM knowledge storage found that deep layers can be deleted without affecting QA performance;
The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning
Youssef Allouah (École Polytechnique Fédérale de Lausanne), Sanmi Koyejo (Stanford University)
OptimizationComputational Efficiency
🎯 What it does: This paper analyzes the fundamental utility, time and space complexity trade-offs of machine forgetting, and proposes a new algorithm to handle out-of-distribution forgetting data.
The Value of Sensory Information to a Robot
Arjun Krishna (University of Pennsylvania), Dinesh Jayaraman (University of Pennsylvania)
Robotic IntelligenceReinforcement LearningTabular
🎯 What it does: This paper proposes a value information-based metric (VoSI) for quantitatively assessing the necessity of robots perceiving environmental information at different time points while executing lookahead strategies. It also analyzes the impact of perceived information on various robotic tasks and strategy architectures through experimental analysis.