ICLR 2026 Papers with Code — Page 14

Omni-Weather: A Unified Multimodal Model for Weather Radar Understanding and Generation

Zhiwang Zhou (Tongji University), LEI BAI

🎯 What it does: Propose Omni-Weather, a unified multimodal foundation model that simultaneously performs weather forecasting (e.g., radar nowcasting, satellite radar inversion) and weather understanding (e.g., radar report generation, question answering)

OmniCT: Towards a Unified Slice-Volume LVLM for Comprehensive CT Analysis

Tianwei Lin (Zhejiang University), Yingda Xia (DAMO Academy, Alibaba Group)

🎯 What it does: Proposed a unified CT slice-volume large vision-language model, OmniCT, which integrates 2D slice and 3D volume information for comprehensive CT image analysis.

OmniMouse: Scaling properties of multi-modal, multi-task Brain Models on 150B Neural Tokens

Konstantin Friedrich Willeke (Stanford University), Andreas S. Tolias (Stanford University)

🎯 What it does: Trained and evaluated a large-scale, multimodal, multitask Transformer model called OmniMouse for predicting mouse visual cortex neuron activity, behavior, and conditional responses to video stimuli.

OmniPortrait: Fine-Grained Personalized Portrait Synthesis via Pivotal Optimization

Dongxu Yue (Tsinghua University), Chun Yuan (Tsinghua University)

🎯 What it does: Proposed OmniPortrait, a fine-grained identity customization portrait synthesis framework based on diffusion;

OmniVideoBench: Towards Audio-Visual Understanding Evaluation for Omni MLLMs

Caorui Li (Nanjing University), Jiaheng Liu (Nanjing University)

🎯 What it does: Propose OmniVideoBench, an audio-visual collaborative reasoning benchmark for multimodal large language models, containing 628 real videos, 1000 high-quality question-answer pairs, and their atomic reasoning chains.

On Discriminative vs. Generative classifiers: Rethinking MLLMs for Action Understanding

Zhanzhong Pang (National University of Singapore), Angela Yao (National University of Singapore)

🎯 What it does: This paper compares two learning methods for multi-modal large language models (MLLM) in closed-set action understanding tasks—generative classifiers and discriminative classifiers—and proposes a Generative-Aided Discriminative (GAD) framework to integrate their advantages.

On Entropy Control in LLM-RL Algorithms

Han Shen (Ant Group)

🎯 What it does: This paper investigates the dilemma of entropy control in large language model reinforcement learning (LLM-RL). To address the poor performance of traditional entropy regularization in scenarios with large vocabularies and sparse optimal outputs, we propose an adaptive entropy regularization method called AEnt;

On Predictability of Reinforcement Learning Dynamics for Large Language Models

Cai Yuchen (University of Science and Technology of China), Junfeng Fang (National University of Singapore)

🎯 What it does: Conduct a systematic analysis of parameter updates in large language models (LLMs) during reinforcement learning (RL) training, revealing that the main performance gains are attributable to a rank-1 subspace of the update matrix, which evolves linearly with training; based on this, propose the AlphaRL framework, which predicts the final updates using an early training window, thereby significantly accelerating RL training.

On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations

Jianing Guo (Beihang University), Simin Li (Beihang University)

🎯 What it does: Systematic evaluation of Vision-Language-Action (VLA) models under 17 perturbations across four modalities (action, observation, environment, language), and propose RobustVLA which enhances multimodal robustness by imposing semantic consistency constraints between worst-case action noise in outputs and inputs.

On the Design of KL-Regularized Policy Gradient Algorithms for LLM Reasoning

Yifan Zhang (Princeton University), Andrew C Yao (Tsinghua University)

🎯 What it does: Proposed a unified KL regularization policy gradient framework (RPG), achieving stable offline reinforcement learning on LLM inference tasks.

On the Design of One-step Diffusion via Shortcutting Flow Paths

Haitao Lin (Westlake University), Stan Z. Li (Westlake University)

🎯 What it does: Propose a general 'one-step shortcut diffusion' design framework, decomposing components such as discrete/continuous, path, and time sampling, and improving training methods within this framework to ultimately achieve high-quality single-step generation.

On the Expressive Power of GNNs for Boolean Satisfiability

Saku Peltonen (ETH Zürich), Roger Wattenhofer (ETH Zürich)

🎯 What it does: This paper studies the expressive power of graph neural networks (GNNs) in Boolean satisfiability problems, proving that the full Weisfeiler–Leman hierarchy cannot distinguish between satisfiable and unsatisfiable instances, and analyzing the distinguishability of different instance families.

On the Generalization Capacities of MLLMs for Spatial Intelligence

Gongjie Zhang (DAMO Academy, Alibaba Group), Ran Xu (DAMO Academy, Alibaba Group)

🎯 What it does: Built a camera-aware multimodal large language model (Camera-Aware MLLM) specifically for spatial intelligence tasks, addressing the geometric ambiguity issue in traditional RGB-only MLLMs during cross-camera generalization.

On the identifiability of causal graphs with multiple environments

Francesco Montagna (Institute of Science and Technology Austria)

🎯 What it does: Proposed a method to uniquely identify any nonlinear structural causal graph using observational data from only two different environments under the Gaussian noise assumption.

On the Interaction of Compressibility and Adversarial Robustness

Melih Barsbey (Imperial College London), Tolga Birdal (Imperial College London)

🎯 What it does: This paper investigates the impact of structural compression (neuron-level compression and spectral compression) on the adversarial robustness of neural networks, establishing theoretical upper bounds and verifying them through various experiments.

On the Predictive Power of Representation Dispersion in Language Models

Yanhong Li (Allen Institute for AI), Jiawei Zhou (Stony Brook University)

🎯 What it does: Investigate and demonstrate that the representational diversity (average cosine distance) of language models is negatively correlated with predictive performance (perplexity), and apply it to unlabeled diagnosis, model and layer selection, and enhance training effectiveness through an auxiliary 'push-away' loss.

On the Shelf Life of Fine-Tuned LLM-Judges: Future-Proofing, Backward-Compatibility, and Question Generalization

Janvijay Singh (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)

🎯 What it does: Investigated the long-term usability of fine-tuned LLM discriminators (LLM-as-a-judge), systematically evaluating their performance when facing stronger model-generated text, old model text, and new unseen questions.

On The Surprising Effectiveness of a Single Global Merging in Decentralized Learning

Tongtian Zhu (Zhejiang University), Can Wang (Zhejiang University)

🎯 What it does: Studied and empirically analyzed time allocation strategies for communication budgets in decentralized learning, finding that concentrating communication in the later stages of training with a single global merge significantly improves model performance.

On the Tension Between Optimality and Adversarial Robustness in Policy Optimization

Haoran Li (University of Chinese Academy of Sciences), Nan Jiang (University of Illinois Urbana-Champaign)

🎯 What it does: Studied the trade-off between standard policy optimization and adversarial robust policy optimization in reinforcement learning, and proposed the BARPO two-layer framework to balance optimality and robustness.

On the Theoretical Limitations of Embedding-Based Retrieval

Orion Weller (Google DeepMind), Jinhyuk Lee (Google DeepMind)

🎯 What it does: This paper investigates the theoretical limits of unidirectional vector embedding models in retrieval tasks and proves that, given a certain dimensionality, there exist certain document combinations that cannot appear as retrieval results. Subsequently, the theory is validated through optimal vector optimization experiments, leading to the construction of the LIMIT dataset based on this theory. The dataset was tested on various state-of-the-art embedding models, revealing that even with high dimensions, achieving ideal recall remains challenging. Finally, the paper discusses alternative approaches such as multi-vector, sparse, and cross-encoders.

On the Wings of Imagination: Conflicting Script-based Multi-role Framework for Humor Caption Generation

Wenbo Shang (Hong Kong Baptist University), Xin Huang (Hong Kong Baptist University)

🎯 What it does: Proposed a multi-role LLM framework called HOMER based on the GTVH theory to generate humorous captions for images.

On-the-Fly Adaptation to Quantization: Configuration-Aware LoRA for Efficient Fine-Tuning of Quantized LLMs

Rongguang Ye (Southern University of Science and Technology), Edith C. H. Ngai (University of Hong Kong)

🎯 What it does: Propose CoA-LoRA, a configuration-aware LoRA tuning method that can dynamically generate low-rank adapters according to arbitrary quantization configurations without repeated fine-tuning, improving deployment efficiency of quantized LLMs.

One Model for All Tasks: Leveraging Efficient World Models in Multi-Task Planning

Yuan Pu (Shanghai Artificial Intelligence Laboratory), Hongsheng Li (Chinese University of Hong Kong)

🎯 What it does: Proposed the ScaleZero unified world model and combined it with Dynamic Parameter Scaling (DPS) to achieve efficient learning for multi-task planning

One protein is all you need

Anton Bushuiev (Czech Technical University), Josef Sivic (Massachusetts Institute of Technology)

🎯 What it does: Propose a self-supervised on-the-fly customization method called ProteinTTT, which can instantly fine-tune existing protein language models with only a single protein sequence, thereby improving the accuracy of structure, compatibility, and function prediction.

One step further with Monte-Carlo sampler to guide diffusion better

Minsi Ren (Westlake University), Tailin Wu (Westlake University)

🎯 What it does: Proposed an ABMS method that improves conditional guidance in training-agnostic diffusion models by incorporating additional backward denoising steps and Monte-Carlo sampling, thereby reducing estimation errors and enhancing generation quality.

One-Prompt Strikes Back: Sparse Mixture of Experts for Prompt-based Continual Learning

Minh Le (Trivita AI), Nhat Ho (Northeastern University)

🎯 What it does: Propose SMoPE, a prompt-based continual learning framework that combines sparse Mixture of Experts (MoE) with Prefix Tuning;

One-Step Flow for Image Super-Resolution with Tunable Fidelity-Realism Trade-offs

Yuanzhi Zhu (Nanjing University), Kai Zhang (Nanjing University)

🎯 What it does: Proposed OFTSR, a flow-based single-step image super-resolution framework that can generate high-resolution images with adjustable realism and fidelity in a single forward pass.

Online Minimization of Polarization and Disagreement via Low-Rank Matrix Bandits

Federico Cinus (Intesa Sanpaolo AI Research), Francesco Bonchi (Intesa Sanpaolo AI Research)

🎯 What it does: Proposes an online low-rank matrix bandit method to minimize polarization and disagreement under the Friedkin-Johnsen model in the absence of prior knowledge about individual's innate opinions.

Online Navigation Refinement: Achieving Lane-Level Guidance by Associating Standard-Definition and Online Perception Maps

Jiaxu Wan (BUAA), Yifan Yang (BUAA)

🎯 What it does: Proposed the Online Navigation Refinement (ONR) task, achieving the transition from road-level navigation to lane-level navigation by associating standard definition (SD) maps with real-time perception (OP) maps.

Only Brains Align with Brains: Cross-Region Alignment Patterns Expose Limits of Normative Models

Larissa Höfling (University Clinics Tübingen), Katrin Franke (University Clinics Tübingen)

🎯 What it does: This paper benchmarks 47 visual models on the BOLD Moments video fMRI dataset, finding that many models achieve similar scores on traditional point-to-point alignment metrics (RSA/LP); to address this limitation, we propose Cross-Region Alignment Pattern Analysis (APA), which distinguishes models that are truly brain-similar by comparing alignment pattern similarity between model layers and brain regions.

OpenEstimate: Evaluating LLMs on Reasoning Under Uncertainty with Real-World Data

Alana Marzoev (MIT), Jacob Andreas (MIT)

🎯 What it does: Designed and implemented the OPENESTIMATE benchmark to evaluate probability estimation and uncertainty reasoning of language models under real-world conditional statistics.

Opponent Shaping in LLM Agents

Marta Emili García Segura, Mirco Musolesi (University College London)

🎯 What it does: Studied opponent shaping of large language model agents in multi-agent environments and proposed the ShapeLLM method, validated through experiments in various games.

Optimal Brain Restoration for Joint Quantization and Sparsification of LLMs

Hang Guo (ETH Zurich), Yawei Li (ETH Zurich)

🎯 What it does: Proposed the Optimal Brain Restoration (OBR) framework, achieving joint compression of quantization and sparsification, and efficiently compressing large language models without requiring additional training.

Optimal Sparsity of Mixture-of-Experts Language Models for Reasoning Tasks

Taishi Nakamura (Institute of Science Tokyo), Rio Yokota (Institute of Science Tokyo)

🎯 What it does: Study the impact of sparsity on memory and inference capabilities of Mixture-of-Experts (MoE) language models under a fixed computational budget, systematically investigating the relationship between total parameters, active parameters, top-k routing, and other factors on pre-training loss and downstream task accuracy.

Optimal transport unlocks end-to-end learning for single-molecule localization

Romain Seailles (Ecole normale superieure), Julien Mairal (Universite Grenoble Alpes)

🎯 What it does: Proposed an end-to-end deep learning method based on optimal transport for single-molecule localization microscopy (SMLM), leveraging the optical system model through an iterative network to achieve detection and localization without NMS and with a single threshold;

OptimalThinkingBench: Evaluating Over and Underthinking in LLMs

Pranjal Aggarwal (FAIR at Meta), Swarnadeep Saha (FAIR at Meta)

🎯 What it does: Proposed a unified evaluation framework called OptimalThinkingBench, which can simultaneously measure over-thinking of LLMs on simple tasks and under-thinking on complex tasks, along with corresponding evaluation metrics;

Optimistic Task Inference for Behavior Foundation Models

Thomas Rupf (ETH Zürich), Andreas Krause (ETH Zürich)

🎯 What it does: Propose an online task inference framework OpTI-BFM, which actively collects reward data by interacting with the environment, forms confidence ellipsoids using success features (USFs) in the behavior foundation model (BFM) and least squares task embedding, and achieves rapid inference through optimistic UCB strategy;

Optimizing ID Consistency in Multimodal Large Models: Facial Restoration via Alignment, Entanglement, and Disentanglement

Yuran Dong (Wuhan University), Mang Ye (Wuhan University)

🎯 What it does: Propose a training-free, plug-and-play framework called EditedID to achieve facial identity-consistent editing and restoration in multimodal large models (e.g., GPT-4o+, Flux.1 Kontext, etc.).

Orak: A Foundational Benchmark for Training and Evaluating LLM Agents on Diverse Video Games

Dongmin Park (KRAFTON), Jaewoong Cho (KRAFTON)

🎯 What it does: Proposed the Orak benchmark to evaluate LLM performance in 12 video games, providing a pluggable interface and fine-grained assessment dimensions.

Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory

Xuan Zhang (Texas A&M University), Xiaofeng Qian (Texas A&M University)

🎯 What it does: Propose the OrbEvo model, which utilizes SO(2)-equivariant graph transformers to learn the evolution of electronic wave functions over time in real-time TDDFT, and can predict wave function coefficients, dipole moments, and absorption spectra.

OrchestrationBench: LLM-Driven Agentic Planning and Tool Use in Multi-Domain Scenarios

Aelim Ahn (Kakao Corp), Jihoon kang

🎯 What it does: Proposed OrchestrationBench, a multilingual (English-Korean) LLM orchestration evaluation benchmark targeting multi-domain, multi-step tasks;

OrderDP: A Theoretically Guaranteed Lossless Dynamic Data Pruning Framework

Chenhan Jin (Chinese University of Hong Kong), Tieyong Zeng (Beijing Normal-Hong Kong Baptist University)

🎯 What it does: Proposed OrderDP, a dynamic data pruning framework based on random candidate batches and loss sorting, which can significantly reduce training costs while maintaining near-lossless performance;

OrthAlign: Orthogonal Subspace Decomposition for Non-Interfering Multi-Objective Alignment

Liang Lin (China Telecom), Kun Wang (Nanyang Technological University)

🎯 What it does: Propose OrthAlign, which addresses parameter conflicts in multi-objective preference alignment through orthogonal subspace decomposition, ensuring gradient updates for different objectives occur within mutually non-interfering subspaces.

Otters: An Energy-Efficient Spiking Transformer via Optical Time-to-First-Spike Encoding

Zhanglu Yan (National University of Singapore), Weng-Fai Wong (Westlake University)

🎯 What it does: Developed a Time-to-First-Spike (TTFS) Spiking Transformer model called Otters based on optoelectronic hetero-synapses, achieving significant energy efficiency improvements through hardware-software co-design.

Out of the Memory Barrier: A Highly Memory-Efficient Training System for LLMs with Million-Token Contexts

Wenhao Li (Xiamen University), Rongrong Ji (Xiamen University)

🎯 What it does: Propose a long-context LLM training framework named OOMB, which adopts block recursive training and activation recomputation to significantly reduce activation memory usage;

Out-of-Distribution Graph Models Merging

Yidi Wang (Great Bay University), Xiao Luo (University of Wisconsin-Madison)

🎯 What it does: Studied how to merge graph neural network (GNN) models pre-trained on different domains without using source domain data, constructing a unified model capable of generalizing to new domains.

Overcoming Joint Intractability with Lossless Hierarchical Speculative Decoding

Yuxuan Zhou (Independent Researcher), Zhi-Qi Cheng (University Of Washington)

🎯 What it does: Proposed Hierarchical Speculative Decoding (HSD), significantly increasing the number of acceptable tokens and inference speed while maintaining the integrity of the target distribution.

Overlap-Adaptive Regularization for Conditional Average Treatment Effect Estimation

Valentyn Melnychuk (LMU Munich), Stefan Feuerriegel (LMU Munich)

🎯 What it does: Proposed an adaptive regularization method called OAR based on overlap weights to improve the CATE estimation performance of meta-learners in low-overlap regions.

Overlap-weighted orthogonal meta-learner for treatment effect estimation over time

Konstantin Hess (LMU Munich), Stefan Feuerriegel (LMU Munich)

🎯 What it does: Propose an overlapping weighted orthogonal meta-learner (WO-learner) for estimating time-varying heterogeneous treatment effects.

Oversmoothing, "Oversquashing'', Heterophily, Long-Range, and more: Demystifying Common Beliefs in Graph Machine Learning

Adrian Arnaiz-Rodriguez (ELLIS Alicante), Federico Errica (NEC Laboratories Europe)

🎯 What it does: Systematically question nine common beliefs in graph machine learning (over-smoothing, over-compression, homogeneity/heterogeneity, and long-distance tasks) and refute them using theoretical and experimental counterexamples.

Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling

Shuyang Jiang (Fudan University), Yu Wang (Shanghai Jiao Tong University)

🎯 What it does: To address the overthinking problem in large-scale reasoning models during the generation process, this paper proposes the DECS framework, which significantly compresses the inference path length by separating token-level rewards from batch scheduling.

Overtone: Cyclic Patch Modulation for Clean, Efficient, and Flexible Physics Emulators

Payel Mukhopadhyay (University of Cambridge), Miles Cranmer (University of Cambridge)

🎯 What it does: This paper proposes the Overtone framework, which improves the long-term prediction accuracy of PDE surrogates and achieves adjustable computational resource allocation by dynamically controlling the patch size during the autoregressive inference process.

OWL : Geometry-Aware Spatial Reasoning for Audio Large Language Models

Subrata Biswas (Worcester Polytechnic Institute), Bashima Islam (Worcester Polytechnic Institute)

🎯 What it does: Designed and implemented a large language model for spatial audio, OWL, along with its geometry-aware encoder, SAGE, and constructed a large-scale dataset, BiDepth, supporting multi-source localization and multi-step spatial reasoning.

OwlEye: Zero-Shot Learner for Cross-Domain Graph Data Anomaly Detection

Lecheng Zheng (Virginia Tech), Jingrui He (University Of Illinois Urbana Champaign)

🎯 What it does: Developed a zero-shot cross-domain graph anomaly detection framework named OWLEYE, capable of detecting anomalies on unseen graphs without requiring retraining;

P-GenRM: Personalized Generative Reward Model with Test-time User-based Scaling

Pinyi Zhang (East China Normal University), Kai Zhang (East China Normal University)

🎯 What it does: Propose P-GenRM, a personalized generative reward model that converts user mixed preference signals into a structured evaluation chain (including dynamic user persona and scoring rubrics), and improves scoring accuracy through user-level and prototype-level scaling during testing.

P2P: Automated Paper-to-Poster Generation and Fine-Grained Benchmark

Tao Sun (ByteDance), Zhoujun Li (Intelligent Strong Technology Co.,Ltd)

🎯 What it does: Designed and implemented a multi-agent framework named P2P for automatically converting research papers into high-quality, HTML-rendered academic posters, along with the corresponding instruction dataset P2PINSTRUCT and a dual evaluation benchmark combining fine-grained (based on manual checklists) and general (XGBoost-simulated subjective aesthetics) criteria named P2PEVAL.

P3D: Highly Scalable 3D Neural Surrogates for Physics Simulations with Global Context

Benjamin Holzschuh (Technical University of Munich), Nils Thuerey (Technical University of Munich)

🎯 What it does: This paper proposes the P3D hybrid CNN-Transformer architecture as a neural agent for high-resolution 3D physical simulations, demonstrating its performance on various PDEs, turbulence, and channel flow tasks.

PaAno: Patch-Based Representation Learning for Time-Series Anomaly Detection

Jinju Park (Sungkyunkwan University), Seokho Kang (Sungkyunkwan University)

🎯 What it does: Developed a lightweight time series anomaly detection method called PaAno based on patch representation learning.

PALC: Preference Alignment via Logit Calibration

SANGHYUN LEE, Hoh Peter In (Korea University)

🎯 What it does: Propose the PALC framework, achieving test-time preference alignment for LLMs in the vocabulary (logit) space via calibration vectors, by simply inserting a lightweight Calibration Module outside the frozen model.

Panda: A pretrained forecast model for chaotic dynamics

Jeffrey B. Lai (University of Texas at Austin), William Gilpin (University of Texas at Austin)

🎯 What it does: We propose a pre-trained prediction model called Panda specifically for chaotic dynamics, and construct a large-scale dataset containing approximately 20,000 chaotic ODE trajectories using evolutionary algorithms;

Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning

Minju Seo (KAIST), Sung Ju Hwang (KAIST)

🎯 What it does: Propose a multi-stage, multi-agent LLM framework called PaperCoder, which automates the generation of complete executable code repositories from machine learning papers.

Paradigm Shift of GNN Explainer from Label Space to Prototypical Representation Space

Jun Yin (Central South University), Chengqi Zhang (Beijing University of Posts and Telecommunications)

🎯 What it does: Propose to migrate GNN interpreter optimization from the graph label space to the graph representation space, and design a general IDEA framework that utilizes prototype graph representation space for discovering and aligning interpretable subgraphs;

Parallel Token Prediction for Language Models

Felix Draxler (University of California Irvine), Stephan Mandt (University of California Irvine)

🎯 What it does: Propose Parallel Token Prediction (PTP), enabling language models to generate multiple tokens at once, eliminating the sequential bottleneck in autoregressive inference.

ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMs

Wonjun Kang (FuriosaAI), Kangwook Lee (UW-Madison)

🎯 What it does: Proposed the PARALLELBENCH evaluation framework for parallel decoding, and empirically validated through information-theoretic analysis the quality degradation of diffusion LLMs (dLLMs) during parallel decoding;

Parameter-Efficient Reinforcement Learning using Prefix Optimization

Itamar Rocha Filho (Harvard University), Samy Jelassi (Harvard University)

🎯 What it does: Studied methods that perform reinforcement learning optimization only on the top k tokens before the answer, proposing Prefix-Clustering and Prefix-RL, and verified that they can significantly improve the accuracy of large language models on mathematical reasoning tasks.

Parameters vs. Context: Fine-Grained Control of Knowledge Reliance in Language Models

Baolong Bi (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences)

🎯 What it does: Proposed the CK-PLUG plugin, which dynamically controls large language models' dependence on parameter knowledge and retrieved context during retrieval-augmented generation (RAG) through token-level probability modulation.

ParaRNN: Unlocking Parallel Training of Nonlinear RNNs for Large Language Models

Federico Danieli (Apple), Luca Zappella (Apple)

🎯 What it does: The ParaRNN framework achieves the first large-scale language model-level training of nonlinear RNNs by reconstructing the recurrence relations of nonlinear RNNs as a system of nonlinear equations and implementing sequence parallelization using Newton iteration combined with parallel prefix scan.

PARD: Accelerating LLM Inference with Low‑Cost PARallel Draft Model Adaptation

Zihao An (Advanced Micro Devices Inc), Emad Barsoum (Advanced Micro Devices Inc)

🎯 What it does: Propose PARD, a target-agnostic parallel draft model that significantly reduces the adaptation cost of LLM inference.

Part-level Semantic-guided Contrastive Learning for Fine-grained Visual Classification

Zhijian Lin (Xidian University), Hong Han (Xidian University)

🎯 What it does: Proposes the Part-level Semantic-guided Contrastive Learning (PSCL) framework for fine-grained visual classification, combining part localization, cross-scale multi-branch reasoning, and visual-language contrastive learning.

Partial Soft-Matching Distance For Neural Representational Comparison With Partial Unit Correspondence

Chaitanya Kapoor (University of California San Diego), Meenakshi Khosla (University of California San Diego)

🎯 What it does: Proposed a 'partial soft-matching' distance based on partial optimal transport to compare unit correspondences in neural networks or neural recordings, allowing partial unit mismatches to enhance robustness and interpretability.

Patch-as-Decodable-Token: Towards Unified Multi-Modal Vision Tasks in MLLMs

Yongyi Su (South China University of Technology), Xun Xu (Institute for Infocomm Research, A*STAR)

🎯 What it does: A unified framework PaDT was constructed by introducing Visual Reference Tokens (VRT) and a lightweight decoder, enabling multimodal large language models to directly generate text and visual outputs;

PathChat-SegR1: Reasoning Segmentation in Pathology via SO-GRPO

Zelin Liu (Shanghai Jiao Tong University), Lichi Zhang (Shanghai Jiao Tong University)

Pay Less Attention to Function Words for Free Robustness of Vision-Language Models

Qiwei Tian (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)

🎯 What it does: Propose Function-word De-Attention (FDA), which reduces the attention interference caused by function words by parallel computing and subtracting the attention between function words and images in multi-head cross-attention, thereby enhancing the robustness of visual-language models against cross-modal adversarial attacks.

PCB-Bench: Benchmarking LLMs for Printed Circuit Board Placement and Routing

Jindong Li (Hong Kong University of Science and Technology), Menglin Yang (Hong Kong University of Science and Technology)

🎯 What it does: This study constructs PCB-Bench—a comprehensive benchmark covering text reasoning, multimodal image-text reasoning, and real-world PCB design understanding—to evaluate the performance of large language models in PCB layout and routing tasks.

PCLR: Progressively Compressed LoRA for Multimodal Continual Instruction Tuning

Weicheng Meng (Harbin Institute of Technology), Yuan Xie (Shanghai Innovation Institute)

🎯 What it does: Propose PCLR, combining LoRA Rank Pool (fine-grained Mixture-of-Experts) with Compression-Integration-Learning (CIL) pipeline, which can balance stability, plasticity, and memory efficiency in multi-modal continuous instruction tuning.

PCPO: Proportionate Credit Policy Optimization for Preference Alignment of Image Generation Models

Jeongjae Lee (KAIST), Jong Chul Ye (KAIST)

🎯 What it does: Proposes the PCPO framework, addressing training instability and model collapse in text-to-image model alignment through proportional credit assignment.

PE-SGD: Differentially Private Deep Learning via Evolution of Gradient Subspace for Text

Tianyuan Zou (Tsinghua University), Sergey Yekhanin (Microsoft)

🎯 What it does: Propose a differential privacy training framework PE-SGD based on gradient projection and evolution, which can dynamically update the projection subspace using evolved synthetic data in scenarios with limited private data and tight privacy budgets, and achieve better gradient approximation by adding noise to the projection coefficients.

Peak-Return Greedy Slicing: Subtrajectory Selection for Transformer-based Offline RL

Zhiwei Xu, Bin Zhang (Institute of Automation Chinese Academy of Sciences)

🎯 What it does: Propose the PRGS framework, which explicitly selects and trains high-quality sub-trajectories through the MMD return estimator, greedy sub-trajectory splitting, and adaptive historical truncation to enhance the performance of Transformer-based offline reinforcement learning.

PEAR: Phase Entropy Aware Reward for Efficient Reasoning

Chen Huang (Singapore University of Technology and Design), Wenxuan Zhang (Singapore University of Technology and Design)

🎯 What it does: By incorporating the token entropy of the thinking phase and the answer phase into the reward function, large reasoning models are trained to automatically compress redundant steps during chain-of-thought generation, producing more concise answers.

Peng's Q($\lambda$) for Conservative Value Estimation in Offline Reinforcement Learning

Byeongchan Kim (Seoul National University), Min-hwan Oh (Seoul National University)

🎯 What it does: Proposed a conservative Peng's Q(λ)-based offline multi-step reinforcement learning algorithm called CPQL, which efficiently utilizes complete trajectories from offline data for value estimation.

PepBenchmark: A Standardized Benchmark for Peptide Machine Learning

Jiahui Zhang (Zhongguancun Academy), Yang Wang (University of Science and Technology of China)

🎯 What it does: Established PepBenchmark, a unified and reproducible machine learning benchmark for protein peptides, comprising three components: data, preprocessing, and evaluation.

PepTri: Tri-Guided All-Atom Diffusion for Peptide Design via Physics, Evolution, and Mutual Information

Ngoc-Quang Nguyen (AIGEN Sciences), Jaewoo Kang (AIGEN Sciences)

🎯 What it does: Proposed a tri-guide diffusion framework PepTri that co-designs in an SE(3)-equivariant latent space for generating physically stable, evolutionarily feasible, and sequence-structure consistent peptides.

Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward

Tong Xiao (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)

🎯 What it does: By introducing visual perception rewards in RLVR training, significantly enhancing the visual perception and reasoning capabilities of multimodal large language models (MLLM).

PerfGuard: A Performance-Aware Agent for Visual Content Generation

Zhipeng Chen (Beijing University of Posts and Telecommunications), Yi-Zhe Song (University of Surrey)

🎯 What it does: Propose the PerfGuard framework, utilizing a multi-dimensional performance assessment system (PASM), adaptive preference update (APU), and capacity-aligned planning optimization (CAPO) to achieve tool performance boundary modeling and task planning in visual content generation;

PERSISTENCE SPHERES: BI-CONTINUOUS REPRESENTATIONS OF PERSISTENCE DIAGRAMS.

Matteo Pegoraro (Universit' a della Svizzera Italiana)

🎯 What it does: Proposed a novel Persistence Spheres function representation, utilizing the support function of lift zonoids to map persistence diagrams to spherical functions, and provided explicit computational formulas and efficient implementations;

PersonaX: Multimodal Datasets with LLM-Inferred Behavior Traits

Loka Li (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)

🎯 What it does: Created two multimodal datasets (CelebPersona and AthlePersona) by inferring behavioral traits (Big Five) of public figures using LLMs, and combining them with facial image embeddings and structured biological information; simultaneously proposed a two-layer analytical framework, including statistical independence tests on structured data and causal representation learning based on multimodal multi-measurement data;

PerSpectra: A Scalable and Configurable Pluralist Benchmark of Perspectives from Arguments

Shangrui Nie (University of Bonn), Charles Welch (McMaster University)

🎯 What it does: Proposed PERSPECTRA, an expandable multi-perspective benchmark that integrates Kialo's structured debates with Reddit's natural language diversity;

Perturbation-Induced Linearization: Constructing Unlearnable Data with Solely Linear Classifiers

Jinlin Liu (Huazhong University of Science and Technology), Xiaojin Zhang (Huazhong University of Science and Technology)

🎯 What it does: Propose a method using a linear model to generate unlearnable examples, called Perturbation-Induced Linearity (PIL), to achieve data protection.

pFedMMA: Personalized Federated Fine-Tuning with Multi-Modal Adapter for Vision-Language Models

Sajjad Ghiasvand (UC Santa Barbara), Ramtin Pedarsani (UC Santa Barbara)

🎯 What it does: Proposes the pFedMMA framework, which personalizes large vision-language models through federated fine-tuning using multi-modal adapters.

PhaseFormer: From Patches to Phases for Efficient and Effective Time Series Forecasting

Yiming Niu (Beihang University), Yongxin Tong (Beihang University)

🎯 What it does: Propose phased tokenization and a lightweight PhaseFormer model for efficient time series prediction.

PHyCLIP: $\ell_1$-Product of Hyperbolic Factors Unifies Hierarchy and Compositionality in Vision-Language Representation Learning

Daiki Yoshikawa (Hokkaido University), Takashi Matsubara (Hokkaido University)

🎯 What it does: Proposed a vision-language model named PHyCLIP, which simultaneously learns hierarchical and compositional semantic structures using multiple hyperbolic factors in an ℓ1-product metric space.

Physics vs Distributions: Pareto Optimal Flow Matching with Physics Constraints

Giacomo Baldan (Technical University of Munich), Nils Thuerey (Technical University of Munich)

🎯 What it does: This paper proposes the Physics-Based Flow Matching (PBFM) method, seamlessly integrating physical constraints (PDE residuals or algebraic constraints) into flow matching generative models, achieving a Pareto optimal trade-off between physical accuracy and distribution consistency;

Physics-Informed Audio-Geometry-Grid Representation Learning for Universal Sound Source Localization

Min-Sang Baek (Hanyang University), Joon-Hyuk Chang (Hanyang University)

🎯 What it does: Proposes a framework based on audio-geometry-grid representation learning (AGG-RL) to achieve flexible grid and geometry-invariant sound source localization; further enhances generalization and interpretability by introducing physics-informed learnable non-uniform discrete Fourier transform (LNuDFT) and relative microphone position encoding (rMPE).

Physics-Inspired All-Pair Interaction Learning for 3D Dynamics Modeling

Kai Yang (Shanghai Jiao Tong University), Qitian Wu (Shanghai Jiao Tong University)

🎯 What it does: Propose the PAINET model, which utilizes a physics-inspired energy function to achieve fully symmetric interaction learning while maintaining SE(3) equivariance, addressing the limitations of traditional GNNs that only consider observed structures.

pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation

Hansheng Chen (Stanford University), Sai Bi (Adobe Research)

🎯 What it does: This paper proposes π-Flow, a few-step generation framework that uses network output policies for multi-step ODE integration.

PI-Light: Physics-Inspired Diffusion for Full-Image Relighting

Zhexin Liang (S-Lab Nanyang Technological University), Xingang Pan (Tencent)

🎯 What it does: Propose a physics-informed diffusion model called PI-Light to achieve full-image relighting, and provide two-stage inverse rendering and forward rendering.

PiCa: Parameter-Efficient Fine-Tuning with Column Space Projection

Junseo Hwang (Seoul National University), Taesup Kim (Seoul National University)

🎯 What it does: Propose a parameter-efficient fine-tuning method called PiCa based on principal column space projection of pre-trained weights, which balances gradient projection and weight sharing;

PICS: Pairwise Image Compositing with Spatial Interactions

Hang Zhou (University of Alberta), Li cheng

🎯 What it does: Propose a parallel paired image synthesis framework PICS, which can synthesize two objects simultaneously in a single forward pass while maintaining spatial interactions between objects and background consistency.

Pinet: Optimizing hard-constrained neural networks with orthogonal projection layers

Panagiotis D. Grontas (ETH Zürich), John Lygeros (ETH Zürich)

🎯 What it does: Propose an output layer called Π net, which enforces neural network outputs to satisfy convex constraints through projection, applicable to parameterized constrained optimization problems.

Pisces: Cryptography-based Private Retrieval-Augmented Generation with Dual-Path Retrieval

Xiaojian Liang (Ant International, Ant Group), Pu Duan (Ant International, Ant Group)

🎯 What it does: This paper designs and implements Pisces, a cryptography-based retrieval-augmented generation (RAG) framework that supports both semantic and lexical retrieval paths while preserving the privacy of queries and documents.