ICLR 2026 Papers — Page 35

PatchDNA: A Flexible and Biologically-Informed Alternative to Tokenization for DNA

Alice Del Vecchio (Relation Therapeutics), Cristian Regep (Relation Therapeutics)

🎯 What it does: Propose PatchDNA, which uses dynamic patching instead of traditional DNA vocabulary encoding, constructs a DNA language model based on the Byte Latent Transformer, and introduces conservation-guided patching and post-training re-patching mechanisms.

PatchRefiner V2: Fast and Lightweight Real-Domain High-Resolution Metric Depth Estimation

Zhenyu Li (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)

🎯 What it does: Propose the PatchRefiner V2 framework to achieve fast and lightweight high-resolution monocular depth estimation.

PateGAIL++: Utility Optimized Private Trajectory Generation with Imitation Learning

Yingjie Ma (San Diego State University), Joann Qiongna Chen (San Diego State University)

🎯 What it does: Propose a differential privacy framework PATEGAIL++ for mobile trajectory generation, achieving a better privacy-utility balance through adaptive noise injection based on trajectory sensitivity;

Path Channels and Plan Extension Kernels: a Mechanistic Description of Planning in a Sokoban RNN

Mohammad Taufeeque (FAR.AI), Adrià Garriga-Alonso (FAR.AI)

🎯 What it does: This paper reverse-engineers a Sokoban DRC (3,3) convolutional RNN trained with model-agnostic reinforcement learning, revealing that certain channels in the hidden state directly encode future paths, termed path channels, and elucidates the specific implementation of forward/backward path expansion and the winner-takes-all mechanism through convolution kernel analysis.

Path Matters: Unveiling Geometric Implicit Bias via Curvature-Aware Sparse View Optimization

Canran Xiao (Shenzhen Campus of Sun Yat-sen University, University of Hong Kong), Peilai Yu (Ludwig Maximilian University of Munich)

🎯 What it does: A complete method is proposed for 3D Gaussian Splatting (3DGS) reconstruction under sparse viewpoints. It first optimizes the camera trajectory using curvature-aware and smooth constraints, then synthesizes pseudo-views along the trajectory to enhance training data density, ultimately achieving more detailed geometric and image reconstruction.

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

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

Patronus: Interpretable Diffusion Models with Prototypes

Nina Weng (Technical University of Denmark), Siavash Bigdeli (Technical University of Denmark)

🎯 What it does: Designed and implemented an interpretable diffusion model called Patronus, embedding a prototype network into the diffusion model, and achieving 'what, when, where' explanations of the generation process through conditional guidance using prototype activation vectors.

Pay Attention to CTC: Fast and Robust Pseudo-Labelling for Unified Speech Recognition

Alexandros Haliassos (NatWest AI Research), Stavros Petridis (Imperial College London)

🎯 What it does: Propose USR 2.0 — a unified speech recognition framework that improves pseudo-label generation through CTC-driven teacher forcing and hybrid sampling techniques, significantly enhancing training efficiency and robustness to out-of-distribution inputs.

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.

PD$^{2}$GS: Part-Level Decoupling and Continuous Deformation of Articulated Objects via Gaussian Splatting

Haowen Wang (Anhui University), Jian Tang (Beijing Innovation Center of Humanoid Robotics)

🎯 What it does: A self-supervised 3D Gaussian Splatting framework called PD GS is proposed to model continuous deformation and achieve part-level decoupling for multi-part movable objects, supporting unsupervised estimation of geometric, appearance, and motion parameters.

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.

Pedagogically-Inspired Data Synthesis for Language Model Knowledge Distillation

Bowei He (Mohamed bin Zayed University of Artificial Intelligence), Chen Ma (City University of Hong Kong)

🎯 What it does: Proposes a three-stage (Identifier-Organizer-Adapter) data synthesis framework based on educational theory to efficiently distill knowledge from large language models (LLMs) into smaller models.

PEERING INTO THE UNKNOWN: ACTIVE VIEW SELECTION WITH NEURAL UNCERTAINTY MAPS FOR 3D RECONSTRUCTION

Zhengquan Zhang (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)

🎯 What it does: Propose a proactive view selection method called PUN, which uses a lightweight network UPNet to predict a neural uncertainty map (UMap), and aggregates the predicted UMaps to efficiently select the next optimal view, achieving high-quality 3D reconstruction using only half the number of views.

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-Aware Policy Optimization for Multimodal Reasoning

Zhenhailong Wang (University of Illinois Urbana Champaign), Heng Ji (University of Illinois Urbana Champaign)

🎯 What it does: Propose a new policy gradient algorithm called PAPO, which encourages multimodal language models to better rely on visual information during the reasoning process.

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;

PerFit: Exploring Personalization Shifts in Representation Space of LLMs

Jiahong Liu (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)

🎯 What it does: For personalized adaptation of large language models, the PerFit two-stage method directly fine-tunes in the hidden representation space.

PERK: Long-Context Reasoning as Parameter-Efficient Test-Time Learning

Zeming Chen (École Polytechnique Fédérale de Lausanne), Antoine Bosselut (École Polytechnique Fédérale de Lausanne)

🎯 What it does: Proposes PERK, a parameter-efficient inference method that compresses long context into a low-rank adapter through gradient updates during testing.

Permutation-Consistent Variational Encoding for Incomplete Multi-View Multi-Label Classification

Chengliang Liu (University of Macau), Jie Wen (Harbin Institute of Technology)

🎯 What it does: Proposed a permutation consistency variational encoding framework (PCVE) for multi-view multi-label classification under dual missingness (view missingness and label missingness).

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;

Person-Centric Annotations of LAION-400M: Auditing Bias and Its Transfer to Models

Leander Girrbach (Technical University of Munich), Zeynep Akata (Technical University of Munich)

🎯 What it does: Perform human-centric annotation on the complete LAION-400M dataset, generating approximately 200 million human bounding boxes, gender, and race/ethnicity labels, along with personalized text descriptions for each individual;

Persona Features Control Emergent Misalignment

Miles Wang, Daniel P Mossing

🎯 What it does: Investigate the widespread misleading behaviors exhibited by language models after fine-grained fine-tuning, and explore their mechanisms and mitigation methods.

PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra

Xiachong Feng (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)

🎯 What it does: This study proposes the PERSONA framework, enabling dynamic and composable personality control in large language models through activation space vectors without requiring gradient updates.

Personalized Collaborative Learning with Affinity-Based Variance Reduction

Chenyu Zhang (Massachusetts Institute of Technology), Navid Azizan (Massachusetts Institute of Technology)

🎯 What it does: This paper proposes a novel multi-agent collaborative learning framework (Personalized Collaborative Learning, PCL), which enables different agents to adaptively collaborate and improve learning efficiency while maintaining personalized objectives through Affinity-Based Variance Reduction (AffPCL).

Personalized Feature Translation for Expression Recognition: An Efficient Source-Free Domain Adaptation Method

Masoumeh Sharafi (ETS Montreal), Eric Granger (ETS Montreal)

🎯 What it does: This paper proposes a source-free domain adaptation method for facial expression recognition (FER) called SFDA-PFT, which achieves model personalization by leveraging personalized feature translation in the feature space under conditions where source data is unavailable and only target neutral expression data is available.

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.

Perturbed Dynamic Time Warping: A Probabilistic Framework and Generalized Variants

Xiangqian Sun (Xi'an Jiaotong-Liverpool University), Wei Zhang (Xi'an Jiaotong-Liverpool University)

🎯 What it does: Proposed Perturbed-DTW (Perturbed Dynamic Time Warping) and its generalization ns-DTW (nested soft-DTW), applying them to mean computation (Barry center) of time series, nearest centroid classification, and clustering tasks.

PETRI: Learning Unified Cell Embeddings from Unpaired Modalities via Early-Fusion Joint Reconstruction

Ryan W Conrad (insitro), Emily Fox (insitro)

🎯 What it does: A Transformer-based PETRI model with early fusion is constructed to learn unified cell embeddings from unpaired cell image and transcriptome data.

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.

PGRF-Net: A Prototype-Guided Relational Fusion Network for Diagnostic Multivariate Time-Series Anomaly Detection

Jahoon Jeong (Yonsei University), Hyunsoo Yoon (Yonsei University)

🎯 What it does: Propose a diagnostic-oriented multi-factor evidence extraction and gated fusion network (PGRF-Net) for multivariate time series anomaly detection.

Phantom-Data: Towards a General Subject-Consistent Video Generation Dataset

Zhuowei Chen (Bytedance), Xinglong Wu (Bytedance)

🎯 What it does: Proposes Phantom-Data, a cross-pair, general-object-oriented video generation training dataset, to address copy-paste issues and improve text alignment and video quality.

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.

PHAT: Modeling Period Heterogeneity for Multivariate Time Series Forecasting

Jiaming Ma (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)

🎯 What it does: Proposes PHAT, a Transformer model that captures heterogeneous cycles in multivariate time series, utilizing periodic bucket structures and positive-negative attention to model variations across different cycles.

Photon: Speedup Volume Understanding with Efficient Multimodal Large Language Models

Chengyu Fang (Tsinghua University), Minfeng Xu (Duke University)

🎯 What it does: Propose the Photon framework, combining a 3D vision encoder with an LLM decoder using instruction-conditioned token scheduling (ITS) and surrogate gradient propagation (SGP), to achieve efficient 3D medical image question answering;

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.

PhyScensis: Physics-Augmented LLM Agents for Complex Physical Scene Arrangement

Yian Wang (University Of Massachusetts Amherst), Chuang Gan (University Of Massachusetts Amherst)

🎯 What it does: Developed a physics scene generation framework called PhyScensis based on LLM agents, capable of automatically constructing high-density, complex, and physically feasible 3D interactive environments according to natural language instructions.

Physically Valid Biomolecular Interaction Modeling with Gauss-Seidel Projection

Siyuan Chen (University of British Columbia), Peter Yichen Chen (University of British Columbia)

🎯 What it does: Proposed a differentiable Gauss-Seidel projection module that embeds physical validity constraints into the training and inference processes of diffusion models for all atomic molecular interaction modeling, significantly reducing the required denoising steps.

Physically-Guided Optical Inversion Enable Non-Contact Side-Channel Attack on Isolated Screens

Zhiwen Zheng (Hangzhou Dianzi University), Xingru Huang (Hangzhou Dianzi University)

🎯 What it does: Propose a physics-constrained optical projection side-channel attack model called IR4Net, which reconstructs display screen content using wall-scattered light spot images under conditions without direct line-of-sight or RF monitoring.

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-Constrained Fine-Tuning of Flow-Matching Models for Generation and Inverse Problems

Jan Tauberschmidt (German Research Center for Artificial Intelligence), Andrew B. Duncan (Imperial College London)

🎯 What it does: Post-training on existing flow-matching generative models using weak-form PDE residuals as rewards to approximate physics-consistent solutions and hidden parameters.

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-Informed Inference Time Scaling for Solving High-Dimensional Partial Differential Equations

Zexi Fan (Peking University), Yiping Lu (Northwestern University)

🎯 What it does: Propose the SCaSML framework, which performs fine-tuning correction without retraining the pre-trained PDE approximator during inference stage through the defect correction method.

Physics-informed learning under mixing: How physical knowledge speeds up learning

Anna Scampicchio (Chalmers University of Technology), Melanie Zeilinger

🎯 What it does: The study derives convergence rates dependent on complexity for estimating vector-valued functions under mixed (non-i.i.d.) data using regularization methods with physical prior knowledge.

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.

PhysLLM: Harnessing Large Language Models for Cross-Modal Remote Physiological Sensing

Yiping Xie (Shenzhen University), Zitong YU

🎯 What it does: Proposes the PhysLLM framework, integrating large language models with rPPG-specific components to enable cross-modal remote physiological measurements.

Pi-CCA: Prompt-Invariant CCA Certificates for Replay-Free Continual Multimodal Learning

Jiayu Zhang (Peking University), Wenshuo Wang (South China University of Technology)

🎯 What it does: Propose a replay-free continual learning framework PI-CCA, which directly maintains the geometric invariance of vision-text alignment by utilizing dimension-constrained CCA certificates;

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;

PICABench: How Far are We from Physical Realistic Image Editing?

Yuandong Pu (Shanghai Jiao Tong University), Yihao Liu (Shanghai AI Laboratory)

🎯 What it does: Proposed PICABench to evaluate the physical realism of image editing, and constructed the PICAEval evaluation protocol along with the PICA-100K video generation dataset.

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.

PINFDiT: Energy-Based Physics-Informed Diffusion Transformers for General-purpose Time Series Tasks

Defu Cao (University of Southern California), Yan Liu (University of Southern California)

🎯 What it does: Proposed PINFDiT, a general-purpose temporal task model that injects physical constraints into diffusion Transformers;

PIRN: Prototypical-based Intra-modal Reconstruction with Normality Communication for Multi-modal Anomaly Detection.

YITING LI (Institute for Infocomm Research), Fayao Liu (Institute for Infocomm Research)

🎯 What it does: Propose the PIRN framework, achieving few-shot multimodal anomaly detection through prototype-driven internal reconstruction and cross-modal normality communication;

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.

Pitfalls in Evaluating Language Model Forecasters

Daniel Paleka (ETH Zurich), Florian Tramèr (ETH Zurich)

🎯 What it does: This paper systematically analyzes the evaluation methods of large language models in event prediction tasks, identifies various evaluation flaws such as time leakage, logical leakage, and retrieval bias, and proposes corresponding improvement recommendations.

Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling

Long Peng (USTC), Zheng-Jun Zha (USTC)

🎯 What it does: Propose the Pixel-to-Gaussian method, which directly reconstructs continuous high-resolution signals from low-resolution images using 2D Gaussian scattering, achieving arbitrary-scale super-resolution within a single model and significantly improving efficiency.

Pixel-Level Residual Diffusion Transformer: Scalable 3D CT Volume Generation

Zhenkai Zhang (University of Melbourne), Tom Drummond (University of Melbourne)

🎯 What it does: This paper proposes Pixel-Level Residual Diffusion Transformer (PRDiT), which directly generates high-resolution 3D CT volumes at the voxel level through a two-stage structure (local MLP denoising + global Transformer residual learning).

Pixel-Perfect Puppetry: Precision-Guided Enhancement for Face Image and Video Editing

Yan Li (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)

🎯 What it does: Proposed a unified FlowGuide framework that enables precise facial attribute editing in images and videos while preserving identity information and temporal coherence.

Pixel3DMM: Versatile Screen-Space Priors for Single-Image 3D Face Reconstruction

Simon Giebenhain (Technical University of Munich), Matthias Nießner (Technical University of Munich)

🎯 What it does: Propose the Pixel3DMM model, which uses ViT to predict per-pixel surface normals and UV coordinates, and optimizes the FLAME 3D facial model through these priors to achieve 3D reconstruction from a single image.

PixelCraft: A Multi-Agent system for High-Fidelity Visual Reasoning on Structured Images

Shuoshuo Zhang (Microsoft Research Asia), Rui Wang (Microsoft Research Asia)

🎯 What it does: Proposed PixelCraft, a multi-agent system that employs high-precision pixel-level tool agents and components such as a planner, critic, and image memory to perform multi-step, scalable visual reasoning on structured images (e.g., charts and geometric diagrams).

PixelVLA: Advancing Pixel-level Understanding in Vision-Language-Action Model

Wenqi Liang (University of Trento), Yang Cong (South China University of Technology)

🎯 What it does: Propose PixelVLA, a visual-language-action model with pixel-level understanding and multimodal prompting

PixNerd: Pixel Neural Field Diffusion

Shuai Wang (Nanjing University), Limin Wang (Nanjing University)

🎯 What it does: Directly construct the diffusion transformer PixNerd in the pixel space, achieving efficient single-stage generation through large patches and neural field decoders

PLAGUE: Plug-and-play framework for Lifelong Adaptive Generation of mUlti-turn jailbrEaks

Neeladri Bhuiya (University of Massachusetts Amherst), Diptanshu Purwar (A10 Networks, Inc.)

🎯 What it does: Propose the PLAGUE framework, which implements multi-round jailbreak through a three-stage process (Planner-Primer-Finisher), supporting lifecycle learning and modular pluggability;

Plan and Budget: Effective and Efficient Test-Time Scaling on Reasoning Large Language Models

Junhong Lin, Dawei Zhou

🎯 What it does: Propose the PLAN-AND-BUDGET framework, combining planning decomposition queries with adaptive token budget based on uncertainty, significantly improving inference efficiency during LLM reasoning.

Plan then Act: Bi-level CAD Command Sequence Generation

Qiangya Guo (South China University of Technology), Tianshui Chen (Zhuzhou CRRC Times Electric Co., Ltd)

🎯 What it does: Propose the PTA two-layer CAD command sequence generation method, first using LLM to plan high-level operation plans, then generating low-level executable command sequences with a demand-aware mechanism.

Plan-Answer-Refine-on-Graph: Structured Planning and Self-Refinement for Large Language Model Reasoning on Knowledge Graphs

Yuxin Shi (Zhejiang University), Jianling Sun (Zhejiang University)

🎯 What it does: Proposed the PARoG framework for large language model reasoning on knowledge graphs;

Plan-R1: Safe and Feasible Trajectory Planning as Language Modeling

Xiaolong Tang (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)

🎯 What it does: Propose the Plan-R1 two-stage framework: first pre-train a general trajectory predictor on expert driving data, then fine-tune with rule-based reinforcement learning to make trajectory planning explicitly adhere to safety, comfort, and rule principles.

PLANETALIGN: A Comprehensive Python Library for Benchmarking Network Alignment

Qi Yu (University of Illinois Urbana Champaign), Hanghang Tong (Texas Aamp University)

🎯 What it does: Proposed and implemented the PLANETALIGN library for unified experimentation and evaluation of network alignment methods

Planned Diffusion

Daniel Mingyi Israel (University of California, Los Angeles), Michael Carbin (MIT)

🎯 What it does: Designed and implemented a hybrid text generation framework combining autoregressive planning with parallel discrete diffusion (planned diffusion).

Planner Aware Path Learning in Diffusion Language Models Training

Fred Zhangzhi Peng (Duke University), Alexander Tong (AITHYRA)

🎯 What it does: Propose the Planner Aware Path Learning (PAPL) method, which aligns the training of discrete diffusion language models with planning inference using Planner-Aware ELBO (P-ELBO), achieving efficient training with a single-line modification;

Planning with an Embodied Learnable Memory

Priyam Parashar (Meta), Roozbeh Mottaghi (Meta)

🎯 What it does: Propose Embodied Perception Memory (EPM), which employs vision-language models to maintain dynamic scene memory with readable text in real-time, and combines LLM planning with DDAFT reinforcement learning to achieve long-sequence mobile manipulation tasks.

PlantRSR: A New Plant Dataset and Method for Reference-based Super-Resolution

Hongyang Zhou, Xu-Cheng Yin (University Of Science And Technology Beijing)

🎯 What it does: This paper proposes a reference image super-resolution method specifically for plant images and constructs a large-scale plant reference dataset called PlantRSR.

Play to Generalize: Learning to Reason Through Game Play

Yunfei Xie (Rice University), Chen Wei (NVIDIA)

🎯 What it does: Post-training a multi-modal large language model (MLLM) using reinforcement learning in simple games (Snake, Rotation) to enhance its performance in multi-modal reasoning tasks such as mathematics, geometry, and multi-disciplinary question answering.

PLoP: Precise LoRA Placement for Efficient Finetuning of Large Models

Soufiane Hayou (UC Berkeley), Bin Yu (UC Berkeley)

🎯 What it does: Proposes the PLoP method, which automatically selects the module type for inserting LoRA adapters by calculating the NFN (Normalized Feature Norm) score of pre-trained models on downstream tasks, thus achieving efficient parameterized fine-tuning.

Plug-and-Play Compositionality for Boosting Continual Learning with Foundation Models

Weiduo Liao (Zhejiang University), Ying Wei (Zhejiang University)

🎯 What it does: Introduce unsupervised Slot Attention and primitive selection in continual learning, embedding visual concepts as low-dimensional feature representations into the model;

Plug-and-Play Fidelity Optimization for Diffusion Transformer Acceleration via Cumulative Error Minimization

Tong Shao (Harbin Institute of Technology), Jingyong Su (Harbin Institute of Technology)

🎯 What it does: Propose a training-agnostic, plug-and-play cumulative error minimization plugin called CEM to optimize the caching strategy of Diffusion Transformer and improve generation quality.

Plug, Play, and Fortify: A Low-Cost Module for Robust Multimodal Image Understanding Models

Siqi Lu (National University of Defense Technology), Jianhang Yao (National University of Defense Technology)

🎯 What it does: Propose a frequency-domain based modality preference metric (FRM), and design the MWAM module using FRM to dynamically balance gradients/losses of the multimodal network during training, thereby enhancing robustness in scenarios with missing modalities.

PluriHarms: Benchmarking the Full Spectrum of Human Judgments on AI Harm

Jing-Jing Li (University of California Berkeley), Sydney Levine (New York University)

🎯 What it does: This paper proposes and publicly releases the PLURIHARMS benchmark, systematically collecting 150 AI safety prompts. Each prompt is scored for harm by 100 annotators on a 0–100 scale, with demographic, psychological characteristics of annotators, and explainable harm and value features of prompts recorded.

PM-KVQ: Progressive Mixed-precision KV Cache Quantization for Long-CoT LLMs

Tengxuan Liu, Yu Wang (Infinigence-Ai)

🎯 What it does: This paper proposes a post-training KV cache quantization method called Progressive Mixed-Precision KV Cache Quantization (PM-KVQ), specifically designed for large language models engaged in long-chain reasoning (long-CoT).

PMark: Towards Robust and Distortion-free Semantic-level Watermarking with Channel Constraints

Jiahao Huo, Mingxun Zhou (Hong Kong University of Science and Technology)

🎯 What it does: Proposed and implemented a semantic-level distortion-free watermarking scheme called PMARK based on proxy functions and multi-channel median sampling.

PMDformer: Patch-Mean Decoupling Information Transformer for Long-term Forecasting

Ao Hu (Southwestern University of Finance and Economics), Zenglin Xu (Shanghai Academy of AI for Science)

🎯 What it does: Propose the PMDformer model, addressing the challenge of shape similarity recognition caused by scale differences in long-term time series prediction through modules such as patch-mean decoupling, proximal variable attention, and trend restoration attention, thereby achieving more accurate multivariate forecasting.

POEMetric: The Last Stanza of Humanity

Bingru Li (University of Birmingham), Hazel Wilkinson (University of Birmingham)

🎯 What it does: Constructed and validated POEMetric, a comprehensive framework for evaluating poetry generation, which assessed 30 large language models (LLMs) and human poets' poems across multidimensional criteria ranging from adherence to poetic forms to creativity, emotional resonance, imagery, and literary techniques.

PoinnCARE: Hyperbolic Multi-Modal Learning for Enzyme Classification

Kun XIE, Biaobin Jiang (Tencent)

🎯 What it does: Proposes PoinnCARE, a hyperbolic (Poincaré ball) embedding framework based on bimodal data (sequence, structure, active site), utilizing graph diffusion and cross-modal alignment for enzyme EC number prediction.

Point Prompting: Counterfactual Tracking with Video Diffusion Models

Ayush Shrivastava (University of Michigan), Andrew Owens (University of Michigan)

🎯 What it does: Leverage pre-trained video diffusion models by adding color marker points on the first frame of the video and using SDEdit for reverse generation, achieving zero-shot tracking of motion trajectories in videos;

Point-Focused Attention Meets Context-Scan State Space: Robust Biological Visual Perception for Point Cloud Representation

Kanglin Qu (Nanjing University of Aeronautics and Astronautics), Yuanhao Sun (Beijing University of Posts and Telecommunications)

🎯 What it does: Proposes PointLearner, a point cloud representation learning network that mimics the foveal vision of the human retina, capable of simultaneously capturing local details and global context.

Point-MoE: Large-Scale Multi-Dataset Training with Mixture-of-Experts for 3D Semantic Segmentation

Xuweiyi Chen (University of Virginia), Zezhou Cheng (University of Virginia)

🎯 What it does: Propose Point-MoE, which employs a sparse MoE structure for large-scale joint training of 3D point cloud semantic segmentation without using dataset labels.

Point-UQ: An Uncertainty-Quantification Paradigm for Point Cloud Few-Shot Class Incremental Learning

Xiangqi Li (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Yongjun Xu (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences)

🎯 What it does: Propose a point cloud few-shot incremental learning framework Point-UQ based on uncertainty quantification, which uses dynamic decision-making instead of frequent fine-tuning to maintain base class knowledge while learning new classes.

Point-wise Anomaly Detection via Fold-bifurcation ODE

Sheo yon Jhin, Noseong Park (KAIST)

🎯 What it does: This paper proposes a point-level anomaly detection framework called FOLD based on folded bifurcation dynamics. It utilizes sensitivity and uncertainty signals generated by a predictive model as time-varying 'pressure' signals. By simulating pressure accumulation and critical point transitions through an ODE model, FOLD achieves anomaly detection without requiring labels or additional training.

Point2RBox-v3: Self-Bootstrapping from Point Annotations via Integrated Pseudo-Label Refinement and Utilization

Teng Zhang (Shanghai Jiao Tong University), Xue Yang (East China Normal University)

🎯 What it does: Proposes a self-supervised rotation object detection framework named Point2RBox-v3 based on point annotations, integrating dynamic pseudo-labels and mask augmentation.