π― What it does: This paper proposes the Unbalanced Graph Collection (UA) paradigm, applies it to self-supervised state representation learning, improves ST-DIM to DIM-UA, and conducts experiments on AtariARI and CIFAR10.
Stochastic Gradient Descent for Gaussian Processes Done Right
Jihao Andreas Lin (University of Cambridge), David Janz (University of Alberta)
CodeOptimizationGraph Neural NetworkTabular
π― What it does: A stochastic dual descent (SDD) algorithm is proposed for Gaussian process regression and sampling to solve large-scale linear systems, directly replacing traditional conjugate gradient or variational methods.
π― What it does: For the task of camouflage object detection (COD), this paper proposes two key technologies: first, the Camouflageator adversarial training framework from the perspective of the 'predator', which uses an auxiliary generator to synthesize more difficult-to-detect camouflage images to enhance the generalization ability of the detector; second, the ICEG detector from the perspective of the 'prey', which includes an internal consistency module (CFC) for segmentation and an edge-guided separation calibration module (ESC) to eliminate fuzzy boundaries.
π― What it does: This paper proposes a method for structural inference using Variational Dynamic Encoder (VDE) and Partial Correlation Coefficient (PCOR);
π― What it does: A new contrastive learning framework called CARE is proposed, which utilizes linear isometric transformations such as rotation to map input augmentations to orthogonal transformations in the embedding space, thereby learning representations that can distinguish samples while maintaining geometric interpretability on unlabeled data.
Stylized Offline Reinforcement Learning: Extracting Diverse High-Quality Behaviors from Heterogeneous Datasets
Yihuan Mao (Institute for Interdisciplinary Information Sciences), Chongjie Zhang (Washington University in St. Louis)
CodeReinforcement LearningTabular
π― What it does: This paper proposes a two-step framework called Stylized Offline RL (SORL) for extracting both high-quality and diverse policies from heterogeneous offline datasets.
Manish Prajapat (ETH Zurich), Andreas Krause (ETH Zurich)
CodeReinforcement Learning
π― What it does: A submodular reinforcement learning framework called SUBRL is proposed, along with a greedy policy gradient algorithm SUBPO, which can handle non-additive, history-dependent rewards.
π― What it does: A deep probabilistic circuit (NPC2) is proposed to achieve subtractive mixing through a square mixture model, enabling non-negative and interpretable distribution modeling.
Successor Heads: Recurring, Interpretable Attention Heads In The Wild
Rhys Gould (University of Cambridge), Arthur Conmy (Independent)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: This paper discovers and explains the 'successor head' in large language modelsβa type of attention head that can increment words in a sequence (such as numbers, months, days of the week), and reveals its internal implementation through mechanistic interpretability analysis.
π― What it does: This paper studies how noise recurrent neural networks can automatically generate offline reactivation in the absence of external stimuli after task optimization, and provides sufficient conditions for this phenomenon.
π― What it does: A new interpretable deep generative model called Sum-Product-Set Networks (SPSNs) is proposed for representing and reasoning about the probability distribution of tree-structured graphs.
Supervised Knowledge Makes Large Language Models Better In-context Learners
Linyi Yang (Westlake University), Yue Zhang (Westlake University)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: In the inference phase of large language models (LLMs), the output of a small supervised learning model (SLM) is incorporated to enhance out-of-distribution (OOD) generalization and reduce hallucination.
SuRe: Summarizing Retrievals using Answer Candidates for Open-domain QA of LLMs
Jaehyung Kim (Carnegie Mellon University), Jinwoo Shin (KAIST AI)
CodeRetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Using zero-shot prompts to first generate answer candidates, then generating conditional summaries for each candidate based on retrieved paragraphs, and verifying and selecting the most suitable answer through the effectiveness of the summaries and their mutual ranking, in order to improve the accuracy of open-domain question answering.
SWAP-NAS: Sample-Wise Activation Patterns for Ultra-fast NAS
Yameng Peng (RMIT University), Xiaojun Chang (University of Technology Sydney)
CodeNeural Architecture SearchImage
π― What it does: This paper proposes a training-free network performance evaluation metric called SWAP-Score, and based on this metric, implements an ultra-fast NAS method named SWAP-NAS.
SWAP: Sparse Entropic Wasserstein Regression for Robust Network Pruning
Lei You (Technical University of Denmark), Hei Victor Cheng (Aarhus University)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: This paper proposes a network pruning method called SWAP based on entropy-regularized Wasserstein regression (EWR), which reduces gradient noise and retains covariance information using OT distance.
π― What it does: A meta-learning framework based on symbolic equation learning (SYMBOL) is proposed, which constructs a black-box optimizer by automatically generating closed-form update rules and uses reinforcement learning for meta-learning.
Symmetric Basis Convolutions for Learning Lagrangian Fluid Mechanics
Rene Winchenbach (Technical University Munich), Nils Thuerey (Technical University Munich)
CodeGraph Neural NetworkGraphPhysics Related
π― What it does: A symmetric Fourier basis continuous convolution (SFBC) framework is proposed and implemented for learning physical simulations in Lagrangian fluid dynamics.
π― What it does: This paper proposes Symphony, an autoregressive 3D molecular generation model that utilizes higher-order E(3)-equivariant features and spherical harmonic projections.
T-Rep: Representation Learning for Time Series using Time-Embeddings
Archibald Felix Fraikin, Stephanie Allassonniere (Universite Paris Cite)
CodeAnomaly DetectionRepresentation LearningConvolutional Neural NetworkTime Series
π― What it does: A self-supervised time series representation learning framework T-Rep has been developed to capture fine-grained temporal features by learning time embeddings and pre-training tasks.
TAB: Temporal Accumulated Batch Normalization in Spiking Neural Networks
Haiyan Jiang (Mohamed bin Zayed University of Artificial Intelligence), Huan Xiong (Harbin Institute of Technology)
CodeSpiking Neural NetworkImage
π― What it does: A batch normalization method named TAB (Temporal Accumulated Batch Normalization) is proposed for directly training Spiking Neural Networks (SNNs), addressing the temporal covariance shift (TCS) problem by utilizing temporal accumulated statistics.
π― What it does: TACTiS-2 is proposed, a Transformer-based attention copula model that can flexibly perform multivariate time series prediction, interpolation, and their combination tasks;
Tag2Text: Guiding Vision-Language Model via Image Tagging
Xinyu Huang (Fudan University), Lei Zhang (International Digital Economy Academy)
CodeGenerationRetrievalTransformerVision Language ModelImageTextMultimodality
π― What it does: This paper proposes a visual-language pre-training framework called Tag2Text, which embeds image label learning to utilize label information in image-text pairs to guide multimodal feature learning.
Nayoung Lee (University of Wisconsin Madison), Dimitris Papailiopoulos (University of Wisconsin Madison)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: The study explores how to efficiently learn basic arithmetic operations (addition, multiplication, square root, sine, etc.) starting from random initialization on small Transformer models (such as NanoGPT, GPT-2) using only the next word prediction as the target.
Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMs
Qingru Zhang (Georgia Institute of Technology), Tuo Zhao (Georgia Institute of Technology)
CodeGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A post-attention steering method called PASTA is proposed and implemented, which utilizes user-emphasized markers such as bold/italic in the text to reweight the multi-head attention of large language models (LLMs) to guide the model's focus on user-specified information during inference.
TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting
Defu Cao (University of Southern California), Yan Liu (University of Southern California)
CodeTransformerLarge Language ModelPrompt EngineeringMultimodalityTime SeriesFinance Related
π― What it does: A time series prediction framework TEMPO based on a pre-trained generative Transformer is proposed, utilizing the trend, seasonality, and residual decomposition of time series, and fine-tuning the model through soft prompts to achieve zero-shot transfer and multimodal fusion prediction.
Test-Time Adaptation with CLIP Reward for Zero-Shot Generalization in Vision-Language Models
Shuai Zhao (University of Technology Sydney), Yi Yang (Zhejiang University)
CodeClassificationRetrievalDomain AdaptationTransformerReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes a testing-time adaptive framework RLCF based on CLIP rewards, which utilizes reinforcement learning to dynamically update the parameters of the visual-language model (VLM) on a single test sample to enhance the generalization ability of zero-shot tasks.
Test-Time Training on Nearest Neighbors for Large Language Models
Moritz Hardt (Max Planck Institute for Intelligent Systems), Yu Sun (Stanford University)
CodeRetrievalOptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Proposes a test-time training method that utilizes nearest neighbor retrieval results for a single gradient update of the language model (Test-Time Training on Nearest Neighbors, TTT-NN).
TESTAM: A Time-Enhanced Spatio-Temporal Attention Model with Mixture of Experts
Hyunwook Lee (Ulsan National Institute of Science and Technology), Sungahn Ko (Ulsan National Institute of Science and Technology)
CodeGraph Neural NetworkTransformerMixture of ExpertsTime Series
π― What it does: A multi-expert spatiotemporal attention model (TESTAM) is proposed, achieving adaptive spatial and temporal modeling in different traffic scenarios.
The All-Seeing Project: Towards Panoptic Visual Recognition and Understanding of the Open World
Weiyun Wang (Shanghai AI Laboratory), Yu Qiao (Tsinghua University)
CodeRecognitionObject DetectionTransformerLarge Language ModelContrastive LearningImageText
π― What it does: A large-scale regional-level annotated dataset AS-1B with 120 million annotations has been constructed, and a unified All-Seeing Model (ASM) has been proposed to achieve panoramic visual recognition and understanding for open-world scenarios.
The Curse of Diversity in Ensemble-Based Exploration
Zhixuan Lin (Mila - Quebec AI Institute, Universite de Montreal), Aaron Courville (Mila - Quebec AI Institute, Universite de Montreal)
CodeReinforcement LearningSequential
π― What it does: This study investigates the 'diversity curse' that arises when using data-sharing diversified ensemble exploration strategies in deep reinforcement learning, where the performance of individual members is significantly lower than that of a single agent; it proposes the use of Cross-ensemble Representation Learning (CERL) auxiliary tasks to alleviate this issue and enhance the performance of the aggregated strategy.
The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models
Yan Liu (Chinese University of Hong Kong), Tsung-Yi Ho (National University of Singapore)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes the Integrated Gap Gradients (IG2) method to accurately locate the social bias neurons in pre-trained language models, and based on this, designs a training-independent Bias Neuron Suppression (BNS) to achieve debiasing.
π― What it does: This study investigates how the intrinsic properties of datasets (intrinsic dimension and label sharpness) affect the generalization ability and adversarial robustness of neural networks across different image domains (natural images and medical images).
The Effective Horizon Explains Deep RL Performance in Stochastic Environments
Cassidy Laidlaw (University of California), Anca Dragan (University of California)
CodeReinforcement LearningTabular
π― What it does: The study explains the efficiency of deep RL in stochastic environments through random exploration and function approximation, proposes the SQIRL algorithm, and proves that its sample complexity is only exponential in the effective horizon under low effective horizons.
π― What it does: A novel adversarial training framework called FOMO is proposed, which alleviates robust overfitting and enhances the model's robust generalization ability by periodically forgetting part of the weights (random reinitialization) and relearning.
The False Promise of Imitating Proprietary Language Models
Arnav Gudibande (University of California Berkeley), Dawn Song (University of California Berkeley)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: In this paper, the authors fine-tune various open-source language models (such as GPT-2, LLaMA 7B/13B) by collecting outputs from ChatGPT as training data, aiming to mimic the behavior of ChatGPT and evaluate whether 'imitation learning' can achieve the functionality of closed-source models at a lower cost.
Tycho F. A. van der Ouderaa, Tijmen Blankevoort (University of Amsterdam)
CodeCompressionTransformerLarge Language ModelText
π― What it does: The LLM Surgeon framework is proposed, which can perform unstructured, semi-structured, and structured pruning on large language models, supporting multi-step pruning and low-rank one-shot gradient correction.
The Reasonableness Behind Unreasonable Translation Capability of Large Language Model
Tingchen Fu (Renmin University of China), Rui Yan (Renmin University of China)
CodeTransformerLarge Language ModelText
π― What it does: This paper explores the reasons why multilingual large language models can acquire translation capabilities without parallel corpora, systematically analyzing the impact of different granularities of non-parallel bilingual data, such as sentence-level and word-level bilingual contamination and code-switching, on translation performance.
The Reversal Curse: LLMs trained on βA is Bβ fail to learn βB is Aβ
Lukas Berglund (Vanderbilt University), Owain Evans (University of Oxford)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Through fine-tuning and empirical evaluation of LLMs, it was found that they cannot automatically infer 'B is A' after learning 'A is B', leading to the phenomenon termed 'reverse curse'.
π― What it does: This paper proposes an algorithm called Wasserstein Belief Updater (WBU), which can construct reliable Bayesian state estimates in partially observable environments by learning latent models and performing Bayesian updates, thus achieving effective compression of history without using RNNs.
π― What it does: This paper presents a theoretical analysis of robust overfitting in wide DNNs during adversarial training using the NTK method. It proves that wide networks can be approximated linearly, provides a closed-form dynamic for adversarial training, and reveals the phenomenon of 'AT degradation,' where long-term adversarial training leads the network to degrade to a state without adversarial training. Based on this, the first infinite-width adversarial training algorithm, Adv-NTK, is designed.
Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph
Jiashuo Sun (IDEA Research), Jian Guo
CodeGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraph
π― What it does: A new LLMβKG coupling paradigm is proposedβThink-On-Graph (ToG), which allows large language models to perform adaptive beam search on knowledge graphs and make reasoning decisions;
THOUGHT PROPAGATION: AN ANALOGICAL APPROACH TO COMPLEX REASONING WITH LARGE LANGUAGE MODELS
Junchi Yu (Institute of Automation Chinese Academy of Sciences), Zhitao Ying
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: A Thought Propagation framework is proposed, utilizing solutions to similar problems to improve multi-step reasoning in large language models.
π― What it does: This study investigates the rate information and temporal information of SNNs, proposing a hybrid attack method called HART that efficiently attacks SNNs in various attack scenarios.
π― What it does: A time-continuous CLIP training benchmark (TIC-DataComp, TIC-YFCC, TIC-RedCaps) is proposed, and dynamic retrieval and classification evaluation tasks are designed on it.
Adam Lechowicz (University of Massachusetts Amherst), Mohammad Hajiesmaili (University of Massachusetts Amherst)
CodeOptimizationTabular
π― What it does: This paper studies the time fairness of the online knapsack problem, introduces Ξ±-conditional time-independent fairness, designs threshold-based deterministic and learning-enhanced algorithms, and conducts experimental evaluations.
Time Travel in LLMs: Tracing Data Contamination in Large Language Models
Shahriar Golchin (University of Arizona), Mihai Surdeanu (University of Arizona)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Using guided instruction to have large language models complete partial instances and assess data contamination in the training or test set through matching evaluation.
π― What it does: A stateful policy gradient estimation method called S2PG is proposed, which does not use BPTT. It utilizes a stochastic internal state transition kernel to decompose the stateful policy into a stochastic kernel and a non-stateful policy, optimizing them jointly to address the issues of gradient vanishing, explosion, and sequence computation bottlenecks associated with traditional BPTT.
Time-LLM: Time Series Forecasting by Reprogramming Large Language Models
Ming Jin (Monash University), Qingsong Wen (Alibaba Group)
CodeTransformerLarge Language ModelPrompt EngineeringTime Series
π― What it does: The TIME-LLM framework is proposed, which reprograms time series data into text prototypes and adds Prompt-as-Prefix in front of the LLM, enabling the frozen LLM to perform general time series prediction tasks.
TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting
Shiyu Wang (Ant Group), JUN ZHOU
CodeTime Series
π― What it does: This paper proposes a multi-scale hybrid framework called TimeMixer based on a fully MLP architecture to address the complex non-stationarity issues in time series forecasting.
π― What it does: This study investigates the relationship between generalization and memory in deep learning, particularly on algorithm datasets with corrupted labels, using interpretable models to analyze the differences between generalization representations and memory representations.
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes a tool-augmented reward model framework called Themis, which enables the reward model to dynamically call external tools (such as calculators, search engines, etc.) during the reasoning process to obtain information and evaluate the quality of outputs.
TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning
Dongming Wu (Beijing Institute of Technology), Jianbing Shen
CodeObject DetectionAutonomous DrivingPoint Cloud
π― What it does: A complete process is proposed that combines query-based 3D lane detection with 2D traffic element detection, using a simple MLP to achieve topological reasoning between lanes and traffic elements.
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Designed and trained a series of Tool Integrated Reasoning Agents (TORA) to solve complex mathematical problems through interaction with external computational/symbolic tools using natural language reasoning.
TorchRL: A data-driven decision-making library for PyTorch
Albert Bou (Acellera), Vincent Moens (Meta)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: This paper presents TorchRL, a general decision-making and control library based on PyTorch, featuring a composable and modular component design.
π― What it does: This study investigates how to protect images from imitation attacks based on diffusion models and proposes a method for generating adversarial perturbations against potential threats.
Towards 3D Molecule-Text Interpretation in Language Models
Sihang Li (University of Science and Technology of China), Qi Tian (Huawei Cloud)
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
π― What it does: This paper proposes 3D-MoLM, which integrates a 3D molecular encoder with a language model to achieve the parsing and generation of 3D molecules into text.
Towards Aligned Layout Generation via Diffusion Model with Aesthetic Constraints
Jian Chen (University at Buffalo), Changyou Chen (University at Buffalo)
CodeGenerationDiffusion modelImage
π― What it does: A unified continuous space diffusion model LACE is proposed for controllable layout generation, capable of handling various tasks such as unconditional generation, category/size conditional generation, partial layout completion, and refinement.
π― What it does: The researchers propose to directly estimate domain adversarial counterfactuals using domain labels within the framework of Invertible Latent Causal Models (ILD), avoiding the need for complete causal structure recovery.
π― What it does: An algorithm is proposed to analytically recover soft labels (label smoothing, mixup) and corresponding fully connected layer features from a single sample gradient inversion.
Towards Energy Efficient Spiking Neural Networks: An Unstructured Pruning Framework
Xinyu Shi (Peking University), Zhaofei Yu (Peking University)
CodeSpiking Neural NetworkImage
π― What it does: A new framework that combines unstructured weight pruning and unstructured neuron pruning is proposed, aiming to maximize the sparsity of Spiking Neural Networks (SNN) on neuromorphic hardware to enhance energy efficiency.
Towards Faithful Explanations: Boosting Rationalization with Shortcuts Discovery
Linan Yue (University of Science and Technology of China), Yanqing An (University of Science and Technology of China)
CodeClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningText
π― What it does: A semi-supervised interpretability method SSR based on the discovery and utilization of 'shortcut keys' is proposed, significantly improving prediction accuracy and interpretability in text classification tasks.
Towards Faithful XAI Evaluation via Generalization-Limited Backdoor Watermark
Mengxi Ya (Tsinghua University), Shu-Tao Xia (Tsinghua University)
CodeOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: This paper proposes a 'Generalized Limited Backdoor Watermark' (GLBW) to improve the automated evaluation of saliency-based representation visualization (SRV).
π― What it does: Multi-task fine-tuning of the base model is conducted, theoretically analyzing its advantages in downstream tasks with a small number of labels, and proposing a task selection algorithm based on diversity and consistency.
Towards Foundation Models for Knowledge Graph Reasoning
Mikhail Galkin (Intel AI Lab), Zhaocheng Zhu (Mila University of Montreal)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes ULTRA, a foundational model that enables zero-shot reasoning and fine-tuning across different knowledge graphs by constructing a relation graph and learning conditional relation representations to achieve cross-graph knowledge inference.
Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets
Dominique Beaini (Mila - Quebec AI Institute), Dominic Masters (Graphcore)
CodeDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
π― What it does: A large-scale multi-task molecular dataset covering nearly 100 billion molecules and over 3000 tasks has been developed, along with the introduction of the Graphium library for efficient training.
Towards Generative Abstract Reasoning: Completing Ravenβs Progressive Matrix via Rule Abstraction and Selection
Fan Shi (Fudan University), Xiangyang Xue (Fudan University)
CodeGenerationData SynthesisAuto EncoderImage
π― What it does: A generative Ravenβs Progressive Matrix solver called RAISE is proposed, which is based on a conditional deep latent variable model. It can abstract attribute concepts from contextual images and select corresponding atomic rules to generate missing images at any position.
π― What it does: This paper proposes an unsupervised domain translation method based on diversified distribution matching, addressing the identifiability issues present in traditional methods.
CodeRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTime SeriesSequentialBenchmarkPhysics Related
π― What it does: This paper proposes an unsupervised pre-training framework called LLM4QPE, which first pre-trains a Transformer+LSTM model using a large amount of unlabeled quantum measurement records, and then fine-tunes it on a limited amount of labeled data to achieve task-independent estimation of quantum system properties.
π― What it does: A Difficulty-Aligned Trajectory Matching (DATM) method is proposed to achieve lossless dataset distillation under both low IPC and high IPC conditions, enhancing distillation stability through soft label learning and sequential generation.
Alexander Theus (ETH Zurich), Sidak Pal Singh (ETH Zurich)
CodeOptimizationConvolutional Neural NetworkImage
π― What it does: A new structured pruning method called Intra-Fusion is proposed, which utilizes Optimal Transport to fuse the pruned neurons instead of simply discarding them, thereby maintaining model performance without data or fine-tuning.
π― What it does: A feature reshaping method is proposed that can be trained without OOD samples to enhance the generalization ability of OOD detection.
π― What it does: This paper constructs offline pre-trained representations based on video data and theoretically analyzes their feasibility in reinforcement learning (RL). It proves that under a Block MDP without external noise, forward modeling and temporal contrastive learning can learn latent state representations that are useful for efficient downstream RL. It also provides a lower bound showing that when external noise is present, the sample complexity of video pre-training deteriorates exponentially compared to trajectory pre-training. Experiments are then conducted in visual environments such as GridWorld and ViZDoom to validate these findings.
π― What it does: A new backdoor trigger inversion (BTI) method called BTI-DBF is proposed to address backdoor attacks in third-party models. This method recovers triggers by decoupling normal features rather than directly decoupling backdoor features, and based on this, two types of backdoor defenses (removal and preprocessing) are designed.
π― What it does: A Cloud-Edge Elastic Model Adaptation (CEMA) framework is proposed, where edge devices only perform forward inference, while the cloud is responsible for self-supervised entropy minimization, replay-based knowledge distillation, and parameter updates, achieving dynamic and low-cost online model adaptation.
Towards Robust Multi-Modal Reasoning via Model Selection
Xiangyan Liu (National University of Singapore), Tao Lin (Westlake University)
CodeOptimizationRepresentation LearningGraph Neural NetworkVision Language ModelMultimodalityBenchmark
π― What it does: A framework named MΒ³ is proposed for dynamic model selection during multi-modal multi-step reasoning processes, thereby enhancing overall reasoning robustness.
Towards Robust Offline Reinforcement Learning under Diverse Data Corruption
Rui Yang (Hong Kong University of Science and Technology), Tong Zhang (University of Texas at Austin)
CodeReinforcement LearningTabular
π― What it does: This paper evaluates the robustness of offline reinforcement learning algorithms under various types of data contamination (state, action, reward, dynamics) and proposes Robust IQL (RIQL) to improve performance in these contaminated scenarios.
π― What it does: A method that integrates global and local features, called SelaVPR, is proposed to achieve seamless transfer of the pre-trained visual foundation model (DINOv2) for the visual place recognition (VPR) task.
Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion
Alexandru Meterez (ETH Zurich), Hadi Daneshmand (MIT)
CodeImage
π― What it does: This paper proposes a multilayer perceptron (MLP) using orthogonal random weights under batch normalization (BN) and demonstrates that it can avoid gradient explosion while maintaining good signal propagation.
Krzysztof Kacprzyk (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeExplainability and InterpretabilityRecurrent Neural NetworkTime SeriesOrdinary Differential Equation
π― What it does: A time series transparent prediction framework called TIMEVIEW based on static features is proposed, utilizing two-layer (trend and attribute) transparency for interpretation;
Towards Understanding Factual Knowledge of Large Language Models
Xuming Hu (Hong Kong University of Science and Technology), Zhijiang Guo (University of Cambridge)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Proposed the Pinocchio benchmark, constructing 20,713 multilingual, multi-domain, and multi-temporal factual knowledge and reasoning questions to systematically evaluate the capabilities of LLMs in factual memory and reasoning;
π― What it does: A fully sparse Bayesian neural network training framework SSVI is proposed, which maintains high sparsity during both training and inference.
π― What it does: This paper proposes a time-dependent importance reweighting (TIW) diffusion model that corrects dataset bias by estimating the time-varying density ratio and achieves unbiased score matching.
π― What it does: This paper proposes Transformer-VQ, a decoder Transformer that utilizes vector quantization (VQ) to achieve linear time self-attention.
π― What it does: This paper studies and proves that using supervised pre-trained Transformers can perform reinforcement learning in new environments (ICRL) and can approximate classic algorithms such as LinUCB, Thompson sampling, and UCB-VI.
π― What it does: This paper proposes a framework that unifies optimal transport and variational inference, and based on this, designs the Controlled Monte Carlo Diffusion (CMCD) sampler, which can simultaneously learn the forward and backward diffusion processes.
π― What it does: Proposes Tree Cross Attention (TCA) and ReTreever architecture to achieve logarithmic token retrieval and efficient inference through cross-attention;
Tree Search-Based Policy Optimization under Stochastic Execution Delay
David Valensi (Technion), Gal Dalal (Nvidia Research)
CodeOptimizationReinforcement LearningSequential
π― What it does: A framework for MDPs with stochastic execution delays (SED-MDP) has been designed and implemented, and within this framework, the DEZ algorithm is proposed, which utilizes delay queues and effective decision time to achieve state-agnostic policy learning and inference.
True Knowledge Comes from Practice: Aligning Large Language Models with Embodied Environments via Reinforcement Learning
Weihao Tan (Nanyang Technological University), Bo An (Nanyang Technological University)
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This study proposes the TWOSOME framework, which aligns large language models with embodied environments online through reinforcement learning to efficiently complete decision-making tasks.
Turning large language models into cognitive models
Marcel Binz (Max Planck Institute for Biological Cybernetics), Eric Schulz (Max Planck Institute for Biological Cybernetics)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By fine-tuning the final layer of the large language model LLaMA and generating text embeddings using psychological experimental data, an interpretable cognitive model called CENTaUR was constructed to simulate human decision-making behavior.
π― What it does: This paper proposes Texture UV Radiance Fields (TUVF), which decouples texture from 3D shapes by generating textures in a learnable UV spherical space, enabling texture editing and cross-shape transfer.
π― What it does: A test-time normalization layer named UnMix-TNS is proposed, which can achieve adaptive normalization in label time-dependent non-i.i.d. test streams, enhancing model robustness.