Conference on Neural Information Processing Systems Β· 1874 papers
Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data
Fan Zhang (Georgia Institute of Technology), Hongyu Zhao (Yale University)
CodeDomain AdaptationSupervised Fine-TuningBiomedical Data
π― What it does: A semi-supervised cross-omics single-cell data transfer framework called DANCE is proposed to transfer cell type labels between scRNA-seq and scATAC-seq data, addressing the issue of label scarcity at both the source and target ends.
π― What it does: A semidefinite programming (SDP) relaxation is proposed to solve the Gromov-Wasserstein (GW) distance, which can obtain globally optimal transport plans in most instances and provide a proof of global optimality.
Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation
Guillaume Huguet (Dreamfold), Joey Bose
CodeGenerationData SynthesisProtein Structure PredictionTransformerLarge Language ModelReinforcement LearningFlow-based ModelBiomedical Data
π― What it does: A flow-matching based SE(3)-invariant model FOLDFLOW-2 was developed for conditional protein backbone generation, integrating large-scale protein language models with structural information.
Sequential Decision Making with Expert Demonstrations under Unobserved Heterogeneity
Vahid Balazadeh (University of Toronto), Vasilis Syrgkanis (Stanford University)
CodeReinforcement LearningSequential
π― What it does: The study utilizes a framework for online sequential decision-making assisted by expert demonstrations under unobserved contextual heterogeneity and proposes the ExPerior algorithm.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: We propose SEQUOIA, a scalable and robust explicit sampling decoding framework that can significantly accelerate inference for large language models.
Set-based Neural Network Encoding Without Weight Tying
Bruno Andreis (KAIST), Sung Ju Hwang (KAIST)
CodeTransformerImage
π― What it does: This paper proposes a set-based encoding method for neural network weights called SNE, which can predict network performance and attributes using only model parameters and can transfer across architectures and datasets.
SfPUEL: Shape from Polarization under Unknown Environment Light
Youwei Lyu (Beijing University of Posts and Telecommunications), Boxin Shi (Peking University)
CodeSegmentationDepth EstimationTransformerImage
π― What it does: This paper proposes the SfPUEL framework, which simultaneously estimates surface normals and segments metallic/dielectric materials using a single polarized image under unknown ambient light.
SGLang: Efficient Execution of Structured Language Model Programs
Lianmin Zheng (University of California Berkeley), Ying Sheng (Stanford University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelImageVideoTextRetrieval-Augmented Generation
π― What it does: This paper presents SGLang, a domain-specific language and runtime for efficient programming and execution of structured language model programs (LM Programs).
Shadowcast: Stealthy Data Poisoning Attacks Against Vision-Language Models
Yuancheng Xu (University of Maryland), Furong Huang (University of Illinois Urbana-Champaign)
CodeAdversarial AttackData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: This paper proposes Shadowcast, an invisible data poisoning attack targeting visual language models, which induces misleading responses from the model using visually consistent text-image pairs.
Shaping the distribution of neural responses with interneurons in a recurrent circuit model
David Lipshutz (Flatiron Institute), Eero P Simoncelli
CodeImage
π― What it does: A feasible neural circuit model based on optimal transport theory is proposed, which achieves nonlinear mapping of input signals to target distributions (such as Gaussian distributions) through the collaborative adjustment of synapses, activation functions, and gain by plastic interneurons, thereby optimizing the neural response distribution.
Sharing Key Semantics in Transformer Makes Efficient Image Restoration
Bin Ren (University of Pisa), Nicu Sebe (University of Trento)
CodeRestorationSuper ResolutionTransformerImage
π― What it does: This paper proposes SemanIR, an image restoration framework that shares a key semantic dictionary within a Transformer, achieving efficient self-attention computation by focusing only on the most semantically related patches.
π― What it does: This study investigates the trade-off relationship between sharpness and diversity in deep ensembles and proposes the SharpBalance method, which utilizes an adaptively selected subset to balance both, thereby improving ID and OOD performance.
π― What it does: A self-supervised hierarchical makeup transfer method (SHMT) based on latent diffusion models is proposed, achieving makeup style transfer by splitting and reconstructing content and makeup information from facial images.
π― What it does: An adaptive aggregation strategy based on feature similarity and structural neighbors (SNAPS) is proposed, which utilizes the inconsistency scores of same-label nodes to improve the efficiency of the prediction set in the segmentation conformal prediction of graph neural networks.
CodeTransformerLarge Language ModelDiffusion modelTextBiomedical Data
π― What it does: This paper proposes and implements a Masked Discrete Diffusion Language Model (MDLM), which significantly improves the log-likelihood of diffusion models in language modeling tasks through a simplified variational objective and improved sampling methods.
π― What it does: A simplified and general mask diffusion model MD4 is proposed, and based on this, a state-dependent GenMD4 is introduced; a concise integral form of the continuous-time ELBO is provided, significantly simplifying the training and sampling process.
π― What it does: This paper studies how to simplify constraint inference through inverse reinforcement learning, reducing the traditional three-layer optimization structure of inverse constraint reinforcement learning to two layers, and achieving safe constraint learning based on this.
π― What it does: A world model called Parsimonious Latent Space Model (PLSM) is proposed, which makes the impact of actions on latent states more predictable through information bottleneck;
π― What it does: A flow matching-based framework for non-simulated training is proposed to learn deterministic mappings on continuous depth models (NODE).
SimVG: A Simple Framework for Visual Grounding with Decoupled Multi-modal Fusion
Ming Dai (Southeast University), Wankou Yang (Southeast University)
CodeRecognitionObject DetectionKnowledge DistillationTransformerVision Language ModelImageMultimodality
π― What it does: A concise and efficient visual grounding framework named SimVG is proposed, which decouples multimodal fusion from downstream tasks and directly uses a pretrained multimodal model for feature interaction.
π― What it does: A dual-stream interactive Transformer for single image reflection separation is proposed, utilizing global and local prior interactions and introducing a dual attention module to achieve inter-layer and intra-layer feature collaboration.
SIRIUS : Contexual Sparisty with Correction for Efficient LLMs
Yang Zhou (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper studies the use of Contextual Sparsity technology in large language model inference to reduce computational load and proposes the SIRIUS mechanism to correct erroneous tokens generated during the sparse model generation process, thereby restoring inference quality.
π― What it does: A low-memory Fisher information matrix approximation method is proposed for pre-trained neural networks, and it is used to calculate uncertainty scores (SLU).
π― What it does: A two-stage SkMM method is proposed for data selection in fine-tuning tasks, which explores the parameter space through gradient sketching and then performs moment matching to balance variance and bias.
π― What it does: This paper studies a context encoder based on skill-aware mutual information, aimed at enhancing the zero-shot generalization ability of Meta-RL across different tasks.
π― What it does: A new local update algorithm SLo wcal-SGD is proposed for heterogeneous distributed stochastic convex optimization, significantly improving communication efficiency and convergence speed in multi-machine training.
SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM
Ming Nie (Fudan University), Li Zhang (Fudan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoBenchmark
π― What it does: The SlowFocus mechanism is proposed, which enhances the understanding and reasoning of fine-grained temporal information in video LLMs through high-frequency sampling, temporal encoding, and multi-frequency mixed attention during query-related periods.
SLTrain: a sparse plus low rank approach for parameter and memory efficient pretraining
Andi Han (RIKEN Artificial Intelligence Project), Bamdev Mishra (Microsoft)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: The SLTrain method is proposed, which decomposes the weight matrix into low-rank and fixed random sparse parts during the pre-training phase of large language models, achieving high efficiency in both parameters and memory.
π― What it does: A new smoothing operator Sm is proposed to enhance the localization capability of multi-instance learning (MIL) in medical image classification, particularly in instance-level predictions.
SmallToLarge (S2L): Scalable Data Selection for Fine-tuning Large Language Models by Summarizing Training Trajectories of Small Models
Yu Yang (University of California), Baharan Mirzasoleiman (University of California)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
π― What it does: Proposed the SMALLTOLARGE (S2L) method: train a small model to obtain the loss trajectory for each sample, cluster the trajectories, and then uniformly sample from each cluster to obtain a subset for supervised fine-tuning of a large model.
SMART: Scalable Multi-agent Real-time Motion Generation via Next-token Prediction
Wei Wu (Tsinghua University), Yuheng KAN
CodeGenerationAutonomous DrivingTransformerLarge Language ModelSequential
π― What it does: A self-regressive motion generation framework called SMART based on discrete sequences is proposed, which uses a GPT-style decoder-only transformer to directly predict the next action or road vector label.
CodeTransformerSupervised Fine-TuningTabularBiomedical DataElectronic Health Records
π― What it does: The SMART model is proposed, utilizing self-supervised missingness-aware pre-training to enhance predictive performance on EHR data.
Riccardo Cadei (Institute of Science and Technology Austria), Francesco Locatello (Institute of Science and Technology Austria)
CodeTransformerVideoBenchmark
π― What it does: This study investigates how training design and prediction processes when using pre-trained deep learning models for causal downstream tasks may lead to biased causal effect estimates, validated based on a newly constructed high-dimensional observational dataset called ISTAnt.
π― What it does: A training-free, unconditional Smoothed Energy Guidance (SEG) method is proposed, which utilizes Gaussian blur on self-attention weights to reduce energy curvature, thereby enhancing the image generation quality of diffusion models.
SnapKV: LLM Knows What You are Looking for Before Generation
Yuhong Li (University of Illinois Urbana-Champaign), Deming Chen (University of Illinois Urbana-Champaign)
CodeGenerationCompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes SnapKV, a KV cache compression method that does not require fine-tuning, which can identify and retain the most important attention features in advance during the generation phase, significantly reducing time and memory consumption during long text inference.
SocialGPT: Prompting LLMs for Social Relation Reasoning via Greedy Segment Optimization
Wanhua Li (Harvard University), Hanspeter Pfister (Harvard University)
CodeOptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: Proposes the SocialGPT framework, which utilizes visual foundation models to extract image information, generates symbolic social stories, and then uses large language models for social relationship reasoning to provide interpretable answers.
Soft ascent-descent as a stable and flexible alternative to flooding
Matthew J. Holland (Osaka University), Kosuke Nakatani (Osaka University)
CodeClassificationOptimizationImage
π― What it does: A soft adaptive gradient method called SoftAD is designed and evaluated to improve the generalization and model complexity of traditional Flooding and SAM in classification tasks.
Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space
Leo Schwinn (Technical University of Munich), Stephan GΓΌnnemann (Technical University of Munich)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes and evaluates embedding space adversarial attacks on open-source large language models, demonstrating their ability to efficiently bypass safety alignment, recover 'forgotten' information, and extract pre-training data.
π― What it does: A soft superpixel neighborhood attention (SNA) module is proposed for image denoising tasks, which re-weights attention weights using pixel-level superpixel probabilities to better capture the variable boundaries of objects.
SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion
Lu Han (Nanjing University), De-Chuan Zhan (Nanjing University)
CodeComputational EfficiencyRecurrent Neural NetworkGraph Neural NetworkTransformerReinforcement LearningTime Series
π― What it does: A multivariate time series forecasting model SOFTS based on MLP is proposed, which introduces the STAR module to capture inter-channel correlations for efficient prediction.
π― What it does: A framework for solving inverse problems based on discrete optimal control (Diffusion Optimal Control) is constructed, treating the reverse diffusion process as controllable dynamics and directly searching for samples in the control space that satisfy observation constraints.
π― What it does: A two-stage compression framework for the multi-output regression (SHORE) problem with high-dimensional sparse outputs is proposed, which first compresses the outputs through random projection and then trains and performs projection gradient descent for prediction.
Source Code Foundation Models are Transferable Binary Analysis Knowledge Bases
Zian Su (Purdue University), Xiangyu Zhang (Purdue University)
CodeGenerationAI Code AssistantTransformerLarge Language ModelMultimodalityRetrieval-Augmented Generation
π― What it does: This paper proposes a 'probe-and-recover' framework called ProRec, which combines a binary-source code cross-modal alignment encoder-decoder (Prober) with a black-box LLM (Recoverer) to automatically generate symbol-rich source code snippets as context, enhancing the effectiveness of binary reverse engineering tasks (summary and function name recovery).
Sourcerer: Sample-based Maximum Entropy Source Distribution Estimation
Julius Vetter (University of TΓΌbingen), Jakob H. Macke (Max Planck Institute for Intelligent Systems)
CodeBiomedical Data
π― What it does: This paper proposes a source distribution estimation method (Sourcerer) based on the maximum entropy principle, which infers the parameter distribution of scientific simulators from observational data without requiring explicit likelihood.
π― What it does: The SpaFL framework is proposed, which achieves structured sparsity through trainable thresholds, only communicating thresholds in federated learning rather than parameters, significantly reducing communication and computation costs while improving model accuracy.
π― What it does: This paper proposes a Sparse Bayesian Generative Modeling approach for linear inverse problems in compressed sensing, which can learn from a small number of noisy compressed samples and directly solve inverse problems without optimization.
Kartikeya Bhardwaj (Qualcomm AI Research), Markus Nagel (Qualcomm AI Research)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageText
π― What it does: A highly sparse high-rank adapter (SHiRA) is proposed, which can complete various generation tasks by fine-tuning only 1-2% of the parameters of the pre-trained model.
Sparse maximal update parameterization: A holistic approach to sparse training dynamics
Nolan Simran Dey (Cerebras Systems), Joel Hestness (Cerebras Systems)
CodeTransformerLarge Language ModelText
π― What it does: Proposed Sparse Maximum Update Parameterization (SΒ΅Par) to stabilize the training dynamics of sparse networks and achieve performance surpassing that of dense networks.
SparseLLM: Towards Global Pruning of Pre-trained Language Models
Guangji Bai (Emory University), Liang Zhao (Argonne National Laboratory)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: Proposes the SparseLLM framework, which breaks down the global pruning of large-scale LLMs into manageable subproblems, achieving globally optimal pruning under low resources;
π― What it does: An efficient spatiotemporal interactive network (STIR) is proposed for reconstructing high-quality intermediate frames from the binary spatiotemporal pulse stream of an integrated neuromorphic vision camera.
Speaking Your Language: Spatial Relationships in Interpretable Emergent Communication
Olaf Lipinski (University of Southampton), Timothy J. Norman (University of Southampton)
CodeRecognitionExplainability and InterpretabilityRecurrent Neural NetworkAgentic AISequential
π― What it does: In a reference game environment, agents communicate spatial relationships of target positions in a sequence through discrete messages and analyze their interpretability.
π― What it does: An algorithm named SPEAR is proposed, which can accurately reconstruct input data with a batch size greater than 1 in federated learning, challenging the previous assumption that this is not achievable in an honest-but-curious setting.
π― What it does: Proposes the Spectral Adapter, which achieves low-parameter fine-tuning by performing incremental or orthogonal rotations in the singular vector space of the pre-trained weights;
Spectral Editing of Activations for Large Language Model Alignment
Yifu QIU, Shay B Cohen
CodeGenerationOptimizationTransformerLarge Language ModelText
π― What it does: This paper proposes a training-agnostic activation editing methodβSpectral Editing of Activations (SEA), which guides the model to generate outputs that better align with human preferences (more accurate and fair) by projecting the internal representations of LLM during inference onto directions that are maximally correlated with positive examples (e.g., true) and minimally correlated with negative examples (e.g., false).
Spectral Learning of Shared Dynamics Between Generalized-Linear Processes
Lucine L Oganesian, Maryam Shanechi
CodeOptimizationTime Series
π― What it does: This paper proposes a multi-stage covariance-based system identification algorithm PGLDM, which is used to simultaneously model two generalized linear time series and explicitly separate shared and private dynamics.
Spectral-Risk Safe Reinforcement Learning with Convergence Guarantees
Dohyeong Kim (Seoul National University), Songhwai Oh (Seoul National University)
CodeOptimizationSafty and PrivacyReinforcement LearningTabular
π― What it does: A safety reinforcement learning algorithm based on spectral risk measures is proposedβSpectral-Risk-Constrained Policy Optimization (SRCPO), which ensures convergence and optimality in discrete (tabular) environments through bi-level optimization and achieves optimal performance in continuous control tasks.
SpeechForensics: Audio-Visual Speech Representation Learning for Face Forgery Detection
Yachao Liang (Institute of Information Engineering Chinese Academy of Sciences), Weiqing Huang (Institute of Information Engineering Chinese Academy of Sciences)
π― What it does: Proposes an unsupervised facial forgery detection method that identifies forgeries by detecting semantic mismatches between lip movements in the video and audio through audiovisual speech representation learning from real videos.
Weiyu Guo (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
CodeRecognitionDomain AdaptationSpiking Neural NetworkTime Series
π― What it does: A SpGesture framework based on spiking neural networks is proposed, which includes Spiking Jaccard Attention and Source-Free Domain Adaptation, specifically addressing the distribution shift problem in sEMG gesture recognition.
π― What it does: A LiDAR point cloud semantic segmentation method based on Spherical Frustum, SFCNet, is proposed, which eliminates the loss of geometric information caused by quantization in traditional spherical projection.
π― What it does: A self-supervised spike-guided motion deblurring framework S-SDM has been developed, which can utilize low-resolution spike flow to recover a continuous sequence of clear frames from a single blurred image.
π― What it does: A spike graph neural network (MSG) that operates in Riemannian geometric spaces is proposed, achieving energy-efficient learning of graph data.
π― What it does: An adaptive event stream segmentation method called SpikeSlicer is proposed, utilizing low-energy pulse neural networks (SNN) as event triggers.
π― What it does: A novel event-driven friendly spiking neural network architecture called STMixer is proposed, which achieves high-performance inference at T=1 by implementing token mixing and information-preserving patch segmentation using operations such as convolution, fully connected layers, and residual paths that are only supported on asynchronous hardware.
π― What it does: A NeRF-based inverse rendering framework is proposed, which utilizes split sum approximation to separate lighting and material, and learns pre-integrated lighting and occlusion factors through a single MLP, allowing for the simultaneous estimation of object geometry, material properties, and environmental lighting within a few hours.
π― What it does: This paper proposes Sequential Monte Carlo Policy Optimisation (SPO), which achieves adaptive improvement of policies by combining SMC planning with the EM framework.
π― What it does: A new offline imitation learning algorithm called SPRINQL is proposed, which can simultaneously utilize expert and multi-level suboptimal demonstrations for learning.
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A new structured linear transformation method SS1 is proposed, which accelerates the inference of the linear layers in deep learning models using random parameter sharing and GPU-friendly computation.
π― What it does: This paper proposes a pixel-level classifier called SSA-Seg, which is based on semantic and spatial adaptation. It uses rough masks to guide the original prototypes to shift towards the semantic and spatial centers of the test images, and enhances performance through online multi-domain distillation.
π― What it does: This paper proposes a diffusion model called SSDiff based on spatial-spectral subspace decomposition, which uses a dual-branch network to learn spatial details and spectral features separately, and achieves high-quality remote sensing image fusion through an alternating projection fusion module.
Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective
Yongxin Zhu (University of Science and Technology of China), Lidong Bing (University of Science and Technology of China)
CodeGenerationTransformerAuto EncoderImage
π― What it does: A discrete image tokenizer based on a stable latent space has been constructed, and a GPT-style causal Transformer generative model DiGIT has been trained in this space.
Stabilizing Linear Passive-Aggressive Online Learning with Weighted Reservoir Sampling
Skyler Wu (Booz Allen Hamilton), James Holt (Laboratory for Physical Sciences)
CodeTabular
π― What it does: A WRS-Augmented Training (WAT) method is proposed that stabilizes any passive-aggressive online learning algorithm with single-pass training and without the need for a hold-out set.
π― What it does: Developed the Stable-Pose adapter within the Stable Diffusion framework, achieving fine control of human poses through a coarse-to-fine hierarchical pose masking attention mechanism, enabling pose-guided text-to-image generation.
Star-Agents: Automatic Data Optimization with LLM Agents for Instruction Tuning
Hang Zhou (Huawei Noah's Ark Lab), Yunhe Wang (Huawei Noah's Ark Lab)
CodeOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
π― What it does: The Star-Agents framework is proposed, which generates diverse and high-quality instruction data through multi-agent collaboration and conducts dual model evaluation to enhance the instruction tuning effect of LLMs.
π― What it does: This paper proposes a domain generalization framework called START based on a state space model, which utilizes saliency-driven token-aware transformations to suppress domain features in the input dependency matrix, thereby enhancing the model's generalization ability to unseen domains.
π― What it does: A new state-time representation (SCR) method is proposed, which enhances the generalization ability of state representation in reinforcement learning by incorporating extensive temporal information into the update steps of dual-similarity metric learning.
π― What it does: This paper proposes a regularized variant of the original Kernel Kullback-Leibler (KKL) divergence, providing its closed-form expression and gradient, and implements Wasserstein gradient flow optimization on discrete measures; theoretical analysis gives bias bounds, finite sample bounds, and convergence relations with the original KKL; the effectiveness of the method is subsequently validated on synthetic datasets.
CodeOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes and implements the method of 'Stealth Edits', which can precisely correct the hallucination outputs of language models under specific prompts through fine-grained weight updates without retraining the model; it also reveals the model's susceptibility to malicious stealth attacks.
π― What it does: A compensation method called StepbaQ is proposed to correct the accumulation of quantization errors in quantized diffusion models, which can improve generation quality without modifying the quantization settings.
Stepwise Alignment for Constrained Language Model Policy Optimization
Akifumi Wachi (LY Corporation), Youhei Akimoto (University of Tsukuba)
CodeOptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper proposes a stepwise alignment approach for optimizing constrained language models (SACPO), which first aligns reward metrics and then aligns safety metrics, thereby achieving dual alignment of human values and safety constraints.
π― What it does: A new concept bottleneck model (SCBM) is proposed, which explicitly models the correlation between concepts using a multivariate normal distribution, and based on this, designs an intervention strategy based on confidence intervals.
π― What it does: This paper introduces random kernel regularization on deep kernel machines, enhancing their generalization ability for image classification tasks.
π― What it does: Construct a stochastic optimal control theory for diffusion bridges in infinite-dimensional Hilbert spaces, and based on this, design learning algorithms for bridge matching and Bayesian inference, solving the problem of density without explicit form caused by the lack of equivalent Lebesgue measure;
π― What it does: A stochastic Taylor derivative estimator (STDE) based on infinitesimal Taylor mode automatic differentiation is proposed, which can efficiently randomize differential operators of any order and any dimension;
π― What it does: A submodular optimization-based active 3D object detection framework called STONE is proposed, which significantly reduces labeling costs using a two-stage subset selection strategy.
π― What it does: This paper proposes a Probability Regret Bound (PRB) stopping rule based on Bayesian optimization, which uses a Gaussian process model to estimate the probability that the current point satisfies the Ξ΅-optimal condition and stops the search once a threshold is reached.
π― What it does: By treating the End-of-Utterance (EoU) delimiters in conversations as 'conversation attention absorption points', long dialogue histories are compressed, only caching these attention absorption points, thereby reducing computational load and memory usage, supporting continuous dialogues of over 200K statements.
Stress-Testing Capability Elicitation With Password-Locked Models
Ryan Greenblatt (Redwood Research), David Krueger (University of Cambridge)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
π― What it does: This paper studies a type of LLM called password-locked models, which only exhibit hidden capabilities when a specific password appears in the input; otherwise, their performance is weakened. The model is used to stress-test capability activation methods based on fine-tuning.
π― What it does: A new perspective synthesis method based on 3D Gaussian projection with few views (SCGaussian) is proposed, which matches prior learned 3D consistent scene structures.
Structured Learning of Compositional Sequential Interventions
Jialin Yu (University College London), Ricardo Silva (University College London)
CodeRecommendation SystemExplainability and InterpretabilityRecurrent Neural NetworkAuto EncoderSequential
π― What it does: This paper proposes an interpretable structured model for predicting behavioral changes under unknown combinations of sequential interventions (such as multi-time point policies or recommendations), and provides its identifiability theory and learnable algorithms.
Structured Matrix Basis for Multivariate Time Series Forecasting with Interpretable Dynamics
Xiaodan Chen (Harbin Institute of Technology), Zhijun Li (Harbin Institute of Technology)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTime Series
π― What it does: A multivariate time series forecasting model SUMBA based on structured matrix bases is proposed, which reduces variance and enhances interpretability by directly parameterizing dynamic spatial structures.