image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 3:Illustration of our motion-conditioned video generation model. The input image and bbox color frames are tokenized via a 3D-VAE encoder and then summed. The resultant tokens are concatenated with other conditional tokens, and fed into the DiT-based video generation model. | 2502.04299v1 | 3 |
|
Figure 7:Results when our method is applied for: (upper) motion transfer, and (bottom) video editing for changing objects, adding and removing objects. | 2502.04299v1 | 7 |
|
Figure 9:Visual comparison of the resulatant videos from DragAnything, MOFA-Video, TrackDiffusion, Ourscoordcoord{}_{\text{coord}}start_FLOATSUBSCRIPT coord end_FLOATSUBSCRIPT, and our MotionCanvas. | 2502.04299v1 | 9 |
|
Figure 13:Illustration of camera-aware transformation and camera-object-aware transformation. In preview videos, dash-line bounding boxes represent the scene-space inputs, while the solid ones with the same color denote the transformed screen-space motion signals. Similarly, point trajectories with white trace indicate scene-space user motion design, while colored ones represent transformed signals. Better investigation in supplementary videos. | 2502.04299v1 | 13 |
|
Figure 14:Illustration of the recomputation process for input motion conditions in our MotionCanvasARAR{}_{\text{AR}}start_FLOATSUBSCRIPT AR end_FLOATSUBSCRIPTduring inference. | 2502.04299v1 | 14 |
|
Figure 16:Designed user study interface. Each participant is required to evaluate 15 video comparisons and respond to three corresponding sub-questions for each comparison. Only one video is shown here due to the page limit. | 2502.04299v1 | 16 |
|
Figure 1:Action-Video Dynamics Model from Heterogeneous Robot Interactions.HMAutilizes heterogeneous datasets comprising over 3 million trajectories (videos) from 40 distinct embodiments to pre-train a full dynamics model with next-set-of-token predictions using masked autoregression. After pre-training, the resulting action-video dynamics model is versatile, supporting applications such as video simulation, policy evaluation, synthetic data generation, and direct adoption as an imitation policy. | 2502.04296v1 | 1 |
|
Figure 2:Dynamics Model.Masked autoregression in the dynamics model generalizes multiple problem settings including policy learning, forward and passive dynamics, and full dynamics. | 2502.04296v1 | 2 |
|
Figure 3:Network Architecture.TheHMAmodel architecture maps low-level video and action sequences across different embodiments into a shared latent space. For actions, embodiment projectors are activated based on the training sample. The spatial-temporal Transformer produces the output video and action tokens for future frames. | 2502.04296v1 | 3 |
|
Figure 4:Pre-trained Video Model Generation.We show that a single unifiedHMAmodel can generate realistic (left 3 columns) and diverse (right 3 columns) videos across multiple embodiment datasets with heterogeneous action spaces. Each group shows three generated frames from a single sequence. | 2502.04296v1 | 4 |
|
Figure 5:Ablation on Pre-training Settings and Architecture.Under the pre-training setting with VQ tokens, we ablate the video generation performance (visual fidelity measured by perplexity and controllability measured by controllability). (a) We find action-conditioned models outperform passive video models. (b) We compare different action conditioning architectures in the masked autoregression framework. The purple color denotes the best model that we use by default. | 2502.04296v1 | 5 |
|
Figure 6:Experiments on Scaling Behaviors ofHMA.We observe positive trends in the scaling performance of heterogeneous video models across axes including the number of datasets, number of trajectories, and model sizes. The evaluation metrics on fidelity (perplexity) and controllability (ΔΔ\Deltaroman_ΔPSNR) are averaged across validation datasets. | 2502.04296v1 | 6 |
|
Figure 7:Qualitative Comparisons Between Tokenizers and Models.Despite longer convergence time, diffusion-based methods (Eq.3) on soft tokens generate better visual quality than on VQ tokens (Eq.2), qualitatively and measured by PSNR. | 2502.04296v1 | 7 |
|
Figure 8:Video Controllability.HMAcan follow user action inputs to generate physically plausible object permanence (top row) and block pushing interactions (bottom row). These video predictions are both at out-of-distribution settings and at a much longer horizon than training (over 100 frames). | 2502.04296v1 | 8 |
|
Figure 9:Policy Evaluation withHMA.By learning the action-video dynamics over both successful and failed examples,HMAcan be used to evaluate policies, similar to a traditional simulator[46]. The autoregressive horizon at inference time is 10 times more than the training time horizon. | 2502.04296v1 | 9 |
|
Figure 1:The crucial role of prompt formatting and its interaction with content.(A): Model-specific format biases: Illustrates the performance sensitivity of two LLMs to different format styles on the GSM8K task, showing substantial variability in the effectiveness of 10 randomly selected formats.(B): For seven different prompt contents evaluated across 24 distinct formats, performance variations show the complex, interdependent relationship between prompt content and structure, demonstrating that no single format universally maximizes effectiveness. | 2502.04295v1 | 1 |
|
Figure 2:Illustration of the CFPO pipeline within a single iteration round. In the initial Component-wise Content Optimization stage, case-diagnosis and Monte-Carlo sampling are employed for content mutation. Subsequently, the Format Optimization stage identifies the most suitable format for each content candidate. The yellow dashed line indicates where the LLM optimizer is employed to guide the optimization process. | 2502.04295v1 | 2 |
|
Figure 4:Built-in formats and rendering effects in our initial format pool. The final format configuration is achieved by selecting and combining elements from both thePrompt Rendererand theQuery Formatcategories. | 2502.04295v1 | 4 |
|
Figure 1:Prediction-powered confidence sequences.The plot shows the p-landscape (i.e., parameter on the x-axis, reciprocal of the e-value on the y-axis) for the confidence sequence generated by our method (green), along with those for inference using only labelled samples (purple) and by using an imputation approach. The 95% confidence intervals for each p-landscape (i.e., region where the p-landscape is above 0.05) is shaded. Our method provides the tightest valid intervals – using only the labelled samples or vanilla PPI(Angelopoulos et al.,2023a)yields weaker inferences, and using imputation fails to cover the true mean. | 2502.04294v1 | 1 |
|
Figure 2:Prediction-powered anytime-valid hypothesis testing.The plot shows the e-values over time for testing two null hypotheses – one on the bottom, which should be rejected, and one on top, which should not be rejected. Our prediction-powered e-values provide the strongest valid signal for rejection (E≥20𝐸20E\geq 20italic_E ≥ 20for a significance level of 95%, marked by the dashed lines), as the imputation approach rejects before the null is actually violated; for non-rejection (E<20𝐸20E<20italic_E < 20), all the methods appear valid, but ours still attains the highest e-value. | 2502.04294v1 | 2 |
|
Figure 3:Prediction-powered change-point detection via e-values.The plot shows the exponential moving average of a time series (in blue), with the few collected labels denoted by the scattered Xs. Our prediction-powered methods detect the change-point accurately, while the base method that only considers the labelled data points does not detect any change-point. | 2502.04294v1 | 3 |
|
Figure 4:Prediction-powered causal discovery with e-values.We compare our prediction-powered causal discovery method with one that uses only labelled data. The lighter nodes correspond to the costly variables, while the darker nodes correspond to cheaper readily-available ones. The standard base method does not detect any edges in the causal graph (denoted by the dashed edges), while ours detects as many edges as the ‘best possible’ method, which uses all the data points regardless of data acquisition costs. | 2502.04294v1 | 4 |
|
Figure 1:Overview of the category-level Pose Estimation Pipeline: (A) Previous methods,i.g.AG-Pose[39]and Second Pose[7], rely on partial features extracted by a neural network (NN) to regress object poses. (B) We introduce a novel approach that leverages a semantic shape reconstruction (SSR) module for global feature extraction. This global context enhances (GCE) the mapping from partial features to NOCS features. | 2502.04293v1 | 1 |
|
Figure 2:Illustration of GCE-Pose: (A) Semantic and geometric features are extracted from an RGB-D input. A keypoint feature detector identifies robust keypoints and extracts their corresponding features. (B) An instance-specific and category-level semantic global feature is reconstructed using our SSR module. (C) The global features are fused with the keypoint features to predict the keypoint NOCS coordinates. (D) The predicted keypoints, NOCS coordinates, and fused keypoint features are utilized for pose and size estimation. | 2502.04293v1 | 2 |
|
Figure 3:Illustration of Deep Linear Semantic Shape Model. A Deep Linear Semantic Shape model is composed of a prototype shapec𝑐citalic_c, a scale network𝒮𝒮\mathcal{S}caligraphic_S, a deformation network𝒟𝒟\mathcal{D}caligraphic_D, a Deformation field𝒱𝒱\mathcal{V}caligraphic_Vand a category-level semantic featurescsemksubscriptsuperscript𝑐𝑘semc^{k}_{\text{sem}}italic_c start_POSTSUPERSCRIPT italic_k end_POSTSUPERSCRIPT start_POSTSUBSCRIPT sem end_POSTSUBSCRIPT. At stage 1, we build a Deep Linear Shape (DLS) model using sampled point clouds from all ground truth instances within each category, training a linear parameterization network to represent each instance. At stage 2, we retrain the DLS model to regress the corresponding DLS parameters from partial point cloud inputs using a deformation and scale network. During testing, the network predicts DLS parameters for unseen objects and reconstructs their point clouds based on the learned deformation field to get semantic reconstruction. | 2502.04293v1 | 3 |
|
Figure 4:Visualization of category-level object pose estimation results on HouseCat6D dataset[27]. Predicted 3D bounding boxes are shown in red, with ground truth in green. Challenging cases are highlighted in pink side squares. Leveraging our global context-enhanced pose prediction pipeline, GCE-Pose outperforms the SOTA AG-Pose[39](DINO), demonstrating robustness to occlusions and strong generalization to novel instances. | 2502.04293v1 | 4 |
|
Figure 5:(A) shows the visualization of input partial points and the output semantic shape reconstructions; (B) visualizes the semantic prototypes of different categories and the aggregated instance semantics. | 2502.04293v1 | 5 |
|
Figure 1:Retrosynthesis problem.Identifying the optimal precursor set for a given target material can be treated as a ranking problem. We use the binary classification probabilities of each set to determine its rank. Checkmarks indicate whether a ranked set corresponds to an experimentally verified synthesis. | 2502.04289v1 | 1 |
|
Figure 2:Learning paradigms for inorganic retrosynthesis(a) Multi-label classification-based approaches, which constitute current state-of-the-art models(Noh et al.,2024), inherently predict known precursorsP𝑃Pitalic_Pfrom a fixed candidate set. (b) Our approach (Retro-Rank-In) overcomes this limitation by embedding both precursor and target materials into a shared latent space and predicting their chemical compatibility in synthetic routes. This enables extrapolation beyond known precursors, allowing the model to propose novel synthesis pathways for unseen materials. The red links highlight an exemplary case for link prediction between a target and a precursor. | 2502.04289v1 | 2 |
|
Figure 3:Comparison of Top-K accuracy.Comparison of Retrieval-Retro and Retro-Rank-Inon the Novel Materials Systems dataset (c). We see the performance gap between both approaches widening, especially for larger K. | 2502.04289v1 | 3 |
|
Figure 4:MTEncoder architecture overview.This diagram illustrates the MTEncoder framework, where material compositions are tokenized and processed through a transformer model. | 2502.04289v1 | 4 |
|
Figure 5:Ablation for layers.Retro-Rank-In tested for various numbers of layers. Results show the robustness of our method regarding hyperparameter choice. | 2502.04289v1 | 5 |
|
Figure 6:PCA Visualization of target embeddings with rank-based coloring.PCA of target embeddings, where each point is color-coded based on rank. Higher-ranked points are shown in warmer tones, while lower-ranked ones appear in cooler shades, illustrating the distribution of rankings in the embedding space. | 2502.04289v1 | 6 |
|
Figure 7:Precursor pair correlation.The plot illustrates the correlation between pairs of precursors. Each point corresponds to a unique precursor pair, sorted along the x-axis by the strength of their correlation. Correlation is quantified here by the logarithm of the ratio of their joint probability to the product of their individual probabilities. | 2502.04289v1 | 7 |
|
Figure 8:Precursor set diversity.Comparison of the number of unique precursor combinations generated by Retrieval-Retro and Retro-Rank-In. | 2502.04289v1 | 8 |
|
Figure 9:Retro-Rank-In achieves higher diversity of predicted precursors.PCA plot of a MTEncoder-encoded target material (red color) and precursors predicted by Retrieval-Retro (green triangles, left) and Retro-Rank-In (blue crosses, right). The intensity/alpha of each point is proportional to the probability assigned by each model. Clearly, Retrieval-Retro assigns high probabilities to a small number of precursors, leading to low diversity. In contrast, Retro-Rank-In assigns significant probabilities to a higher number of precursors, leading to higher diversity. Importantly, this improvement in diversity does not come at the expense of accuracy, as shown in Table2. | 2502.04289v1 | 9 |
|
Figure 10:Distribution of predicted probabilities.A comparison of the top 60 highest predicted precursor probabilities across the test set of the Distinct Reactions dataset. Our approach demonstrates improved probability calibration compared to the previous state-of-the-art(Noh et al.,2024). We attribute this improvement to the class-balanced learning of a pairwise ranker, which translates to enhanced throughout performance at higher values ofK𝐾Kitalic_Kin Top-K𝐾Kitalic_Kexact match accuracy. | 2502.04289v1 | 10 |
|
Figure 2:Simplified illustration of the construction of the optimal hyperplane (line) in SVM. In2D2𝐷2D2 italic_D/3D3𝐷3D3 italic_Dthesupport lines/planespassing through the support vectors and parallel to the optimal hyperplane provide a simple illustration of the regions of classification and the margin in the SVM classification process. | 2502.04273v1 | 2 |
|
Figure 3:Confusion matrix for Quadratic Kernel, SVM, for a single inclusion detection showing 72.9% accuracy in prediction. Correct classifications are shown in blue and light blue along the main diagonal of the confusion matrix, while incorrect ones are displayed in red in the off-diagonal entries of the matrix. The darker the colour (blue if the classification is correct), the higher the number of classifications in that category. In an ideal confusion matrix, we would see a dark diagonal (indicating high accuracy) and light red off-diagonal elements (indicating fewer misclassifications). | 2502.04273v1 | 3 |
|
Figure 4:Confusion matrix for the test data for detecting the number of19.419.419.419.4mm inclusions in a tank showing32.2%percent32.232.2\%32.2 %accuracy in prediction. | 2502.04273v1 | 4 |
|
Figure 5:CF for detection the number of inclusions of radius=38.8absent38.8=38.8= 38.8mm showing31.3%percent31.331.3\%31.3 %accuracy in prediction. | 2502.04273v1 | 5 |
|
Figure 6:This figure depicts the tank and the inclusion present in the dataset. Here we use the opposite injecting pattern. Utilizing voltage and current data pairs, we generate the D-N matrix𝐋σsubscript𝐋𝜎\mathbf{L}_{\sigma}bold_L start_POSTSUBSCRIPT italic_σ end_POSTSUBSCRIPT, which we subsequently map to inclusion radii using our ANN. Our ANN is tasked with detecting whether the D-N matrix corresponds to a specific inclusion radius. | 2502.04273v1 | 6 |
|
Figure 7:Cross-entropy error per epoch for training, validation and test data. We observe that the best performance on the validation dataset is reached at epoch 52. The cross-entropy error on the validation dataset is 1.1253e-07 at this point. The considerable overlap between the validation and training performance curves indicates a well-generalised model with minimal overfitting. The similarity of these two curves suggests that the model is learning patterns that generalise well to unseen data. The green circle indicates the point of optimal performance. | 2502.04273v1 | 7 |
|
Figure 9:Confusion matrix for test and validation data showing 100% accuracy on test data for radii’s detection. | 2502.04273v1 | 9 |
|
Figure 10:Performance accuracy for inclusion radii detection when varying the number of measurementsM𝑀Mitalic_Mtaken from a 16 electrode setup andM=1,2,4,8,12,16𝑀12481216M=1,\>2,\>4,\>8,\>12,\>16italic_M = 1 , 2 , 4 , 8 , 12 , 16. | 2502.04273v1 | 10 |
|
Figure 11:Performance accuracy of inclusion radii detection with varying number of electrodes. Here we takeE𝐸Eitalic_Emeasurements from anE𝐸Eitalic_Eelectrode system yielding𝐋′σ∈ℝE2subscriptsuperscript𝐋′𝜎superscriptℝsuperscript𝐸2\mathbf{L^{\prime}}_{\sigma}\in\mathbb{R}^{E^{2}}bold_L start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_σ end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_E start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT, forE=2,4,8,12,16𝐸2481216E=2,\>4,\>8,\>12,\>16italic_E = 2 , 4 , 8 , 12 , 16. | 2502.04273v1 | 11 |
|
Figure 12:Heatmap of performance accuracy for Inclusion Radii detection. Setups with a single electrode are clearly not feasible, leaving the top row empty. Similarly, having the number of measurementsM𝑀Mitalic_Mexceed the number of electrodesE𝐸Eitalic_Eis not possible, as this would result in non-unique opposite injections. Consequently, the elements above the diagonal remain empty. | 2502.04273v1 | 12 |
|
Figure 13:Schematic representation of the process followed for anisotropy detection. This is an adapted version of Figure6to show the differences in the setup of the two problems. | 2502.04273v1 | 13 |
|
Figure 14:Confusion matrix for anisotropy detection. Class 1 refers to an isotropic tank and inclusion, and Class 2 refers to an anisotropic tank and inclusion. The observed testing accuracy stands at94.2%percent94.294.2\%94.2 %, denoting a94.2%percent94.294.2\%94.2 %detection rate of anisotropy. | 2502.04273v1 | 14 |
|
Figure 1:Schematic layout of the VDD accordion ansatz for 2-5 qubits. At the top of each diagram is the root node, and at the bottom is the terminal node. The VDD has variational parameters at each edge of the diagram. To obtain the amplitude probability of a given element of the canonical basis of the Hilbert space of a system ofn𝑛nitalic_nqubits (a bit string of lengthn𝑛nitalic_n), all that is needed is to take the path from the root node to the terminal node, where if a zero(one) is encountered in the bit string, the left(right) edge is taken; the probability amplitude is the product of all probability amplitudes lying on the edges of the specified path (seedefinitions1and2for details on how the probability amplitudes are defined and computed). The red path in highlighted in the 3 qubit case is related to an example that is explained in the main text. | 2502.04271v1 | 1 |
|
Figure 1:Overview of our approach. (a) We consider a full RLHF training setup, where a language model (LM) policy is iteratively refined through active data collection. Our goal is to develop an optimal response sampling method for preference labeling. (b) We introduce PILAF, which generates responses by interpolating between the current policy and a reference policy, balancing exploration and exploitation. (c) Our theoretical analysis shows that T-PILAF aligns the parameter gradient with the steepest direction for maximizing human values and achieves more favorable convergence in regions of high sensitivity. | 2502.04270v1 | 1 |
|
Figure 2:Reward-KL curve for Iterative DPO. All training runs start from the same model obtained at the end of the first iteration viaVanilla Sampling. Each dot represents an evaluation performed every 50 training steps. | 2502.04270v1 | 2 |
|
Figure 3:Reward-KL curve for Online DPO. Each dot represents an evaluation performed every 50 training steps. | 2502.04270v1 | 3 |
|
Figure 4:Online DPO with an overfitted initial policy. Each dot represents an evaluation performed every 50 training steps. Color saturation indicates the training step, with darker colors representing later steps. | 2502.04270v1 | 4 |
|
Figure 5:Online DPO with an overfitted initial policy. Full results of theFigure4. Each dot represents an evaluation performed every 50 training steps. Color saturation indicates the training step, with darker colors representing later steps. | 2502.04270v1 | 5 |
|
Figure 5:Online DPO with an overfitted initial policy. Full results of theFigure4. Each dot represents an evaluation performed every 50 training steps. Color saturation indicates the training step, with darker colors representing later steps. | 2502.04270v1 | 6 |
|
Figure 1:Cosine similarity between English and translated words across four languages (French, German, Spanish, and Chinese). The Chinese language exhibits notably lower cosine similarity compared to others. | 2502.04269v1 | 1 |
|
Figure 2:t-SNE visualization of word embeddings across five languages: English, French, German, Spanish, and Chinese. Related languages form overlapping clusters, while Chinese forms a distinct, isolated cluster. | 2502.04269v1 | 2 |
|
Figure 3:Sentence similarity of English to other languages in the BLOOM model. Learning performance decreases as layers increase. | 2502.04269v1 | 3 |
|
Figure 4:Sentence similarity of English to other languages in the QWEN model. Arabic shows improved performance compared to the BLOOM model. | 2502.04269v1 | 4 |
|
Figure 5:Hidden state analysis across layers in the BLOOM model. The hidden state value decreases sharply in the deeper layers, indicating degradation of useful representations. | 2502.04269v1 | 5 |
|
Figure 6:Comparison of Accuracy and F1 Score Across Languages for BLOOM-560m and BERT-base Multilingual Cased. | 2502.04269v1 | 6 |
|
Figure 1:Related methods, their principles for knowledge mining, whether using additional priors, and performance on DOTA-v1.0. | 2502.04268v1 | 1 |
|
Figure 2:Visual comparisons with state-of-the-art methods including: Point2Mask (2023)[21], PointOBB (2024)[33], PointOBB-v2 (2025)[38], PointOBB-v3 (2025)[67], and Point2RBox (2024)[66]. The boxes detected by our method (last row) wrap the objects more tightly. | 2502.04268v1 | 2 |
|
Figure 3:The training pipeline of Point2RBox-v2. Gaussian overlap loss and Voronoi watershed loss utilize the spatial layout (see Fig.4), while edge loss (see Sec.3.4), symmetry-aware learning (see Sec.3.5), and copy-paste (see Sec.3.6) further enhance the method. | 2502.04268v1 | 3 |
|
Figure 4:To illustrate the procedure of the three newly proposed loss functions and their impact on the learning results. (a) Gaussian overlap loss (see Sec.3.2). (b) Voronoi watershed loss (see Sec.3.3). (c) Edge loss (see Sec.3.4). | 2502.04268v1 | 4 |
|
Figure 1:Motivation and overview.Left:Theinter-modalcontrastive loss used in pretraining enforces paired images and texts to be at a given distancer𝑟ritalic_r(i.e.rdogandrcatsubscript𝑟𝑑𝑜𝑔andsubscript𝑟𝑐𝑎𝑡r_{dog}\text{ and }r_{cat}italic_r start_POSTSUBSCRIPT italic_d italic_o italic_g end_POSTSUBSCRIPT and italic_r start_POSTSUBSCRIPT italic_c italic_a italic_t end_POSTSUBSCRIPT) but does not encourageintra-modalalignment. Consequently, intra-modal similarity scores might not correspond to those of actual images and texts (i.e.d1<d2subscript𝑑1subscript𝑑2d_{1}<d_{2}italic_d start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT < italic_d start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT).Right:We show that the common practice of individually exploiting only one encoder is suboptimal and approaching intra-modal tasks (e.g.image-to-image retrieval)inter-modallyvia modality inversion improves performance. | 2502.04263v1 | 1 |
|
Figure 2:(a, b)Loss values and retrieval performance overOTIoptimization steps for different numbers of pseudo-tokensR𝑅Ritalic_R.(c)Distribution of pairwise image-image, text-image, andOTI-image cosine similarities. We consider theOTI-inverted features using four pseudo-tokens (R=4𝑅4R=4italic_R = 4) at two distinct optimization steps: theperformance peak(step 17) and thefinal step(step 1000). | 2502.04263v1 | 2 |
|
Figure A1:Distribution of pairwise dog-dog and dog-cat image similarities. Overlap highlights the intra-modal misalignment issue. | 2502.04263v1 | 4 |
|
Figure A2:Captions generated by pre-trained captioning models for an image from theℛℛ\mathcal{R}caligraphic_ROxford dataset. | 2502.04263v1 | 5 |
|
Figure 1:Our estimator achieves the same statistical precision as the standard estimator with up to 20% fewer samples. Each study is subsampled ton=75𝑛75n=75italic_n = 75. The figure shows the percentage sample size reduction needed to match the confidence interval width of the standard estimator using ours. | 2502.04262v1 | 1 |
|
Figure 2:H-Aipwcombines the standardAipwestimator, which relies on experimental data alone, with multiple competing estimators that replace the outcome regression with predictions from foundation models. By leveraging foundation models trained on a much larger sample, rather than estimating the outcome regression with the limited experimental data,H-Aipwsignificantly reduces the finite sample variance of the average treatment effect estimate. | 2502.04262v1 | 2 |
|
Figure 1:Our approach is effective across major Image-to-Image (I2I) architectures, including VQ-GAN[6], diffusion model[7], and autoencoders (see Section 5). The figure also presents a comparison with the state-of-the-art (SOTA) I2I unlearning algorithm. For retain samples, our method generates consistent images before and after unlearning, while the SOTA method generates inconsist output. On forget samples, our approach intentionally produces inaccurate or unreliable outputs, aligning with the expected behavior of realistic unlearning. | 2502.04260v1 | 1 |
|
Figure 2:Overview of our proposed approach. We maximize the loss on the forget samples (Dfsubscript𝐷𝑓D_{f}italic_D start_POSTSUBSCRIPT italic_f end_POSTSUBSCRIPT) to get unlearned model. We get the updated I2I model by further fine-tuning the unlearned model on the retain samples to maintain the performance. | 2502.04260v1 | 2 |
|
Figure 3:Comparison of various unlearning approaches, along with the retrained model, on forget samples that were poisoned with a ’+++’ sign using an AutoEncoder. The results clearly demonstrate that our method effectively unlearns the ’+++’ sign, producing outputs that are most similar to those of the retrained model. | 2502.04260v1 | 3 |
|
Figure 4:T-SNE analysis of the generated images with our unlearning framework. After unlearning in both the cases, the generated image from retain samples closely overlaps the ground truth, while the image generated from forget samples diverge from the ground truth. | 2502.04260v1 | 4 |
|
Figure 5:Results of cropping8×8888\times 88 × 8patch at the center of the image on VQ_GAN models. The results demonstrate that our model effectively removes the embedded ’+++’ pattern, not just replacing it with Gaussian noise. For retain samples, our method shows no signs of the embedded ’+++’ sign from the forget samples, in contrast to other baseline and benchmark methods, which often retain subtle remnants of the pattern. | 2502.04260v1 | 7 |
|
(a)Comparison of various unlearning approaches, along with the retrained model for an AutoEncoder. The results clearly demonstrate that our method produce outputs that are most similar to those of the retrained model. | 2502.04260v1 | 8 |
|
(a)Comparison of various unlearning approaches, along with the retrained model for an AutoEncoder. The results clearly demonstrate that our method produce outputs that are most similar to those of the retrained model. | 2502.04260v1 | 9 |
|
Figure 8:Results of cropping8×8888\times 88 × 8patch at the center of the image on diffusion model. The results demonstrate that our model effectively unlearns by producing incorrect/unreliable results instead of replacing it with Gaussian noise. For retain samples, our method does not reduces the quality of the images in contrast to the benchmark methods I2I SOTA. This happens due to the fine-tuning on the Gaussian noise. | 2502.04260v1 | 11 |
|
Figure 1:Impact of our proposed regularization term (ALR) in both training and fine-tuning on CIFAR-10.Adversarialℓ2subscriptℓ2\ell_{2}roman_ℓ start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTregularization (ALR) significantly improves generalization to the unforeseen StAdv attack when performing adversarial training forℓ2subscriptℓ2\ell_{2}roman_ℓ start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTrobustness. Using ALR when subsequently fine-tuning with only StAdv attack also decreases the drop inℓ2subscriptℓ2\ell_{2}roman_ℓ start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTrobustness. | 2502.04248v1 | 1 |
|
Figure 4.Posterior Student-t𝑡titalic_tprocess (on the right) and posterior Gaussian process (on the left).
We followed the same strategy and used the same parameters introduced to generateFigures2and3. | 2502.04247v1 | 4 |
|
Figure 1:Frame to motion profile strip. In green, the detected horizon line, in red the considered pixel belt. | 2502.04244v1 | 1 |
|
Figure 3:A motion profile with two consecutive lane changes with different shapes in time and space. | 2502.04244v1 | 3 |
|
Figure 1:Overview ofMAGAframework. Our method expands the original corpus through a two-stage synthesis process. Each document is reformulated to 5 new documents, achieving 3.9× token number expansion while maintaining diversity through massive (genre, audience) pairs. | 2502.04235v1 | 1 |
|
Figure 2:Implementation details. From a high-quality corpus, we sample a subset to serve as input for the LLM labeler and judger.
Through iterative filtering, we train and quantize SLM tool models for each stage to improve inference efficiency,
which are used to generate the reformulated corpus. | 2502.04235v1 | 2 |
|
Figure 3:The first and last two figures illustrate the training dynamics of EntireSet and SubSet data recipes, respectively.
Benchmark details are provided inAppendix D. | 2502.04235v1 | 3 |
|
Figure 4:Benchmark results and validation losses in the MAGA-Mix setting. The sensitivity to data repetition varies across capability domains, with knowledge dimension showing greater resilience. | 2502.04235v1 | 4 |
|
Figure 5:t-SNE visualization results. Base (left) maintains a distribution that overlaps with but extends beyond the original data.
Strict (middle) clusters also extend original data, but indicating limited diversity compared to left. Relaxed (right) shows significant distributional shift, explaining its poor performance. | 2502.04235v1 | 5 |
|
Figure 8:Detail evaluation results of EntireSet described inTable 4. MAGACorpus group demonstrats advantages over other groups across most evaluation sets, consistently across models of sizes. | 2502.04235v1 | 8 |
|
Figure 9:Detail evaluation results of Subset described inTable 4. As the model size increases, the performance gap between the upsampling group and MAGACorpus gradually widens in ARC, DROP, GSM8K, RACE, but with some variations observed in TriviaQA and WinoGrande. | 2502.04235v1 | 9 |
|
Figure 11:Random examples sampling from wheremean(lossdiffi)>0.5mean𝑙𝑜𝑠subscriptsuperscript𝑠𝑖diff0.5\text{mean}(loss^{i}_{\text{diff}})>0.5mean ( italic_l italic_o italic_s italic_s start_POSTSUPERSCRIPT italic_i end_POSTSUPERSCRIPT start_POSTSUBSCRIPT diff end_POSTSUBSCRIPT ) > 0.5, the synthetic-trained model fail to predict the tokens in later sequence positions. | 2502.04235v1 | 11 |
|
Figure 12:Corresponding cases sampled from Fineweb-Edu, which align with the loss patterns shown in Figure11, with higher loss by synthetic-trained model highlighted inred. | 2502.04235v1 | 12 |
|
Figure 2:Transformation of a multigraph of flights into a weighted directed graph. The multigraph (left) represents multiple flights between airports. In the weighted graph (right), edges are aggregated to show total flights as weights. | 2502.04233v1 | 2 |
|
Figure 3:Airport multigraph GAT Layer with multi-head attention for three different flights between nodes (SP,RJ), with alternating colors opacity for each flight. | 2502.04233v1 | 3 |
|
Figure 4:Feature importance for the CatBoost model on the airport network dataset, highlighting the relevance of graph-based features. | 2502.04233v1 | 4 |
|
Figure 5:Distribution of predicted delay values (ypredsubscript𝑦predy_{\text{pred}}italic_y start_POSTSUBSCRIPT pred end_POSTSUBSCRIPT) for the regression task using CatBoost. | 2502.04233v1 | 5 |
|
Figure 6:Distribution of actual delay values (ytestsubscript𝑦testy_{\text{test}}italic_y start_POSTSUBSCRIPT test end_POSTSUBSCRIPT) for the test set in the regression task. | 2502.04233v1 | 6 |
|
Figure 7:Airdelayweb-based simulation tool, showing predicted flight delays due to holding maneuvers. | 2502.04233v1 | 7 |
|
Figure 1:Quality-attribution performance trade-off curve across different watermarking strengths and the overall performance comparison on detection and attribution tasks.
Higher values on both axes indicate better performance. | 2502.04230v1 | 1 |
Subsets and Splits