image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Visualization of the attention mechanism.wijsubscript𝑤𝑖𝑗w_{ij}italic_w start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPTis the similarity of Queryi𝑖iitalic_iand Keyj𝑗jitalic_j. The similarities are used to weight the Values. | 2502.01659v1 | 1 |
|
Figure 2:Visualizing different masks from the Longformer (left and center) and BigBird (right) transformers. The white cells indicate masked terms that are not considered. The black cells are local attention, which can be dilated. The blue cells correspond to global attention. The orange cells are uniform random attention. | 2502.01659v1 | 2 |
|
Figure 5:Plotting the average log runtime performance from 15 benchmark runs of PyTorch’s FlashAttention implementation against our local attention with either constant window size (left plot) or constant sparsity (right plot). The x-axis shows increasing context length from left-to-right. | 2502.01659v1 | 10 |
|
Figure 6:Plotting the average runtime performance from 15 benchmark runs of different attention mechanisms from well-known masked transformer models as the context length increases along the x-axis. The left-most plot is Longformer with local and global attention; it was modelled with PyTorch’s SDP, a double kernel call of our local and global, and CSR. The middle plot is Longformer with dilated local and global attention; it was run with PyTorch’s SDP and our CSR function. The right-most plot is BigBird with local, global, and random attention; it was run with PyTorch’s SDP, a sequential kernel call of our local; global; and CSR functions, and just our CSR implementation. | 2502.01659v1 | 11 |
|
Figure 4:Training dynamics whenK=10𝐾10K=10italic_K = 10. The first row shows results for varyingE𝐸Eitalic_E, the second for differentB𝐵Bitalic_B, and the third for variousη𝜂\etaitalic_η. | 2502.00182v1 | 4 |
|
Figure 7:Performance under varying levels of data imbalance. The first row shows results using weighted aggregation, while the second row uses naive averaging. | 2502.00182v1 | 7 |
|
Figure 10:Test accuracy, test loss, and training loss forK=10𝐾10K=10italic_K = 10. The first row shows results for varyingE𝐸Eitalic_E, the second for differentB𝐵Bitalic_B, and the third for variousη𝜂\etaitalic_η. | 2502.00182v1 | 10 |
|
Figure 11:Toy example illustrating the loss landscape and update paths for two clients with IID and Non-IID datasets. | 2502.00182v1 | 11 |
|
Figure 2:Example stimuli illustrating the range of phosphene brightness levels shown to participants. A value of 0 represents complete darkness, 2 is the ideal brightness for a retinal prosthesis user, 5 is overly bright, and 10 is extremely bright, with white filling most of the stimulus area. | 2502.00177v1 | 2 |
|
Figure 3:A)Number of participants who significantly preferredhuman-in-the-loop optimization(HILO)over the naïve encoder (left) and thedeep stimulus encoder(DSE)without HILO (right), based on log odds less than 0 in a linear mixed-effects model.B)Example percepts generated by theHILOencoder, theDSEwithout HILO, and the naïve encoder for three participants across the three experimental conditions.C)Distribution of log odds forHILOacross the three experiments.D)Distribution of the two main user-specific parameters,ρ𝜌\rhoitalic_ρ(phosphene size) andλ𝜆\lambdaitalic_λ(axon-aligned elongation), colored by the log odds indicating preference forHILOin the main experiment.E)Medianmean squared error(MSE)over the course of optimization for each experiment, with shaded regions denoting the interquartile range (IQR).F)Proportion of duels where participant decisions matched those of the simulated agent. | 2502.00177v1 | 3 |
|
Figure 1:Example of an ARC task. All images that include both input and output are demonstration examples, the image that only includes input is the test example. In this particular task, the rule is that the output grid corresponds to representing the most common object, differentiated by different colors, in the input grid. | 2502.00174v1 | 1 |
|
Figure 2:The selected task is derived from the original ARC dataset where the objective is to connect rows where the pixels at both ends share the same color. Importantly, if all examples are rotated by 90 degrees, the goal of the task remains unchanged, except that the connections would then occur across the columns instead of the rows. | 2502.00174v1 | 2 |
|
Figure 3:Average correct matches across 3 distinct batch sizes when trained and tested for horizontal (pink) versus vertical (green) examples using the default CodeT5+ model. | 2502.00174v1 | 3 |
|
Figure 4:Visual illustration of token creation for vertical (top) and horizontal (bottom) examples. Tokens are generated linearly, with the “goal tokens” (representing the red color) appearing closer together in horizontal arrangements and farther apart in vertical ones. | 2502.00174v1 | 4 |
|
Figure 5:Average correct matches across 3 distinct batch sizes when trained and tested for vertical examples with the original PE (green) versus an altered PE which prioritizes relevant tokens (yellow) using the default CodeT5+ model. | 2502.00174v1 | 5 |
|
Figure 8:An illustration of the tokenization process used in the experiments. The original grid is first converted into a list format, where each number corresponds to a given color. In this case, 2 corresponds to red, 3 to green and 0 to black. The tokens are then generated in two ways: bracketed (left), where spatial structure is explicitly marked, and raw (right), where the grid is represented as a raw sequence of tokens. Both versions include start (/s) and end (/e) tokens to mark the boundaries of the sequence. | 2502.00174v1 | 12 |
|
Figure 2:LBGconstructs an open-vocabulary 3D instance segmentation from a sequence of posed RGB images. A generic 2D instance segmentation model is used to segmentobjects,parts, andsubpartsin each RGB image. Semantic feature vectors are extracted for each region, and the masks are lifted to the per-pixel max-contributing Gaussian, generating per-frame 3D object fragments. These fragments are incrementally merged into coherent, scene-level 3D objects. By applying this process hierarchically to the part and subpart masks,LBGproduces a hierarchical decomposition of any 3DGS scene. | 2502.00173v1 | 2 |
|
Figure 3:Qualitative comparison on the LERF dataset for 3D Asset extraction. We show three extracted objects per scene, with two different views for each object. Compared to prior methods, the objects extracted from LBG are much cleaner and have fewer noisy artifacts. 3D objects from SAGA and Gaussian Grouping have missing parts and are of lower quality overall. | 2502.00173v1 | 3 |
|
Figure 4:Qualitative comparison on novel view synthesis for 2D instance masks. Black regions are unassigned. We see that our 2D masks are on par with other methods. LBG picks out instances across segmentation scales better than Gaussian Grouping. Compared to SAGA, our method provides more complete masks. | 2502.00173v1 | 4 |
|
Figure 5:Ablation on using CLIP features for merging. Using only spatial proximity leads to nearby objects being grouped together (red dashed boxes). When using DINO features together with CLIP this error is fixed. | 2502.00173v1 | 5 |
|
Figure 7:Additional results on novel view synthesis for 2D instance masks. For SAGA, we show images rendered at three levels: 0.1 (left), 0.5 (middle), and 1.0 (right). For our method, we show object level (left), part level (middle), and subpart level (right). | 2502.00173v1 | 7 |
|
Figure 8:Qualitative comparison on the 3DOVS dataset.Black regions are unassigned. In the bed scene, Gaussian Grouping merges hand and banana objects together, resulting in segmentation failure. Similarly, LangSplat fails to segment the white sheet due to low contrast in the feature space. Our method shows cleaner boundaries compared to both baselines. | 2502.00173v1 | 8 |
|
Figure 9:LBG Segmentation on 2DGS. 2DGS with colored Gaussians according to instance IDs (left) and individually extracted meshes (right). | 2502.00173v1 | 9 |
|
Figure 11:Visualization of the number of views which see each Gaussian. Notice how many structured floaters are only seen by a single view, showcasing visual artifacts from single-view over-fitting. | 2502.00173v1 | 11 |
|
Figure 12:Lifting DINOv2 features onto Gaussians. Using our Lifting-by-Gaussians approach, we lift DINOv2 features and visualize the first 24 PCA components. | 2502.00173v1 | 12 |
|
Figure 1:SQFA uses a geometric proxy of discriminability.Left. SQFA maps then𝑛nitalic_n-dimensional data into anm𝑚mitalic_m-dimensional
feature space using the linear filters𝐅𝐅\mathbf{F}bold_F. The class-specific
statistics in the feature space are represented as
points in theSPD(m)SPDm\mathrm{SPD(m)}roman_SPD ( roman_m )manifold (which is an open cone).
The distances inSPD(m)SPDm\mathrm{SPD(m)}roman_SPD ( roman_m )are a proxy for discriminability.Right. Each point inSPD(m)SPDm\mathrm{SPD(m)}roman_SPD ( roman_m )(top) corresponds to
a second-moment ellipse (below). As the distance inSPD(m)SPDm\mathrm{SPD(m)}roman_SPD ( roman_m )increases,
the second-order statistics become more different. | 2502.00168v1 | 1 |
|
Figure 2:Toy problem 1. Ellipses represent 6D Gaussian-distributed raw data for each
of three different classes (colors). Each panel shows two data-dimensions,
where the classes are distinguished by different statistical properties
(left: covariances; middle: means; right: neither). We learned two
filters each with SQFA, drLDA, and PCA (arrows). Each
dimensionality reduction technique prefers a different subspace. | 2502.00168v1 | 2 |
|
Figure 3:Toy problem 2. Ellipses represent 4D Gaussian-distributed
raw data for each of three different classes (colors). The second-moment
matrices are almost identical for dimensions 1-2 and 3-4, but
they are slightly more different in dimensions 3-4 than in
dimensions 1-2. The differences in the means make dimensions 1-2
much more discriminative. SQFA learns features that select for the
subspace in which the classes are most discriminable, regardless of
whether the useful information is in the means or covariances.
smSQFA features select for the subspace most discriminative
in terms only of its second-moments. | 2502.00168v1 | 3 |
|
Figure 4:SQFA extracts useful features using class-conditional second-order
statistics, outperforming related methods.Top. Example images
from SVHN.Center. Filters learned with different methods.Bottom. QDA accuracy using the features learned by the
different methods. | 2502.00168v1 | 4 |
|
Figure 5:SQFA can exploit class-conditional first- and second-order information.Top. Example images from MNIST.Center. Filters learned with different methods.Bottom. QDA accuracy using the features
learned by the different methods. | 2502.00168v1 | 5 |
|
Figure 6:SQFA features are close to optimal for quadratic discrimination.Top. 4 example videos with the same speed (left)
and 4 example videos with different speeds (right).
Each video is shown as a 2D space-time plot where the vertical axis is time
and the horizontal axis is space.Center. Filters learned different methods (each image shows
a 2D space-time plot).Bottom. Performance of a QDA classifier using the filters
learned by the different methods. | 2502.00168v1 | 6 |
|
Figure 7:We learned 2 filters in the full 4 dimensional data
space using different distances for the smSQFA objective. The
ellipses show the distributions of the 3 classes, with the
class indicated by color. The filters learned are shown as
arrows overlayed on the data. The distance used to learn the filters
is indicated in the title of each panel. | 2502.00168v1 | 7 |
|
Figure 8:Accuracy of a QDA classifier using features learned by SQFA
with different distances inSPD(m)SPDm\mathrm{SPD(m)}roman_SPD ( roman_m ). The same number of filters
as in the main text were learned for each distance.Left. SVHN dataset.Center. Speed estimation dataset.Right. MNIST dataset.
The distances are
labeled as AIRM (affine-invariant), BW (Bures-Wasserstein),
LE (Log-Euclidean), E (Euclidean) and KL (symmetric Kullback-Leibler). | 2502.00168v1 | 8 |
|
Figure 9:Training times for the different models on the different datasets.
Times are in seconds, and indicated by the numbers on top of each
bar. The y-axis is logarithmic.Left: SVHN.Center:MNIST.Right: Speed estimation dataset.
We note that LMNN was trained on reduced datasets compared
to the other methods (seeSection2.7). Also, LMNN training times
where highly variable, unlike the other methods. | 2502.00168v1 | 9 |
|
Fig. 4:Comparison of test accuracy for each transfer model with pretrained task and models trained from scratch | 2502.00160v1 | 4 |
|
Figure 1:Histogram of the cosine similarity between editing and unlearning task gradients in one epoch (left for TOFU in-profile dataset and right for TOFU out-profile dataset). | 2502.00158v1 | 1 |
|
Figure 6:Stability of performance during training with and without the proposed gradient penalty objective. | 2502.00156v1 | 35 |
|
Figure 1:A diagram of our method, showing how LLM hidden states are converted into compositional neurosymbolic representations. The encoder network converts the LLM hidden state to a neurosymbolic vector which can be queried to obtain the ones, tens, and hundreds digit of each number, as well as the type of problem being asked. This information is used by the neurosymbolic algorithm to find a solution to the problem, which the decoder converts from a neurosymbolic vector into an LLM hidden state vector, which is then added to the original LLM hidden state. | 2502.01657v1 | 1 |
|
Figure 4:Histogram of maximum similarity of queried problem type across all problem types, segregated per training and non-training problems | 2502.01657v1 | 6 |
|
Figure 1.Three paradigms in deep learning: supervised, unsupervised, and reinforcement. (a) In supervised learning, models are trained on labeled data to mostly perform the task of classification (where results are categories) and regression (where results are continuous values); (b) in unsupervised learning, models are trained on unlabeled data to identify associated relationships in a dataset, insightful cluster data group, and reduce high dimensional data form. | 2502.00151v1 | 1 |
|
Figure 2.Neural network models: (a) Fully connected neural network; (b) Convolutional neural network; (c) Recurrent neural network (RNN). | 2502.00151v1 | 2 |
|
Figure 1:Visual examples highlighting challenges in detecting polyps: (a) small polyps, (b) extremely large or abnormal polyps, and (c) irregularly shaped or multiple polyps. | 2502.00133v1 | 1 |
|
Figure 2:The overview of the transfer learning pipeline for deep learning-based polyp detection in colonoscopy images. | 2502.00133v1 | 2 |
|
Figure 1:ProtoSnapapplied to a full tablet by cropping each sign using existing bounding boxes (such as those depicted in unique colors), and matching prototypes of the signs (illustrated in the center). Our technique “snaps” the skeletons of the prototypes to the target images depicting real cuneiform signs.
These aligned results can be used to produce an automatic digitalhand copy(right).
We also show that our approach can be used to boost performance of cuneiform sign recognition. | 2502.00129v1 | 1 |
|
Figure 2:Method Overview. Given a prototype image with annotated skeleton and a target image of a real cuneiform sign,ProtoSnapfirst extractsbest-buddycorrespondences from deep diffusion features (extracted with our fine-tuned SD-model), globally aligning the target image to the skeleton of the prototype. Our method then “snaps” the individual strokes into place with a local refinement stage by optimizing a per-stroke transform. | 2502.00129v1 | 2 |
|
Figure 3:DIFT-Based Best-Buddies Correspondences. Noised images are passed through our fine-tuned denoising diffusion model SD-to extract deep Diffusion Features (DIFT), used to calculate the 4D similarity volumeS𝑆Sitalic_S. For each region(i,j)𝑖𝑗(i,j)( italic_i , italic_j )in the target image,
we examine the 2D sliceS[i,j,⋅,⋅]𝑆𝑖𝑗⋅⋅S[i,j,\cdot,\cdot]italic_S [ italic_i , italic_j , ⋅ , ⋅ ], and determine the indices(k,ℓ)𝑘ℓ(k,\ell)( italic_k , roman_ℓ )which maximize its value.
Symmetrically, for each region(k,ℓ)𝑘ℓ(k,\ell)( italic_k , roman_ℓ )in the prototype we find the corresponding region in the target by maximizing the 2D sliceS[⋅,⋅,k,ℓ]𝑆⋅⋅𝑘ℓS[\cdot,\cdot,k,\ell]italic_S [ ⋅ , ⋅ , italic_k , roman_ℓ ]. If these two regions correspond to each other, they are identified asbest buddies. | 2502.00129v1 | 3 |
|
Figure 5:Qualitative alignment results, aligning the prototypes (first row) to target cuneiform images (second row). We demonstrate the results after performing global alignment (third row), and the final result after local refinement (fourth row). As illustrated above, the global alignment stage provides a coarse placement of the prototype template, while the refinement stage allows each stroke to slightly diverge from the original prototype, resulting in more accurate alignments. | 2502.00129v1 | 5 |
|
Figure 7:Limitations of our method, illustrating examples with significant deformations from the prototype skeleton (left) and structurally different sign variants (middle) and corrupt sign image (right). We visualize correctly-aligned strokes ingreen, and misaligned strokes inred. | 2502.00129v1 | 7 |
|
Figure 8:Examples for data generated using our fine-tuned ControlNet model CN-, where the control is an image of a prototype sign (with two different added transformations). | 2502.00129v1 | 8 |
|
Figure 9:Examples ofProtoSnapapplied on photographed cuneiform signs of varying structure, illumination conditions and degrees of intactness. | 2502.00129v1 | 9 |
|
Figure 10:Results ofProtoSnapon our manually annotated test set, with DIFT and PoseAnything[Hirschorn and Avidan,2023]shown for comparison. We can see that our method produces alignments which are much closer to expert annotations and is generally less sensitive to outliers. | 2502.00129v1 | 10 |
|
Figure 11:ProtoSnapapplied on images from a different dataset and language (Hittite), showing that the method is robust and generalizable to various usages of cuneiform writing system. The 3 images on the left show signs from types (names) unseen in the training data, further emphasizing the generalizability of the method. | 2502.00129v1 | 11 |
|
Figure 4:The movement of the different language and gender samples in
and out of the predominant Spanish language index. | 2502.00127v1 | 8 |
|
Figure 2:(Top) Plot of computation time versus rotation range, given a translation range of±0.2plus-or-minus0.2\pm 0.2± 0.2m. (Bottom) Plot of computation time versus translation range, given a rotation range of±45∘plus-or-minussuperscript45\pm 45^{\circ}± 45 start_POSTSUPERSCRIPT ∘ end_POSTSUPERSCRIPT. | 2502.00115v1 | 2 |
|
Figure 3:Image of the robot described in SectionV-Bpicking up a jug from a grocery shelf, after using DSES to correct the gripper pose. | 2502.00115v1 | 3 |
|
Figure 4:Reference arm tip point cloud (black) vs. observed arm tip point cloud (red) both (a) before pose alignment, (b) after pose alignment with DSES. While the pose delta is small, this difference impacts success of the resulting grasp. | 2502.00115v1 | 4 |
|
Figure 1:Cross entropy loss vs. training time over 4B tokens forGemma-2-2BTopK SAEs withwidth=18,432,L0=64formulae-sequencewidth18432subscriptL064\mathrm{width}=18,432,\mathrm{L_{0}}=64roman_width = 18 , 432 , roman_L start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 64. We find that our method (TopK + LoRA in the plot) outperforms an e2e SAE and vanilla TopK SAE. | 2501.19406v1 | 1 |
|
Figure 2:Visual representation of our method, with a local SAE trained on layer12121212and low-rank adapters trained on MLP and attention components on all layers. | 2501.19406v1 | 2 |
|
Figure 3:Cross entropy loss improvement (Top: absolute,Bottom:percentage of CE loss gap closed) using our method for Gemma Scope SAEs onGemma-2-2B.LeftScaling across sparsity with fixed width=16k and layer=12, we see the largest effect by percentage of our method at lower sparsities, but still substantial effect at higher sparsities as well.Middle:Scaling across width with fixedL0=68subscript𝐿068L_{0}=68italic_L start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 68and layer=12, the highest effect by percentage is at low width but again this is not a large effect.Right:Scaling across layer with fixedL0=68subscript𝐿068L_{0}=68italic_L start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 68and width=16k, the highest effect of our method by percentage is at layer9999but it is mostly unaffected by layer. | 2501.19406v1 | 3 |
|
Figure 4:Cross entropy loss improvement (Top: absolute,Bottom:percentage) for Gemma Scope SAEs of width16k16𝑘16k16 italic_kandL0subscript𝐿0L_{0}italic_L start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTclosest to70707070on Gemma-2-2B, 7B, and 27B. We find that our method works increasingly well on larger models. | 2501.19406v1 | 4 |
|
Figure 5:Cross entropy loss vs. training time forLlama-3.2-1Bwith TopK SAEs ofL0=64subscript𝐿064L_{0}=64italic_L start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 64and width 16384. Our method (TopK + LoRA) achieves lower CE loss sooner than e2e SAE or vanilla TopK SAEs | 2501.19406v1 | 5 |
|
Figure 6:Downstream cross entropy loss when multiple Llama Scope SAEs are inserted intoLlama-3.1-8Bat once. “Base” is the original loss without any fine-tuning, while “LoRA” is the loss after 15M tokens of LoRA training. | 2501.19406v1 | 6 |
|
Figure 7:The distribution of normalized log-likelihood change post-steering is more pronounced on positive examples for the LoRA model. Results shown for the SAE latent responsible for “Donald Trump”. | 2501.19406v1 | 7 |
|
Figure 8:Distribution of loss improvements and loss degredations across validation tokens. We see that more tokens have a loss improvement than degredation (although a substantial number have a degradation) and most loss improvements and degredations happen in a range of about 0.01 to 1ΔΔ\Deltaroman_ΔℒSAEsubscriptℒSAE\mathcal{L}_{\text{SAE}}caligraphic_L start_POSTSUBSCRIPT SAE end_POSTSUBSCRIPTnats. | 2501.19406v1 | 8 |
|
Figure 9:Plot ofℒSAEsubscriptℒSAE\mathcal{L}_{\text{SAE}}caligraphic_L start_POSTSUBSCRIPT SAE end_POSTSUBSCRIPTwhen running LoRA on just a single layer of Gemma Scope 2B. We find that LoRA on layers closer to the SAE layer do better, and that also LoRA on just the next layer achieves much of the loss reduction of training on all layers. | 2501.19406v1 | 9 |
|
Figure 10:Distribution plots of the change in normalized log-likelihood after steering for various SAE latents.Top leftis for machine learning (neuron 8421).Top rightis for San Francisco (neuron 2195).Bottom leftis for Donald Trump (neuron 13677).Bottom rightis for COVID-19 (neuron 17811). | 2501.19406v1 | 10 |
|
Figure 11:Change in average distance to original activations before and after applying LoRA; increases in cosine similarity (Top) and decreases in Euclidean distance (Bottom) are good. | 2501.19406v1 | 11 |
|
Figure 1:The stability ofq^^𝑞\hat{q}over^ start_ARG italic_q end_ARGin different calibration set sizes (results of RT method on CIFAR-10 with ResNet-18). When the calibration set size is greater than2,00020002,0002 , 000, the fluctuation ofq^^𝑞\hat{q}over^ start_ARG italic_q end_ARGcan be kept within a certain range. | 2501.19403v1 | 1 |
|
Figure 1:Cross-domain performance analysis on the Camel-Bench Benchmark.
Our AIN-7B achieves promising performance compared to significantly bigger models (GPT-4o and Gemini-1.5-Pro)
in both domain-specific and aggregate settings. Despite its smaller size, our AIN-7B achieves competitive performance across all 38 sub-domains with significantly superior capabilities on OCR & document understanding. | 2502.00094v2 | 1 |
|
Figure 2:AIN: A versatile LMM excelling in visual and contextual understanding across diverse domains, including VQA on complex topics, OCR for various fonts and handwriting, cultural insights (traditions, food, places), agricultural tasks (crop identification, fruit classification, disease detection), remote sensing (multi-scale objects), medical imaging (various modalities), and video analysis (animation, human activities). | 2502.00094v2 | 2 |
|
Figure 3:AIN compared to existing LMMs across CAMEL-Bench benchmark[1]domains:OCR: “OCR & Document Understanding”,Video: “General Video & Multi-Image Understanding”,RS: “Remote Sensing Understanding”,CDT:“Chart, Diagram & Table Understanding”,Agro.: “Agricultural Image Understanding”,Cultural: “Cultural-Specific Understanding”,Medical: “Medical Image Understanding”. | 2502.00094v2 | 3 |
|
Figure 4:Qualitative results demonstrating AIN’s comprehensive capabilities across diverse domains. The results show its proficiency in handling both multiple-choice and open-ended questions. Our proposed AIN exhibits robust performance in addressing queries related to visual attributes (shape, color, quantity), while maintaining appropriate response formats (single character, word, or complete sentence) according to task requirements. | 2502.00094v2 | 4 |
|
Figure 5:Comparison of AIN with GPT-4o[3]and LLaVA[11]across diverse tasks. The evaluation demonstrates AIN’s proficiency in handling both multiple-choice and open-ended questions while maintaining appropriate response formats. | 2502.00094v2 | 5 |
|
Figure 6:AIN human evaluation survey, illustrating assessment criteria and multi-domain questions designed to evaluate multi-task and complex reasoning capabilities. The survey includes evaluations on specific food items, road signs in low-resolution settings, celebrities, charts, remote sensing tasks, and other diverse topics to comprehensively assess performance across multiple domains and challenges. | 2502.00094v2 | 6 |
|
Figure 7:Nationality of AIN Survey Participants: Participants represent 17 Arab nations, with the highest contributions from Saudi Arabia (30%), followed by Egypt (25%), the UAE (13.3%), and Lebanon (13.3%). | 2502.00094v2 | 7 |
|
Figure 8:User Model Preferences. Participant preferences for the three models in the survey, with Model 1 (AIN (ours)) receiving 76% of the votes, GPT-4o 15%, and LLaVA 9%, demonstrating AIN’s strong performance. | 2502.00094v2 | 8 |
|
Figure 9:User Preferences for MSA and Local Dialects: The majority (74.3%) preferred MSA for reading and writing. An additional 11% are comfortable with MSA but favored their local dialects, while 4.3% found MSA challenging and preferred using their dialect. A further 10.5% reported difficulties unrelated to linguistic aspects. | 2502.00094v2 | 9 |
|
(a)Q1:\<ما هو لون الورقة المصابة بالبقعة البكتيرية؟>Domain:Agricultural Image Understanding / Plant diseases.Purpose:Ability to detect diseased plant areas and identify their color. | 2502.00094v2 | 10 |
|
(a)Q1:\<ما هو لون الورقة المصابة بالبقعة البكتيرية؟>Domain:Agricultural Image Understanding / Plant diseases.Purpose:Ability to detect diseased plant areas and identify their color. | 2502.00094v2 | 11 |
|
(b)Q2:\<ما هو شكل الطعام الموجود في الصورة؟>Domain:Cultural-Specific Image Understanding / Food.Purpose:Ability to recognize food and precisely determine its shape.). | 2502.00094v2 | 12 |
|
(c)Q3:\<كم عدد التقاطعات الموجودة في الصورة؟>Domain:Remote Sensing Image Understanding / Roads & Constructions.Purpose:Ability to identify specific constructions among similar ones. | 2502.00094v2 | 13 |
|
(d)Q4:\<يرجى الإجابة مباشرة بكلمة أو رقم واحد، هل الضوء أخضر؟>Domain:General VQA/ Binary Question.Purpose:Ability to identify tiny details in ambiguous scenes and answer binary questions. | 2502.00094v2 | 14 |
|
(a)Q5:\<كم عدد لافتات ممنوع الانعطاف إلى اليسار الموجودة؟>Domain:General VQA / Traffic Signs.Purpose:Ability to spot traffic signs at a distance and in low resolution. | 2502.00094v2 | 15 |
|
(a)Q5:\<كم عدد لافتات ممنوع الانعطاف إلى اليسار الموجودة؟>Domain:General VQA / Traffic Signs.Purpose:Ability to spot traffic signs at a distance and in low resolution. | 2502.00094v2 | 16 |
|
(b)Q6:\<ما هو النص المكتوب في الصورة؟ >Domain:OCR & Document Understanding.Purpose:Ability to discern Arabic characters and extract text from images. | 2502.00094v2 | 17 |
|
(c)Q7:\<كم عدد قطع ورق العنب الموجودة في الصورة؟>Domain:General VQA / Short Answer Question.Purpose:Ability to pinpoint the required item among several items + provide a short answer as required. | 2502.00094v2 | 18 |
|
(d)Q8:\<ما هو الحالة أو مستوى الحالة في هذه الصورة؟ >Domain:Medical Image Understanding / Diseases Diagnoses.Purpose:Ability to diagnose organ health by reasoning its condition (normal or abnormal) for a specific disease. | 2502.00094v2 | 19 |
|
(a)Q9:\<هل الفنان بداخل المربع الأحمر المحيط يدعى ريتشارد رومانوس؟>Domain:General VQA / Grounding and Celebrities.Purpose:Ability to determine a person’s identity in a specific location. | 2502.00094v2 | 20 |
|
(a)Q9:\<هل الفنان بداخل المربع الأحمر المحيط يدعى ريتشارد رومانوس؟>Domain:General VQA / Grounding and Celebrities.Purpose:Ability to determine a person’s identity in a specific location. | 2502.00094v2 | 21 |
|
(b)Q10:\<ما هي النسبة المئوية للنمو في أفريقيا؟>Domain:Chart, Diagram & Table Understanding / Bar Charts.Purpose:Ability to extract values from charts, even when not explicitly shown. | 2502.00094v2 | 22 |
|
Figure 13:Data verification and filtering pipeline for textual and visual data.Textual data underwent semantic similarity checks using LaBSE[29](80% threshold) and quality evaluation using BLEU[31](60% threshold), METEOR[32](80% threshold), and ROUGE[33](80% threshold). Visual data was screened for toxicity using LLavaGuard[34]policies with GPT-4o[3], discarding unsafe images to ensure quality and safety. | 2502.00094v2 | 23 |
|
Figure 16:Visual Data Toxicity Filtering. Using GPT-4o[3]and LLavaGuard[34]policies, about 96% of the data is classified as safe, while the remainder was deemed unsafe. The unsafe data was distributed across four categories: “Weapon, or Substance Abuse” (3.25%), “Hate, Humiliation, Harassment” (0.55%), “Animal Cruelty” (1.09%), and “Violence, Harm, or Cruelty” (0.55%). | 2502.00094v2 | 30 |
|
Figure 1:PS-LBs: Dynamic Regret vs. Environment Steps,d=5𝑑5d=5italic_d = 5,S=L=1𝑆𝐿1S=L=1italic_S = italic_L = 1,T=50000𝑇50000T=50000italic_T = 50000,100100100100actions. | 2501.19401v1 | 1 |
|
Figure 2:PS-GLBs: Dynamic Regret vs. Environment Steps,d=5𝑑5d=5italic_d = 5,S=L=1𝑆𝐿1S=L=1italic_S = italic_L = 1,T=50000𝑇50000T=50000italic_T = 50000,100100100100actions. | 2501.19401v1 | 2 |
|
Figure 3:PS-SCBs: Dynamic Regret vs. Environment Steps,d=5𝑑5d=5italic_d = 5,S=3𝑆3S=3italic_S = 3,L=4𝐿4L=4italic_L = 4,T=50000𝑇50000T=50000italic_T = 50000,50505050actions. | 2501.19401v1 | 3 |
|
Figure 4:Drifting Non-Stationarity: Dynamic Regret vs. Environment Steps, Left: LBs, Middle: SCBs, Right: GLBs. | 2501.19401v1 | 4 |
|
Figure 5:Final dynamic regret vs. radius of changeδ𝛿\deltaitalic_δ: Drifting Environment LBs,S=L=1𝑆𝐿1S=L=1italic_S = italic_L = 1,T=50000𝑇50000T=50000italic_T = 50000. | 2501.19401v1 | 5 |
|
Figure 1:Vintix: Self-Correction with Many-Shot ICRL.Many-shot ICRL consistently self-corrects online on the training tasks (87) nearing the demonstrators‘ performance across four domains. Optimal number of shots for many-shot ICRL are shown inside the bar for each task. One shot corresponds to one online episode. SeeSection3for more details and comparisons to action models based on expert distillation. | 2501.19400v1 | 1 |
|
Figure 2:Vintix: Approach Overview.Stage 1 (Noise Distillation) - approximating policy improvement trajectory by injecting gradually annealed uniform noise (seeAlgorithm1). Stage 2 (Cross-Domain Dataset) - combining collected multi-episodic sequences into shared cross-domain dataset for subsequent model training. Stage 3 (Algorithm Distillation) - running AD(Laskin et al.,2022)with collated{s,a,r}𝑠𝑎𝑟\{s,a,r\}{ italic_s , italic_a , italic_r }triplets on collected dataset. | 2501.19400v1 | 2 |
|
Figure 3:Continuous Noise Distillation trajectories.Aggregated normalized returns for the collected cross-domain dataset. Returns are normalized with respect to random and demonstrator scores, while trajectory lengths are reported as a fraction of their maximum values. | 2501.19400v1 | 3 |
Subsets and Splits