image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 7.Models from the same family (Llama) exhibit slightly lower population-level variability than models from different families. | 2501.19361v1 | 7 |
|
Figure 8.Humans in prior studies have higher individual originality scores than humans in our study for all three tests.For the AUT and DAT tests, a t-test for a difference in means (alternative hypothesis is that prior study has higher mean than ours) is significant at the0.010.010.010.01level but not the0.0010.0010.0010.001level: AUT hast(5064)=3.21,p=0.001formulae-sequence𝑡50643.21𝑝0.001t(5064)=3.21,p=0.001italic_t ( 5064 ) = 3.21 , italic_p = 0.001and DAT hast(206)=3.32,p=0.001formulae-sequence𝑡2063.32𝑝0.001t(206)=3.32,p=0.001italic_t ( 206 ) = 3.32 , italic_p = 0.001. For FF, the difference more significant:t(892)=6.91,p<0.0001formulae-sequence𝑡8926.91𝑝0.0001t(892)=6.91,p<0.0001italic_t ( 892 ) = 6.91 , italic_p < 0.0001. | 2501.19361v1 | 8 |
|
Figure 9.Our study responses have higher population-level variability than the prior study on the DAT and FF tests, but slightly lower variability on the AUT.We use t-tests to compare means of the two population-level originality distributions. Null hypothesis is that means are equal, and alternative is that they are not. For AUT, the prior study has higher mean variability,t(1046)=11.0,p<0.001formulae-sequence𝑡104611.0𝑝0.001t(1046)=11.0,p<0.001italic_t ( 1046 ) = 11.0 , italic_p < 0.001. For FF, our study has a higher mean,t(366689)=−28.9,p<0.001formulae-sequence𝑡36668928.9𝑝0.001t(366689)=-28.9,p<0.001italic_t ( 366689 ) = - 28.9 , italic_p < 0.001. For the DAT, our study also has higher mean,t(19832)=−19.6,p<0.001formulae-sequence𝑡1983219.6𝑝0.001t(19832)=-19.6,p<0.001italic_t ( 19832 ) = - 19.6 , italic_p < 0.001. | 2501.19361v1 | 9 |
|
Figure 1:Illustration of the energy loss phenomenon.During response generation,❶the energy loss222The energy loss of a given layer is defined as the difference between theL1subscript𝐿1L_{1}italic_L start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTnorms of its input and output hidden states.in RLHF model’s final layer tends to be higher than that in SFT models; ❷hackingRLHF responses exhibitexcessiveenergy loss compared tonormalones. | 2501.19358v2 | 1 |
|
Figure 2:Response comparison betweenEPPOand baselineson AlpacaFarm dataset, as assessed by GPT-4. By focusing solely on energy loss in the LLM’s final layer,EPPObenefits from a broader optimization landscape, outperforming existing RLHF algorithms that directly constrain output spaces, such as PPO with length penalty (PPO w/ LP) and PPO with KL penalty (PPO w/ KL). | 2501.19358v2 | 2 |
|
Figure 3:Energy loss in the final layer of various LLMs duringPPOandEPPOtraining, including Llama3-8B, Mistral-7B, and DeepSeek-7B. Observations:❶Energy loss in the LLM’s final layergradually increases as RL progresses.❷By penalizing the increase in energy loss ,EPPOeffectivelysuppresses the excessive energy loss increaseduring RL process. | 2501.19358v2 | 3 |
|
Figure 5:Win rate dynamics of RLHF models compared to SFT models during RL processacross various LLMs on AlpacaFarm dataset under GPT-4 evaluation. Win rate is calculated aswin+0.5∗tie𝑤𝑖𝑛0.5𝑡𝑖𝑒win+0.5*tieitalic_w italic_i italic_n + 0.5 ∗ italic_t italic_i italic_efor more accurate performance assessment. Observations: Comparison methods either show limited RL performance gains or significantly degrade in later RL stages indicating reward hacking. In contrast,EPPOeffectively mitigates reward hacking while significantly boosting final RLHF performance. | 2501.19358v2 | 5 |
|
Figure 7:The relationship between energy loss in the LLM’s final layer and response-context relevance, measured by the contextual dependency strength metric proposed for general dialogue task(Chen et al.,2024a). Observation:Excessively increased energy loss suppresses the contextual relevance of responses, which aligns with our theoretical analysis that significant energy loss increases tighten the upper bound of contextual relevance, potentially diminishing it in practice. | 2501.19358v2 | 7 |
|
Figure 8:Energy loss in the final layer of various LLMs during the RL process usingPPOand ourEPPOalgorithms.From top to bottom:The LLMs shown are Llama3-8B, Llama2-7B, Mistral-7B, and DeepSeek-7B, respectively.From top to bottom:The tasks are general dialogue and Summerization tasks, respectively. | 2501.19358v2 | 8 |
|
Figure 13:The win rate dynamics of RLHF model compared to SFT model during RL training on Llama3-8B, under GPT-4 evaluation. To better measure performance, we calculate the win rate aswin+0.5∗tie𝑤𝑖𝑛0.5𝑡𝑖𝑒win+0.5*tieitalic_w italic_i italic_n + 0.5 ∗ italic_t italic_i italic_e.From left to right and from top from bottom:The evaluation datasets are AlpacaFarm, Anthropic Helpful, Anthropic Harmless, and Reddit TL;DR datasets, respectively. | 2501.19358v2 | 13 |
|
Figure 16:Win rate (%) of RLHF model compared to SFT model using ourEPPOwith different hyperparameterη𝜂\etaitalic_η, evaluated under GPT-4. To better measure performance, we calculate the win rate aswin+0.5∗tie𝑤𝑖𝑛0.5𝑡𝑖𝑒win+0.5*tieitalic_w italic_i italic_n + 0.5 ∗ italic_t italic_i italic_e. | 2501.19358v2 | 16 |
|
Figure 17:GPT-4 prompts used for general dialogue evaluation, summarization evaluation, and hacking sample identification. | 2501.19358v2 | 17 |
|
Figure 18:Typical hacking samples in general dialogue and summarization tasks, withspecific hacking segments highlighted in orange. The first four samples are from the general dialogue task, while the last is from summarization. A common characteristic of most hacking samples is weakened contextual relevance, i.e., the LLM’s response gradually deviates from the user’s intent. | 2501.19358v2 | 18 |
|
Figure 19:Qualitative Example I of RLHF model outputs with different RLHF algorithms.Specific hacking segments are highlighted in orange, whileinformation covered by ourEPPObut missed by competing methods is highlighted in green. | 2501.19358v2 | 19 |
|
Figure 20:Qualitative example II of RLHF models output with different RLHF algorithms.Specific hacking segments are highlighted in orange, whileinformation covered by ourEPPObut missed by competing methods is highlighted in green. | 2501.19358v2 | 20 |
|
Figure 21:Qualitative example III of RLHF models output with different RLHF algorithms.Specific hacking segments are highlighted in orange, whileinformation covered by ourEPPObut missed by competing methods is highlighted in green. | 2501.19358v2 | 21 |
|
Figure 2:Proto-Elamite tablet MDP 26, 177 (Scheil1935; P008805) alongside its transliteration and converted readings for both numerals. The tablet is read right-to-left: M056∼similar-to\sim∼f is the sign in the top-right corner. Observe that 111.5/44.6 equals the 2.5:1 ratio noted in Section4.1.2. | 2502.00090v1 | 2 |
|
Figure 1:The pipeline ofELREAfor fine-tuning and inference.
The data points (solid and hollow circles) do not necessarily have a geometric correspondence to their gradient directions (arrows). | 2502.00089v1 | 1 |
|
Figure 1:InSciCapChallenge, models generate captions based on the figure and the figure-mentioning paragraph. | 2501.19353v1 | 1 |
|
Figure 2:ROUGE-2 normalized scores of each model across eight arXiv domains, highlighting similar trends and demonstrating the generalizability of the caption generation approaches. | 2501.19353v1 | 2 |
|
Figure 3:Circuitry for a single clause. Literals are denoted withl𝑙litalic_land TA-actions withi𝑖iitalic_i. | 2501.19347v1 | 3 |
|
Figure 2:GMVAE latent manifold of Navier-Stokes flow fields showing a continuous distribution of Reynolds numbers, with clusters corresponding to distinct physical states. | 2502.02605v1 | 2 |
|
Figure 1:Traditional multi-objective optimization forT𝑇Titalic_Tobjectives might select any point along the Pareto frontier, but in some situations like this any Pareto optimal point performs poorly on at least one objective. In situations where multipleK<T𝐾𝑇K<Titalic_K < italic_Tsolutions are allowed (■■\blacksquare■), we can sometimes optimize all objectives well. | 2501.19342v1 | 1 |
|
Figure 2:Diagram of example “coverage optimization” problem.MOCOBOseeks a set ofK=4𝐾4K=4italic_K = 4peptides to coverT=11𝑇11T=11italic_T = 11target pathogenic bacteria. Arrows from peptides to bacteria indicate which peptides successfully inhibit (or “cover”) each bacteria in the target set. We provide results for this problem setting inSection4. | 2501.19342v1 | 2 |
|
Figure 4:In vitroresults for the two best “template free” (TF1, TF2) and two best “template constrained” (TC1, TC2) runs ofMOCOBOfor the peptide design task.
Columns are the best/lowestin vitroMIC among theK=4𝐾4K=4italic_K = 4peptides found byMOCOBOfor each target pathogenic bacteria B1,…,,\ldots,, … ,B11 listed inTableA.5. (-) and (+) indicate Gram negative and Gram positive respectively. TF1 and TC1 correspond to the single runs ofMOCOBOshown inTable1andTableA.1respectively. Methods used to obtainin vitroMICs are provided inSectionA.1. | 2501.19342v1 | 4 |
|
Figure 6:Ablation study comparingMOCOBOto optimization performance where a known “good” partitioning of theT𝑇Titalic_Tobjectives intoK𝐾Kitalic_Ksubsets is available in advance. We individually optimizeK𝐾Kitalic_Ksolutions, one for each partition. | 2501.19342v1 | 6 |
|
Figure A.1:In vitroresults for the two best “template free” (TF1, TF2) and two best “template constrained” (TC1, TC2) runs ofMOCOBOfor the peptide design task.
Here, “best” means runs that achieved highest coverage scores according to the APEX 1.1 model.
Each row is a single peptide found by a run ofMOCOBO.
Row TFi𝑖iitalic_i-j𝑗jitalic_jindicates template free runi𝑖iitalic_i, peptidej𝑗jitalic_j. Similarly, TCi𝑖iitalic_i-j𝑗jitalic_jindicates template constrained runi𝑖iitalic_i, peptidej𝑗jitalic_j.
Columns are thein vitroMICs for each target pathogenic bacteria B1,…,,\ldots,, … ,B11 listed inTableA.5. (-) and (+) indicate Gram negative and Gram positive respectively. TF1 and TC1 correspond to the single runs ofMOCOBOshown inTable1andTableA.1respectively. Methods used to obtainin vitroMICs are provided inSectionA.1. | 2501.19342v1 | 7 |
|
Figure A.2:(Leftmost Panel)Initial hdr church image for the church test image variation of the “image tone mapping” task.(Four Rightmost Panels)Images obtained by transforming the hdr church image using the best covering set ofK=4𝐾4K=4italic_K = 4solutions found by a single run ofMOCOBO. | 2501.19342v1 | 8 |
|
Figure A.3:The 8 panels depict the 8 obstacle courses that the rover must navigate for theT=8𝑇8T=8italic_T = 8variation of the rover task, with obstacles colored in red.
The required starting point for the rover is a green point in the bottom left of each panel. The “goal” end point that the rover aims to reach without hitting any obstacles is the red point in the top right of each panel.
The line in each panel shows the best trajectory for navigating the obstacle course from among theK=2𝐾2K=2italic_K = 2covering trajectories found by a single run ofMOCOBO.
The first trajectory in the covering set is shown in magenta and successfully navigates obstacle courses 5, 6, 7, and 8(Bottom Row). The second is shown in blue and successfully navigates obstacle courses 1, 2, 3, and 4(Top Row). | 2501.19342v1 | 9 |
|
Figure 1:PEAP framework: we investigate the possibility of perceive everything as pixels. This framework aligns better with human perception reducing the need for excessive pre-processing. Evaluated on our benchmarkPixelWorld,PEAPboosts performance on multimodal tasks (e.g., websites, slides, documents) but struggles with complex, text-centric tasks (e.g., reasoning and coding). Larger models achieve better transferability between pixel- and token-based performance compared to smaller ones. We also observed that text and images exhibit similar attention patterns, and reduced the overhead of model reasoning through patch pruning byPEAP-Fast. | 2501.19339v1 | 1 |
|
Figure 2:The performance oftext-onlydatasets. The comparison is made between text input and synthesized image input. Most models demonstrate comparable performance on language understanding datasets such as SuperGLUE, GLUE, and ARC. However, notable performance disparities emerge between text-based input and synthesized image input on mathematical reasoning tasks (e.g., MMLU-Pro, GSM8K) and programming tasks (e.g., MBPP). Phi-3.5-Vision exhibits consistently poor performance across all vision tasks, primarily due to its insufficient instruction-following capabilities. | 2501.19339v1 | 2 |
|
Figure 3:The performance of thestructureddataset. We report all the subsets for the TableBench. In thesemisetting, questions were presented as text, while tables were rendered as synthetic images. We observed that for tasks involving reasoning (numerical reasoning) and coding (visualization subset), synthetic images yielded inferior performance compared to text. However, for tasks emphasizing semantic understanding, such as data analysis and fact checking, synthetic images achieved performance comparable to or even surpassing text. Additionally, we found that the semi approach often performed worse than either text or synthetic images individually, providing insights into potential limitations and future directions for leveraging vision-language models (VLMs). | 2501.19339v1 | 3 |
|
Figure 4:The performance of themultimodaldataset (MMMU-Pro). We adopt the result reported by the origin paper. We can observe that strong models perform better inPEAP. | 2501.19339v1 | 4 |
|
Figure 5:The performance of themultimodaldatasets (except MMMU-Pro). We compare text-only and vision-only subsets in Mathverse, while SlidesVQA and WikiSS-QA are evaluated as VQA tasks. Larger models perform better on text-based tasks with more modalities. GPT-4o tends to generate longer responses in long-context QA, leading to performance degradation on WikiSS-QA. | 2501.19339v1 | 5 |
|
Figure 6:Last Layer Attention Heatmap on QWen2VL-7B between token-based (left) and pixel-based (right) inference. | 2501.19339v1 | 6 |
|
Figure 1:Data was collected using facial stimuli representing two racial and two gender groups. Three uncertainty measures—entropy, perplexity, and probability of differentiation—were calculated for each token position, with aggregate measures compared across groups. | 2501.19337v1 | 1 |
|
Figure 2:Sample facial stimuli representing racial and gender groups. Two faces are shown for each group, arranged from left to right: Black men, Black women, White men, and White women. | 2501.19337v1 | 2 |
|
Figure 3:Comparison of measures of uncertainty in token sampling probability distributions between Black and White Americans, aggregated across 50 token positions. Statistical analysis revealed significantly lower uncertainty measures for Black Americans across all three metrics. | 2501.19337v1 | 3 |
|
Figure 4:Comparison of measures of uncertainty in token sampling probability distributions between men and women, aggregated across 50 token positions. Statistical analysis revealed significantly lower uncertainty measures for women across all three metrics. | 2501.19337v1 | 4 |
|
Figure 5:Comparison of measures of uncertainty in token sampling probability distributions between Black and White Americans across token positions. The plot shows consistently higher uncertainty measures for White Americans throughout the generation sequence, suggesting that the model samples tokens more deterministically when generating texts about Black Americans. | 2501.19337v1 | 5 |
|
Figure 6:Comparison of measures of uncertainty in token sampling probability distributions between men and women across token positions. The plot shows consistently higher uncertainty measures for men throughout the generation sequence, suggesting that the model samples tokens more deterministically when generating text about women. | 2501.19337v1 | 6 |
|
Figure 3:Unemployment durationThe red line marks the 12 month threshold used to classify a jobseeking episode as long-term unemployment (LTU) in Germany. | 2501.19334v1 | 3 |
|
Figure 5:Policy value across different screening capacities (α𝛼\alphaitalic_α) and worst-off fractionsβ𝛽\betaitalic_βevaluated on the test set using theCatBoostregression model. Aβ𝛽\betaitalic_βvalue of0.150.150.150.15corresponds to the12121212-month cutoff used to define long-term unemployment in Germany. | 2501.19334v1 | 5 |
|
Figure 8:TheR2superscript𝑅2R^{2}italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTvalue on the test set for varying training set size (CatBoostRegression). | 2501.19334v1 | 12 |
|
Figure 10:Prediction-Access Ratio forR2=0superscript𝑅20R^{2}=0italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT = 0andΔR2=Δα=0.01subscriptΔsuperscript𝑅2subscriptΔ𝛼0.01\Delta_{R^{2}}=\Delta_{\alpha}=0.01roman_Δ start_POSTSUBSCRIPT italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_POSTSUBSCRIPT = roman_Δ start_POSTSUBSCRIPT italic_α end_POSTSUBSCRIPT = 0.01. | 2501.19334v1 | 14 |
|
Figure 11:V(R2+ΔR2)−V(R2)𝑉superscript𝑅2subscriptΔsuperscript𝑅2𝑉superscript𝑅2V(R^{2}+\Delta_{R^{2}})-V(R^{2})italic_V ( italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT + roman_Δ start_POSTSUBSCRIPT italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_POSTSUBSCRIPT ) - italic_V ( italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT )forCatBoostmodel andΔR2=0.05subscriptΔsuperscript𝑅20.05\Delta_{R^{2}}=0.05roman_Δ start_POSTSUBSCRIPT italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_POSTSUBSCRIPT = 0.05. | 2501.19334v1 | 15 |
|
Figure 1:Comparison with frame independent colorization framework TCVC[9]. Our method shows superiority in color vividness and temporal consistence. | 2501.19331v1 | 1 |
|
Figure 2:Overview of the video colorization pipeline with palette-guidance. We re-purpose a pre-trained video diffusion model and augment it with palette guidance. The left side illustrates three methods for generating palettes. During training, the palette is generated in the manner depicted in (a), with the reference image being a randomly selected frame from each training clip. | 2501.19331v1 | 2 |
|
Figure 4:Visual effects of ablation studies. The first column (NULL) shows the colorization results without palette guidance. The other columns display the colorization results of model variants with different palette generation methods. | 2501.19331v1 | 4 |
|
Figure 1:In this paper, we propose using human freehand sketches to improve image segmentation in challenging scenes—camouflaged object detection. (a) Comparison of SAM and our method with different inputs. While SAM struggles with point, box, and scribble inputs, sketch input (drawing a rough contour of the object) improves performance with extra annotation time. The blue area shows user response time, and green lines represent prompt types. (b) Practical applications of our method. Our method as an alternative to pixel-by-pixel annotation, producing comparable results, whereas SAM struggles with accuracy using point or box prompts. | 2501.19329v1 | 1 |
|
Figure 2:(a) Overall structure of DeepSketchCamo. Sketch augmentation, boundary refinement, and adaptive focal loss are introduced in this sketch-based task and are used in training to get elevated performance. The GT mask is used for supervision. (b) The application of our model. After the model has been trained, we input an image and a rough sketch into the model, the model can generate a predicted mask which can be directly used to train other deep-learning models. | 2501.19329v1 | 2 |
|
Figure 3:Visualization for sketches augmentation. Sketches generated by different amplitudes of variation of the Bezier curve. | 2501.19329v1 | 3 |
|
Figure 4:The visualization results of camouflaged image segmentation, our approach significantly enhances the performance of object segmentation, such as accurately segmenting spider legs, insect antennae, and the body of a turtle. | 2501.19329v1 | 4 |
|
Figure 5:Sketching the “wrong” region. The experiment shows that our network has successfully used the sketch information as the key reference. | 2501.19329v1 | 5 |
|
Figure 6:The analysis ofEmsubscript𝐸𝑚E_{m}italic_E start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPTmetrics on the COD10K dataset aboutθ𝜃\thetaitalic_θin Adaptive Focal Loss. | 2501.19329v1 | 6 |
|
Figure 6:The analysis ofEmsubscript𝐸𝑚E_{m}italic_E start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPTmetrics on the COD10K dataset aboutθ𝜃\thetaitalic_θin Adaptive Focal Loss. | 2501.19329v1 | 7 |
|
Figure 7:Sensitivity Anallysis ofEmsubscript𝐸𝑚E_{m}italic_E start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPTon the COD10K dataset. | 2501.19329v1 | 8 |
|
Figure 3:Sketching the “wrong” region. The experiment shows that our network has successfully used the sketch information as
the key reference information. | 2501.19329v1 | 11 |
|
Figure 4:Sensitivity Anallysis ofEmsubscript𝐸𝑚E_{m}italic_E start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPTon the CHAMELEON, CAMO, COD10K datasets. | 2501.19329v1 | 12 |
|
Figure 5:The analysis ofEmsubscript𝐸𝑚E_{m}italic_E start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT,Fβwsuperscriptsubscript𝐹𝛽𝑤F_{\beta}^{w}italic_F start_POSTSUBSCRIPT italic_β end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_w end_POSTSUPERSCRIPT, MAE metrics on the COD10k dataset. | 2501.19329v1 | 13 |
|
Figure 7:Visualization of the implementation process of Boundary Refinement. (a) original image; (b) GT; (c) predicted segment (pred); (d) boundary of GT; (e) boundary of pred; (f) expanded boundary of GT; (g) expanded boundary of pred; (h) pixel-wise multiplication of masks (d) and (g); (i) pixel-wise multiplication of masks (e) and (f). | 2501.19329v1 | 15 |
|
Figure 8:Different types of annotation camouflaged object detection task. The mask annotation takes about 60 minutes for each image. The point takes just 2 seconds, the box takes about 3 seconds, the scribble takes 5 seconds and sketch is drawn by different people only takes between 10-30s | 2501.19329v1 | 16 |
|
Figure 1:Comparison of six canopy height maps with precise measurements obtained via aerial laser scanning (ALS). The patches contain tall trees exceeding 30 m in height. Our model is the only one that can accurately estimate the height of such trees. | 2501.19328v1 | 1 |
|
Figure 3:Boxplots for each model showing the 2020 mean error in every5m5m5\,\mathrm{m}5 roman_mbin between10m10m10\,\mathrm{m}10 roman_mand40m40m40\,\mathrm{m}40 roman_m. AlthoughLiu et al. (2023),Pauls et al. (2024)andLang et al. (2023)perform well on smaller trees, our models performs especially well for taller trees. | 2501.19328v1 | 3 |
|
Figure 4:Qualitative comparison of canopy height maps for the reference year 2020:Liu et al.,Tolan et al.,Lang et al.,Pauls et al.,Turubanova et al.(2020) and our model,3D-Stack-MultiYear-L. | 2501.19328v1 | 4 |
|
Figure 5:Left: Scatterplots between 2020 GEDI labels and prediction forLang et al.;Liu et al.;Pauls et al.;Tolan et al.;Turubanova et al.and our model includingR2superscript𝑅2R^{2}italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTfor all labels andR72subscriptsuperscript𝑅27R^{2}_{7}italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 7 end_POSTSUBSCRIPTfor labels exceeding7m7m7\,\mathrm{m}7 roman_m. Right: Histograms of GEDI labels and all maps.Turubanova et al.andTolan et al.saturate at28m28m28\,\mathrm{m}28 roman_m, our model is the only one matching above40m40m40\,\mathrm{m}40 roman_m. | 2501.19328v1 | 5 |
|
Figure 6:Temporal maps illustrate the expansion of deforestation from 2019 to 2022. This is observed by comparing differences between each year, visible in both solitary forest patches surrounded by open land and within densely forested areas. | 2501.19328v1 | 6 |
|
Figure 4:Sub-model architecture with input sizeP×2P×C𝑃2𝑃𝐶P\times 2P\times Citalic_P × 2 italic_P × italic_C, where the piece size isP×P𝑃𝑃P\times Pitalic_P × italic_Ppixels andC𝐶Citalic_Cis the number of channels; the architecture includes four convolutional layers with3×3333\times 33 × 3kernels and ReLU activation function; max pooling is applied after the second and third layers, and dropout with a probability of 0.25 is applied after all layers except the first; the final convolutional layer is flattened and passed through a fully-connected layer to compute the compatibility score without an activation function; no biases are used in any layer. | 2501.19325v1 | 12 |
|
Figure 5:Our DLCM architecture, consisting of four sub-models: RGB-Net, Red-Net, Green-Net, and Blue-Net, with the same architecture as Figure4; the DLCM output is the sum of the outputs of all four sub-models. | 2501.19325v1 | 13 |
|
Figure 6:Illustration of tile placement through chromosome pairing in different phases of the proposed genetic algorithm (GA); starting with a single piece, the kernel expands by sequentially placing additional tiles adjacent to its free (piece) edges. In the depicted toy scenario, the kernel consists of four (out of nine possible pieces) while specifically examining the right boundary of tile ’b’.Phase 1.1/1.2:For the left parent chromosome, the average CM scores of pieces ’b’ and ’g’ exceed the chromosome’s overall fitness;Phase 2:Both parent chromosomes share the adjacency of piece ’g’ to the right of piece ’b’;Phase 3:Pieces ’b’ and ’g’ are best buddies in the left parent;Phases 4.1, 4.2,andPhase 5:see details in the figure itself. The new kernel (and resulting offspring chromosome), due to any of the above phases, includes piece ’g’ to the right of piece ’b’. | 2501.19325v1 | 14 |
|
Figure 7:Top-i𝑖iitalic_iaccuracy plots comparing traditional and DL-based CMs with the proposed DLCM for Type-1 (top) and Type-2 (bottom) Portuguese tile panels; DLCM achieves superior performance over all tested methods, including SqueezeNet- and GAN-based CMs. | 2501.19325v1 | 15 |
|
Figure 8:Evolutionary reconstruction of three Portuguese tile panels using the enhanced GA-based solver; each sub-image shows the intermediate solution along with the corresponding ”heatmap” for a specific generation, highlighting the evolutionary nature of the GA; heatmap cells represent fitness (i.e., average CMs) of each tile relative to its local neighborhood; the brighter a cell, the higher its local compatibility. | 2501.19325v1 | 16 |
|
Figure 11:Compatibility measure for strip-cut pair. Left to right: Two strips, adjusted as a strip-cut pair, are divided inton𝑛nitalic_nchunks (with any residue disposed of in subsequent phases); GrayNet assigns a compatibility scoreSisubscript𝑆𝑖S_{i}italic_S start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTto each chunkXisubscript𝑋𝑖X_{i}italic_X start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT, and the final strip-cut compatibility is determined by the sum of the scores. | 2501.19325v1 | 25 |
|
Figure 12:Compatibility map of puzzle tiles: (Top) 256-piece Portuguese tile panel from the MNAz test set, and (bottom) compatibility score matrix; thei𝑖iitalic_i-th row represents anchor piecei𝑖iitalic_i, and thej𝑗jitalic_j-th column represents examined piecej𝑗jitalic_jwith respect to the right edge of the anchor; unless the anchor is positioned at the rightmost puzzle, its correct neighboring piece on the right is piece(i+1)𝑖1(i+1)( italic_i + 1 ),i.e., super diagonal entries of the CM matrix should be the highest (in yellow), while other entries (in dark purple) should be lower. | 2501.19325v1 | 26 |
|
Figure 1:Reward-Guided Speculative Decoding (RSD).This diagram illustrates how RSD improves upon standard speculative decoding (SD) by incorporating reward-guided selection. SD strictly enforces exact token matching between the draft and target model, leading to unnecessary computations when mismatched tokens are discarded. In contrast, RSD evaluates draft outputs based on reward signalsr𝑟ritalic_rand selectively refines them, reducing reliance on exact matching and improving efficiency. The process starts with a small and fast draft model generating preliminary results, followed by a larger and more reliable target model verifying and refining predictions. Darker background regions indicate higher computational costs, showing how SD wastes resources on rejected tokens, whereas RSD reduces unnecessary steps by accepting useful draft outputs even when they do not exactly match, balancing efficiency and accuracy. | 2501.19324v1 | 1 |
|
Figure 3:Left: A comparison of the reward scores for all questions generated by the draft model and the target model within the RSD framework.Middle: A focused comparison of the reward scores for correctly answered questions generated by the draft model and the target model in the RSD framework.Right: The winning rate (in terms of reward) comparison between the draft model and the target model, highlighting the proportion of cases where each model outperforms the other. RSD is configured with Qwen2.5-Math-1.5B-Instruct as the draft model, Qwen2.5-Math-7B-Instruct as the target model, and Skywork-o1-Open-PRM-7B as the PRM. | 2501.19324v1 | 3 |
|
Figure 6:Accuracy of weighting functions from Table1with RSD (1.5B/7B/7B). All settings share a similar inference cost. | 2501.19324v1 | 6 |
|
Figure B.1:The behaviour of RSD (1.5B/7B/7B) for differentδ𝛿\deltaitalic_δs and questions in different complexity levels (the higher the level, the harder the question.).δ=0𝛿0\delta=0italic_δ = 0andδ=1𝛿1\delta=1italic_δ = 1denotes all questions are solved by the draft model alone and the target model only, respectively. Overall, the involvement of the target model improves the accuracy. The improvement is more obvious for harder question, +16 for level 4 and 5. In addition, with an increasing level, the questions solved by the draft model only decrease for the sameδ𝛿\deltaitalic_δ, demonstrating harder questions need more involvement of the target model. | 2501.19324v1 | 7 |
|
Figure B.2:Left: A comparison of the reward scores for all questions generated by the draft model and the target model within the RSD framework.Middle: A focused comparison of the reward scores for correctly answered questions generated by the draft model and the target model in the RSD framework.Right: The winning rate comparison between the draft model and the target model, highlighting the proportion of cases where each model outperforms the other in the RSD framework. RSD is configured with Qwen2.5-Math-1.5B-Instruct as the draft model, Qwen2.5-Math-72B-Instruct as the target model, and Skywork-o1-Open-PRM-7B as the PRM. | 2501.19324v1 | 8 |
|
Figure 4:Correlation matrix between the features in the simulated dataset to perform binary classification. | 2502.00088v1 | 4 |
|
Figure 1:Training pipeline for all modelsXLS-R is fine-tuned to an upstream language. We then prune and train the downstream task for 10 epochs. If the upstream and downstream languages do not match we freeze the encoder and train for 1 extra epoch before unfreezing the encoder and training for 10 epochs | 2501.19321v1 | 1 |
|
Figure 2:Upstream English to multiple downstream Languagesan English upstream model is fine-tuned to downstream English, French, German, Polish and Spanish while pruning from 0% up to 90% sparsity | 2501.19321v1 | 2 |
|
Figure 3:Upstream Spanish to five downstream languagesthe Spanish upstream model is fine-tuned to downstream English, French, German, Polish and Spanish at 70%, 80% and 90% sparsities | 2501.19321v1 | 3 |
|
Figure 4:Upstream Polish to five downstream Languagesthe Polish upstream model is fine-tuned to downstream English, French, German, Polish and Spanish at 70%, 80% and 90% sparsities | 2501.19321v1 | 4 |
|
Figure 5:Mean average results for each language-specific subnetwork when tested on all other downstream languageseach subnetwork is fine-tuned to the four other languages, these results are then averaged and plotted at 70%, 80% and 90% sparsity | 2501.19321v1 | 5 |
|
Figure 6:Mean average results for each downstream languageeach language is trained from an upstream model fine-tuned to each of the four other languages, these results are then averaged and plotted at 70%, 80% and 90% sparsity. Results when the downstream and upstream languages match are not included in the averaging | 2501.19321v1 | 6 |
|
Figure 7:Intersection Over Union (IOU) at 90% sparsityIOU overlap of weights in subnetworks between English, Spanish and Polish subnetworks and the subnetwork found in the base model before fine-tuning. All subnetworks pruned to 90%. | 2501.19321v1 | 7 |
|
Figure 8:Intersection Over Union (IOU) at 90% sparsity for Spanish languages and EnglishOverlap of weights in subnetworks between English with Asturian and Catalan, then Spanish with Asturian and Catalan. Subnetworks pruned to 90%. | 2501.19321v1 | 8 |
|
Figure 1:Reconstruction and Rendering Results.Compared with the 3DGS-based SLAM, our method utilizes 2DGS for geometry-accurate scene representation, producing novel view rendering of high-quality images, view-consistent depth maps, and precise surface normal. | 2501.19319v1 | 1 |
|
Figure 2:Overview of Endo-2DTAM.Our proposed system consists of three modules: the tracking module, the mapping module, and the bundle adjustment. The tracking module takes the incoming RGBD frame as input and tracks the camera pose. Then the frame is added to the candidate list for the pose-consistent keyframe selection. In the mapping module, we first expand 2D Gaussians with the new frame and then update 2D Gaussians with the selected keyframes. The selected keyframes are also used for bundle adjustment for joint optimization of poses and 2D Gaussians. | 2501.19319v1 | 2 |
|
Figure 3:Qualitative Result on C3VD[56].We compare our method with the SOTA EndoGSLAM[38]for dense endoscopic SLAM. Our method generates more robust color and depth reconstruction as shown by results fromcecum_t2_b,sigmoid_t2_a. Our method also estimates a more precise trajectory demonstrated by results fromcecum_t3_a. | 2501.19319v1 | 3 |
|
Figure 4:Surface Normal Comparison of Mapping Ablation.We compare the rendered surface normal with different modalities as supervisions. Results demonstrate that the combination of all supervision with color, depth, and normal achieves the best quality. | 2501.19319v1 | 4 |
|
Figure 1:Overview of the MINDSTORES planning architecture. The left shows the iterative experiential learning pipeline leveraging the experience database. Database-related methods are in orange, planning steps are in green, and Minecraft steps are in red. The right shows an example applies this pipeline to an example task in Minecraft. | 2501.19318v1 | 1 |
|
Figure 2:Interactive planning process for crafting iron boots in Minecraft. The system initially plans to mine iron with a wooden pickaxe but learns from past experience that this will fail. It then updates the plan to include creating a stone pickaxe first, leading to successful iron ore mining. | 2501.19318v1 | 2 |
|
Figure 1:Probing workflow with Coreference Resolution (Coref) as target task and four different source tasks: Relation Extraction (RE), Question Answering (QA), Named Entity Recognition (NER), and Paraphrase Detection (MRPC). | 2501.19316v1 | 1 |
|
Figure 3:Average cosine similarity between the embeddings of the source tasks and the target coreference task, averaged across all tokens for 15 batches | 2501.19316v1 | 3 |
|
(a)Pairs of models: comparison of two embedding aggregation methods, mean and attention, to combine the source task model outputs | 2501.19316v1 | 4 |
|
(a)Pairs of models: comparison of two embedding aggregation methods, mean and attention, to combine the source task model outputs | 2501.19316v1 | 5 |
|
(b)Single models: performance gains by adding attention projections (attention) compared to having no additional parameters (mean) | 2501.19316v1 | 6 |
|
Figure 5:Source task model performance truncated to the best layer (in parentheses) with mean aggregation | 2501.19316v1 | 7 |
Subsets and Splits