image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 8:HTSR metrics predict forecasting accuracy across architectures (FPT vs non-FPT) and within architecture across epochs.P50 quantile test loss by epoch for architectures with a Linear and MLP embedding/decoding layer, respectively. Markers and lines are colored according to theα𝛼\alphaitalic_αmetric (left) and the stable rank metric (right). | 2501.06386v1 | 36 |
|
Figure 9:Representation of univariate patching[36]for time series (left) and multivariate patching for time series (right). In the multivariate setting, the time series is first patched, and then flattened across the patch and covariates prior to embedding. | 2501.06386v1 | 39 |
|
Figure 1:Illustration of the difference between human cognition and AI contextual completion. The original fragment of “One hundred years of solitude"(García Márquez,,1967)(top) has a clear “spontaneous" thought, but the GPT-2’s completion222Just to illustrate, we use the promptPlease continue this short sentence, forgetting about “One hundred Years of Solitude”, since on a real
conversation the LLM would be blind to the final output.(bottom), demonstrates continuity. | 2501.06382v2 | 1 |
|
Figure 2:The feasible set withr⋆superscript𝑟⋆r^{\star}italic_r start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPTandr^,r^M,K,r^m,K^𝑟subscript^𝑟𝑀𝐾subscript^𝑟𝑚𝐾\widehat{r},\widehat{r}_{M,K},\widehat{r}_{m,K}over^ start_ARG italic_r end_ARG , over^ start_ARG italic_r end_ARG start_POSTSUBSCRIPT italic_M , italic_K end_POSTSUBSCRIPT , over^ start_ARG italic_r end_ARG start_POSTSUBSCRIPT italic_m , italic_K end_POSTSUBSCRIPT. AxisB,S,T𝐵𝑆𝑇B,S,Titalic_B , italic_S , italic_Trefer to
the reward values of B, S and T. | 2501.06376v1 | 2 |
|
Figure 3:(Left) The target environment considered in the
experiment.
(Right) Representation ofr⋆superscript𝑟⋆r^{\star}italic_r start_POSTSUPERSCRIPT ⋆ end_POSTSUPERSCRIPTfor the target environment. | 2501.06376v1 | 3 |
|
Figure 4:(Left) Plot ofdp,π1superscript𝑑𝑝superscript𝜋1d^{p,\pi^{1}}italic_d start_POSTSUPERSCRIPT italic_p , italic_π start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT.
(Right) Plot ofdp,π2superscript𝑑𝑝superscript𝜋2d^{p,\pi^{2}}italic_d start_POSTSUPERSCRIPT italic_p , italic_π start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT. | 2501.06376v1 | 4 |
|
Figure 5:The trajectories compared in the first feedback.ω11subscriptsuperscript𝜔11\omega^{1}_{1}italic_ω start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTis on the left, andω12subscriptsuperscript𝜔21\omega^{2}_{1}italic_ω start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTon the right. | 2501.06376v1 | 5 |
|
Figure 6:The trajectories compared in the second feedback.ω21subscriptsuperscript𝜔12\omega^{1}_{2}italic_ω start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTis on the left, andω22subscriptsuperscript𝜔22\omega^{2}_{2}italic_ω start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTon the right. | 2501.06376v1 | 6 |
|
Figure 7:The trajectories compared in the third feedback.ω31subscriptsuperscript𝜔13\omega^{1}_{3}italic_ω start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPTis on the left, andω32subscriptsuperscript𝜔23\omega^{2}_{3}italic_ω start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPTon the right. | 2501.06376v1 | 7 |
|
Figure 8:(Left) The new map considered for the comparisons
feedback. (Right) The new map considered for the demonstrations feedback. | 2501.06376v1 | 8 |
|
Figure 9:The occupancy measures compared in the first comparisons
feedback.π11subscriptsuperscript𝜋11\pi^{1}_{1}italic_π start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTis on the left, andπ12subscriptsuperscript𝜋21\pi^{2}_{1}italic_π start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTon the right. | 2501.06376v1 | 9 |
|
Figure 10:The occupancy measures compared in the second comparisons
feedback.π21subscriptsuperscript𝜋12\pi^{1}_{2}italic_π start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTis on the left, andπ22subscriptsuperscript𝜋22\pi^{2}_{2}italic_π start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPTon the right. | 2501.06376v1 | 10 |
|
Figure 12:(Left) The sequence of rewardsr^m,ksubscript^𝑟𝑚𝑘\widehat{r}_{m,k}over^ start_ARG italic_r end_ARG start_POSTSUBSCRIPT italic_m , italic_k end_POSTSUBSCRIPTcomputed by our algorithm. (Right) The corresponding values of the objective
function. | 2501.06376v1 | 12 |
|
Figure 13:(Left) The sequence of rewardsr^M,ksubscript^𝑟𝑀𝑘\widehat{r}_{M,k}over^ start_ARG italic_r end_ARG start_POSTSUBSCRIPT italic_M , italic_k end_POSTSUBSCRIPTcomputed by our algorithm. (Right) The corresponding values of the objective
function. | 2501.06376v1 | 13 |
|
Figure 14:The rewardsrM,rm,rℱ,gsubscript𝑟𝑀subscript𝑟𝑚subscript𝑟ℱ𝑔r_{M},r_{m},r_{\mathcal{F},g}italic_r start_POSTSUBSCRIPT italic_M end_POSTSUBSCRIPT , italic_r start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT , italic_r start_POSTSUBSCRIPT caligraphic_F , italic_g end_POSTSUBSCRIPTcomputed through a
discretization of the feasible set. | 2501.06376v1 | 14 |
|
Figure 4:Difference in d-CHRF for NLLB-200 (1.3B) before and after finetuning onAfriDoc-MTfor our two domains. | 2501.06374v1 | 10 |
|
Figure 5:Distribution of the quality estimation of of the translated sentences using COMET scores for thehealth(top),tech(bottom). | 2501.06374v1 | 13 |
|
Figure 14:Change in s-BLEU and s-CHRF for sentence evaluation comparing NLLB1.3B before and after supervised finetuning onAfriDoc-MT | 2501.06374v1 | 42 |
|
Figure 15:Change in d-BLEU and d-CHRF for sentence evaluation comparing NLLB1.3B before and after supervised finetuning onAfriDoc-MT | 2501.06374v1 | 47 |
|
Figure 16:d-CHRF scores for some LLMs for sentence-level translation using different prompts when translating into African languages | 2501.06374v1 | 52 |
|
Figure 17:d-CHRF scores for some LLMs for sentence-level translation using different prompts when translating into African languages | 2501.06374v1 | 53 |
|
Figure 18:d-CHRF scores for some LLMs for sentence-level translation using different prompts when translating into English | 2501.06374v1 | 54 |
|
Figure 19:d-CHRF scores for some LLMs for sentence-level translation using different prompts when translating into English | 2501.06374v1 | 55 |
|
Figure 20:dCHRF scores for some LLMs for sentence-level translation using different prompts when translating into African languages | 2501.06374v1 | 56 |
|
Figure 21:d-CHRF scores for some LLMs for sentence-level translation using different prompts when translating into African languages | 2501.06374v1 | 57 |
|
Figure 22:d-CHRF scores for some LLMs for sentence-level translation using different prompts when translating into English | 2501.06374v1 | 58 |
|
Figure 23:d-CHRF scores for some LLMs for sentence-level translation using different prompts when translating into English | 2501.06374v1 | 59 |
|
Figure 2:Visualization of results on three 2D datasets: (a), (c), and (e) represent the three synthetic datasets, while (b), (d), and (f) show the affinity graphs𝐙𝐙\mathbf{Z}bold_Zlearned by the model. For clarity, only edges with weights of 0.001 or higher are displayed. | 2501.06368v2 | 2 |
|
Figure 3:Binarized affinity matrices on high-dimensional SyD4, comparing various SOTA methods with ours. | 2501.06368v2 | 3 |
|
Figure 5:Clustering results for four sequences (people1, cars10, 1R2TCR, 2T3RTCR) from the Hopkins155 database. The top row shows sample images with overlaid tracked points, the second row displays self-representation matrices generated by DKLM, and the bottom row provides 3D visualizations of the clustering outcomes. | 2501.06368v2 | 5 |
|
Figure 8:Result of time series. (a) Discovered distinct regimes for the three time series datasets. (b) True (black) and forecasted (blue) values for the three series, each from a real dataset. (c) The best window size (red line) for the three time series datasets. | 2501.06368v2 | 8 |
|
Figure 1:System overview of the proposed multi-step MPC framework for DED, whereJ𝐽Jitalic_Jis the cost function,w.r.tformulae-sequence𝑤𝑟𝑡w.r.titalic_w . italic_r . italic_tdenotes "with respect to",s.t.formulae-sequence𝑠𝑡s.t.italic_s . italic_t .denotes "subject to", andgisubscript𝑔𝑖g_{i}italic_g start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTrepresents thei𝑖iitalic_ith constraint. | 2501.07601v2 | 1 |
|
Figure 10:Evaluation of TiDE model and training loss: (a) and (c) show the melt pool temperature comparison between the GAMMA simulation result as the ground truth and the predictions of uni-variate and multi-variate TiDE model, and (b) and (d) shows the melt pool depth comparison between GAMMA simulation and multi-variate TiDE model. The TiDE prediction is made in one-shot through out the horizonp=50𝑝50p=50italic_p = 50. (e) and (f) are the loss for training uni-variate and multi-variate TiDE model | 2501.07601v2 | 11 |
|
Figure 11:Simulation result from GAMMA using MPC and PID as controller. (a) Melt pool temperature reference, GAMMA simulation result using PID controller and MPC, with the TiDE prediction used in MPC. (b) Comparison of the controlled laser power using PID and MPC, respectively. (c) Highlight of the beginning of the printing where MPC starts activating. (d) Temperature trajectory at the corner. (e) Temperature trajectory at layer transition. | 2501.07601v2 | 12 |
|
Figure 12:Comparison between the enforcement of constraints on melt pool depth using GAMMA simulation. (a) Melt pool temperature comparison between the GAMMA result of unconstrained and constrained MPC, with the TiDE prediction from constrained MPC. (b) Melt pool depth comparison between the GAMMA result of unconstrained and constrained MPC. (c) Applied laser power comparison for constrained and unconstrained MPC. (d) Melt pool temperature and depth trajectory at the sixth layer. (e) Melt pool temperature and depth trajectory at the transition from the seventh to eighth layer. | 2501.07601v2 | 13 |
|
Figure 13:Histogram of solving time for constrained and unconstrained MPC. The highlighted window shows the counts of instances with solving time greater than 1 second. | 2501.07601v2 | 14 |
|
Figure 4:Comparisons under the linear setting (top row) and non-linear setting (bottom role): a,d) CF metric versus sample size, b,e) cumulative discounted reward versus CF metric (multiple sample sizes) withδ=1𝛿1\delta=1italic_δ = 1, c,f) CF metric versusδ𝛿\deltaitalic_δ. All the results are aggregated over 100 random seeds. The shaded area represents the95%percent9595\%95 %CI. | 2501.06366v2 | 4 |
|
Figure 5:Semi-synthetic experiments: CF metric of different approaches versusη𝜂\etaitalic_ηfor different sensitive attributes: education, sex, ethnicity. All the results are aggregated over 100 random seeds. The shaded area is the95%percent9595\%95 %CI. | 2501.06366v2 | 5 |
|
Figure 2:Overview of the Mix-QViT framework.
(a) Layer Importance Score (ΩΩ\Omegaroman_Ω), calculated offline on a small sample of 256 images from the ImageNet1K validation dataset, based on the Layer-wise Relevance Propagation (LRP) method. (b) Quantization sensitivity score (ΛΛ\Lambdaroman_Λ), calculated offline on 256 images from the ImageNet1K validation dataset. Here, performance changes are recorded between two models: one where all layers are quantized at baseline precision, and another where the target layer in each transformer block is quantized at different precision. (c) Mixed-precision bit allocation strategy, based on the layer importance score and quantization sensitivity score, which, under model constraints, generates an optimal mixed-bit allocation. | 2501.06357v1 | 4 |
|
Figure 3:Layer Importance Score of DeiT-Small. Each value refers to the layer relative importance toward model classification. | 2501.06357v1 | 5 |
|
Figure 6:Model size vs. accuracy trade-off curves for DeiT-S under various fixed- and mixed-precision frameworks. | 2501.06357v1 | 14 |
|
Fig. 1:Without a pixel-level lesion segmentation mask, the currentmask-and-pastegeneration paradigm[6,7]produces noticeable boundary artifacts, creating a clear distinction between the synthetic lesion and its background (middle row). | 2501.06356v1 | 1 |
|
Fig. 3:a.) The visualization ofP(ΔX,ΔY∣X=x,Y=y)P(\Delta X,\Delta Y\mid X=x,Y=y)italic_P ( roman_Δ italic_X , roman_Δ italic_Y ∣ italic_X = italic_x , italic_Y = italic_y ), where the blue dot (x,y𝑥𝑦x,yitalic_x , italic_y) denotes position of the anatomical structure, and b.) the pipeline for creating the lesion ”skeleton” which is used as condition to guide the generative model. | 2501.06356v1 | 3 |
|
Figure 2:Proportion of features shared across languages (intersection over union) among the top 32 features for each morphosyntactic concept.
A significant fraction of the morphosyntactic concept representations are shared across languages in both Llama-3-8B and Aya-23-8B. | 2501.06346v1 | 2 |
|
Figure 3:Examples of the activation patterns of selected features that correspond to cross-lingual representations of grammatical concepts. For example, we locate features that indicate the presence of plural nouns across languages or features that indicate past tense. | 2501.06346v1 | 3 |
|
Figure 4:Performance of the probing classifiers before and after ablating multilingual features. Specifically, we test ablating (a) all multilingual features, and (b) the upper quartile of themost multilingualfeatures. We find that the classifiers crucialy rely on multilingual features to predict the presence of morphosyntactic features. | 2501.06346v1 | 4 |
|
Figure 5:Efficacy in flipping the model behavior on our counterfactual dataset when translating between languages and intervening on asinglemultilingual feature per concept.
For each concept, we translate a sentence from some source language (e.g. German) where some concept (e.g. present tense) to another language (e.g. English) and measure the number of times the model starts to predict some alternative concept value (e.g. past tense).
In each of these settings, we intervene on a single feature and measure the success rate over 64 examples. | 2501.06346v1 | 5 |
|
Figure 6:Fraction of times an intervention on a grammatical concept exclusively impacts that concept, indicating the degree of cross-concept interference. We observe that interference is generally very rare. | 2501.06346v1 | 6 |
|
Figure 8:Distribution of the number of languages across which a given feature associated with masculine gender is shared (Llama 3). | 2501.06346v1 | 8 |
|
Figure 9:Distribution of the number of languages across which a given feature associated with past tense is shared (Llama 3). | 2501.06346v1 | 9 |
|
Figure 10:Distribution of the number of languages across which a given feature associated with accusative case is shared (Llama 3). | 2501.06346v1 | 10 |
|
Figure 11:Distribution of the number of languages across which a given feature associated with masculine gender is shared (Aya-23). | 2501.06346v1 | 11 |
|
Figure 12:Distribution of the number of languages across which a given feature associated with past tense is shared (Aya-23). | 2501.06346v1 | 12 |
|
Figure 13:Distribution of the number of languages across which a given feature associated with accusative case is shared (Aya-23). | 2501.06346v1 | 13 |
|
Figure 2:Existing metrics.A comparison between MEt3R and TSED[46]scores obtained from individual image pairs generated by GenWarp[31]. TSED misses obvious, partial multi-view inconsistencies and is biased to small violations of epipolar geometry. In contrast, MEt3R correctly captures clear 3D inconsistencies and is robust to insignificant artifacts almost invisible to the human eye. | 2501.06336v1 | 2 |
|
Figure 4:Metric comparison.We compare MEt3R against TSED, SED, and FVD by computing average per-frame (/-segment for FVD) scores over a large number of generated sequences. MEt3R is able to capture nuanced differences in consistency of DFM, MV-LDM, and real videos, while TSED rates them all very similar. Unlike MEt3R, SED does not capture increasing inconsistency for PhotoNVS and DFM. MEt3R is able to capture the influence of anchor views in MV-LDM (c.f. Sec.4and appendix Sec.A) as structured high-frequency patterns. For MEt3R, the standard deviation gradually increases, starting from a small value, which is expected behavior due to conditioning on the first frame and is not the case for the other metrics. | 2501.06336v1 | 4 |
|
Figure 5:Qualitative comparison of generated novel views.We compare generated views of the multi-view generation method for the same conditioning view. We can extract certain characteristics: DFM is almost perfectly consistent but has lower image quality. PhotoNVS and MV-LDM are reasonably consistent on a structural scale but fail to produce consistent details. GenWarp fails to keep the structural consistency over the sequence while producing high-quality images. These observations are confirmed by MEt3R in Tab.1and Fig.4. | 2501.06336v1 | 5 |
|
Figure 6:Feature similarity ablation.We compare MEt3R against versions of it that compare RGB projections via PSNR and SSIM. It can be seen that the PSNR versions give better scores to DFM than to real videos. We attribute this to their sensitivity to view-dependent effects, such as lighting. In contrast, MEt3R rates the real video best. Further, the standard deviation of PSNR and SSIM versions are much higher, also for real videos, indicating a lower signal-to-noise ratio. | 2501.06336v1 | 8 |
|
Figure 7:MEt3R on generated videos.Per-image-pair plot for MEt3R across 48 frames and averaged across 100 sequences of RealEstate10K. For I2VGen-XL, we observe large inconsistencies initially as the inputs are out-of-distribution, followed by gradual improvement, indicating the inputs get closer to being in-distribution, also visible qualitatively in Figs.15-19. Meanwhile, Ruyi-Mini-7B shows several periodic spikes indicating abrupt inconsistencies throughout the video sequence, whereas MEt3R for SVD stays relatively low and smooth. | 2501.06336v1 | 9 |
|
Figure 8:Feature backbone ablation.We analyze the effect of different feature backbones on MEt3R. While DINOv2[26]and MaskCLIP[48]can be employed as well, we found DINO features to lead to a more informative separation of models. | 2501.06336v1 | 10 |
|
Figure 9:MV-LDM.Architecture overview of MV-LDM, which consists of a shared 2D UNet initialized from Stable Diffusion 2.1[28]across multiple input views with cross-view attentions (3D attention) in between for modeling multi-view prior. | 2501.06336v1 | 11 |
|
Figure 10:MEt3R at different training iterations.As we continue to train MV-LDM, we see a consistent improvement in 3D consistency, which is an expected behavior. Furthermore, in the beginning, the improvements are large, which slows down and saturates in the later iterations. | 2501.06336v1 | 12 |
|
Figure 11:Anchored vs. autoregressive.Per-image-pair MEt3R on 2 different sampling strategies. For autoregressive sampling, we see significant and periodic spikes becoming larger as we progress and show the effect of compounding error, i.e., sequentially generating new frames and anchors conditioned on the previously generated ones. As illustrated in Fig.12, autoregressive sampling produces several anchor-to-anchor transitions causing these periodic spikes. On the other hand, anchored generation limits the effect of compounding error by generating all anchors in parallel. | 2501.06336v1 | 13 |
|
Figure 12:Anchored vs. Autoregressive sampling schemes.An illustration of the differences in the sampling schemes. In autoregressive sampling, we start from the initial input image and generate a set of target frames. The next set of frames is conditioned both on the input image and on the last frame (anchor) of the previously generated set. With this sampling strategy, we see several anchor-to-anchor transitions and results in large inconsistencies as visible in Fig.11. Whereas using anchored generation i.e., generate anchors first and then sample the remaining conditioned on the closest anchor and the input image. With this strategy, we observe significantly fewer anchor-to-anchor transitions, limited error accumulation, and relatively stable and lower MEt3R across the image pairs. | 2501.06336v1 | 14 |
|
Figure 13:Fixed vs. adjusted projection matrix.With fixed focal length, the projection area varies across different scales of DUSt3R point maps. We automatically adjust the focal length for each example pair to allow maximal projection and, therefore, more pixels for evaluating feature similarity. | 2501.06336v1 | 15 |
|
Figure 14:MEt3R without overlap mask.Per-image-pair MEt3R without normalizing against the overlap mask. Under this setting, DFM[36]is worse than all other baselines in 3D consistency, even though it has a strong inductive bias, which forces its results to be 3D consistent at the expense of blur. Whereas PhotoNVS[46]and GenWarp[31]are similar, both of which increase gradually, whereas MV-LDM stays relatively low with visible spikes due to anchor-to-anchor transitions. | 2501.06336v1 | 16 |
|
Figure 15:Examples of generated multi-views and videos.From Top→→\rightarrow→Down is the increasing frame number with columns for each method. Note that the first row is the input image, the first four columns are the results of multi-view generation models with explicit camera control, whereas the last three columns are generated videos from video diffusion models without any camera control. | 2501.06336v1 | 17 |
|
Figure 16:Examples of generated multi-views and videos.From Top→→\rightarrow→Down is the increasing frame number with columns for each method. Note that the first row is the input image, the first four columns are the results of multi-view generation models with explicit camera control, whereas the last three columns are generated videos from video diffusion models without any camera control. | 2501.06336v1 | 18 |
|
Figure 17:Examples of generated multi-views and videos.From Top→→\rightarrow→Down is the increasing frame number with columns for each method. Note that the first row is the input image, the first four columns are the results of multi-view generation models with explicit camera control, whereas the last three columns are generated videos from video diffusion models without any camera control. | 2501.06336v1 | 19 |
|
Figure 18:Examples of generated multi-views and videos.From Top→→\rightarrow→Down is the increasing frame number with columns for each method. Note that the first row is the input image, the first four columns are the results of multi-view generation models with explicit camera control, whereas the last three columns are generated videos from video diffusion models without any camera control. | 2501.06336v1 | 20 |
|
Figure 19:Examples of generated multi-views and videos.From Top→→\rightarrow→Down is the increasing frame number with columns for each method. Note that the first row is the input image, the first four columns are the results of multi-view generation models with explicit camera control, whereas the last three columns are generated videos from video diffusion models without any camera control. | 2501.06336v1 | 21 |
|
Figure 2:Federated Learning architecture with a central orchestrator
and 3 clients with private datasets. | 2501.06332v1 | 2 |
|
Figure 3:SST2 Dataset (i.i.d. split with balanced classes): Evaluation set accuracy on each client per iteration over a rolling average of 7. | 2501.06332v1 | 3 |
|
Figure 4:MNLI Dataset (i.i.d. split with balanced classes): Evaluation set accuracy on each client per iteration. | 2501.06332v1 | 4 |
|
Figure 5:SST2 Dataset (split with imbalanced classes): Evaluation set accuracy on each client per iteration over a rolling average of 7. | 2501.06332v1 | 5 |
|
Figure 6:MNLI Dataset (split with imbalanced classes): Evaluation set accuracy on each client per iteration. | 2501.06332v1 | 6 |
|
Figure 7:ErrorserrFRA-LoRA𝑒𝑟subscript𝑟FRA-LoRAerr_{\texttt{FRA-LoRA}}italic_e italic_r italic_r start_POSTSUBSCRIPT FRA-LoRA end_POSTSUBSCRIPTanderrFedAvg𝑒𝑟subscript𝑟FedAvgerr_{\texttt{FedAvg}}italic_e italic_r italic_r start_POSTSUBSCRIPT FedAvg end_POSTSUBSCRIPTfor the first 10 iterations of the MNLI imbalanced example. FRA-LoRA errors are an order of magnitude lower in absolute terms. | 2501.06332v1 | 7 |
|
Figure 2:This figure illustrates how CTC loss was integrated to the pipeline with the hope of learning positional alignment of characters and brain data. For each time step the log probabilities of each character in the vocabulary were calculated and then the CTC loss was calculated between the predicted and actual sentence. | 2501.06326v2 | 2 |
|
Figure 3:This figure shows the architecture of a Conformer Neural Network as proposed by Gulati et Al.Gulati et al.(2020)and how Convolution Layer can be integrrated with a Multi Headed Attention Layer. | 2501.06326v2 | 3 |
|
Figure 4:This Figure shows the proposed architecture for a Wave2Vec2 model training regime as proposed byBaevski et al.(2020a) | 2501.06326v2 | 4 |
|
Figure 1.Example question-answering application of LLM-based multi-agent collaborative system. In the first collaboration channel, two LLMs are collaborating through a debate against each other, given the input by the user with a turn-based strategy. In the second channel, the Oppose Agent cooperates and leverages information from Research Agents, and provides the final response to the user. | 2501.06322v1 | 1 |
|
Figure 2.Our proposed framework for LLM-based multi-agent collaborative system. Each agent consists of a language modelm𝑚mitalic_mas the neural processor, current objectiveo𝑜oitalic_o, environmente𝑒eitalic_e, input perceptionx𝑥xitalic_xand corresponding output/actiony𝑦yitalic_y. The framework’s central focus is on collaboration channels𝒞𝒞\mathcal{C}caligraphic_Cthat facilitate coordination and orchestration among agents. These channels are defined by their actors (agents involved), type, structure, and strategy. Our framework is flexible, accommodating previous approaches and enabling the analysis of diverse MASs under a unified structure. | 2501.06322v1 | 2 |
|
Figure 3.Illustrative examples of collaboration types, where each agenta𝑎aitalic_ais equipped with different tools or capabilities through their system promptr𝑟ritalic_r. In scenario a), agents cooperate by leveraging their individual specialties (e.g., writing, translation, research) to achieve a shared goal (academic writing). In scenario b), agents compete and debate against each other fo
r their own goals. In scenario (c) of coopetition, agents compromise with each other, compete on one aspect while agreeing on another. | 2501.06322v1 | 3 |
|
Figure 4.Different types of collaboration strategies, illustrated by multiple use cases. In the rule-based example, agents debate and participate in majority voting. Software project is an instance of role-based protocol. In games, agents communicate and perform probabilistic decision-making in uncertain environments. | 2501.06322v1 | 4 |
|
Figure 5.Summary of communication structures of MAS. Figure (a) illustrates the centralized structure, which can be categorized into two types. In the first type, the LLM resides on distributed agents, with FedIT serving as an example. In the second type, the LLM is hosted on a central agent, as exemplified by AutoAct. Figure (b) depicts the decentralized structure, with AgentCF as an example. Finally, Figure (c) represents the hierarchical structure, with the CAMEL architecture. | 2501.06322v1 | 5 |
|
Figure 6.LLM-based MAS-enabled semantic communication system framework, leveraging LLMs directly to the physical layer coding and decoding of communication system(Wang et al.,2024c). | 2501.06322v1 | 6 |
|
Figure 2:Representative examples of statistical and conceptual figures, with their sources detailed in the Appendix. | 2501.06317v1 | 2 |
|
Figure 3:User interface for the figure caption writing task, showing: (1) ‘Target Paper’ - the hyperlink to the redacted PDF, (2) ‘Target Figure’ - displaying the figure image for the writing task, with the figure number and the page number in the redacted PDF, (3) ‘Your Captions’ - User input area for caption writing, and (4) ‘Suggested Captions’ - AI-generated captions from 3 different configurations (Unlimited,Text-Only,30-Word), presented in randomized order for each caption item. | 2501.06317v1 | 3 |
|
Figure 4:Average number of interactions with AI-generated captions per writing task. Participants interacted with AI-generated captions an average of 2.13 times per session for statistical figures (SD=1.32) and 2.00 times for conceptual figures (SD=1.65). | 2501.06317v1 | 4 |
|
Figure 5:Distribution of the four AI integration activity types for statistical and conceptual figures. | 2501.06317v1 | 5 |
|
Figure 6:Distribution of the number of AI-generated captions used in a single writing task during the figure caption writing process in our study. | 2501.06317v1 | 6 |
|
Figure 7:Distribution of AI text usage degrees—Full,Sentence,Phrase, andTerm—in participants’ initial interactions with AI-generated captions. One session (out of 34) was excluded from this analysis as the participant did not use AI-generated captions. | 2501.06317v1 | 7 |
|
Figure 8:Distribution of GPT-4o configurations ranked by participants, with Rank 1 indicating the best and Rank 3 the worst (N=34). | 2501.06317v1 | 8 |
|
Figure 9:GPT-4o generated captions in three different configurations: (i) GPT-4o (image+text) with a 30-word limit, (ii) GPT-4o (text-only) with unlimited length, and (iii) GPT-4o (image+text) with unlimited length. The figure is sourced from Ramanan[38]. | 2501.06317v1 | 9 |
|
(b)Example ofAdapt- Incorporates AI suggestion’s core concept while modifying its linguistic expression | 2501.06317v1 | 12 |
|
Figure 1:Supporting Research through Knowledge Graphs and Directed Acyclic Graphs.A. A knowledge graph showcases the current state-of-the-art for a researcher to start from a research question and navigate through relevant data, tools, and methods to independently run their analysis. B. A typical bioinformatics workflow represented as a Directed Acyclic Graph (DAG), showcasing the intricate dependencies between tasks such as data preprocessing, genome assembly, annotation, and analysis, where each node represents a computational step and edges indicate the flow of data or control. | 2501.06314v1 | 1 |
|
Figure 3:Comparison of system and expert performance across conceptual genomics and code generation tasks.The top row evaluates conceptual genomics tasks, with separate panels for accuracy (left) and completeness (right). The bottom row evaluates code generation tasks, similarly split into accuracy (left) and completeness (right). For conceptual genomics tasks, the system demonstrates comparable performance to human experts across all levels of difficulty. In code generation tasks, the system matches expert performance on easier tasks, but shows a decline in accuracy and completeness for medium and hard tasks, highlighting opportunities for improvement in addressing complex challenges. | 2501.06314v1 | 6 |
|
(a) BioAgents on the Hard Workflow, broken down into two key parallel workflows, De novo Assembly and Variant calling. | 2501.06314v1 | 7 |
|
(a) BioAgents on the Hard Workflow, broken down into two key parallel workflows, De novo Assembly and Variant calling. | 2501.06314v1 | 8 |
|
(b) Experts on the Hard Workflow, broken down into three parallel workflows, De novo Assembly, Variant calling, and Sequence evolution. | 2501.06314v1 | 9 |
Subsets and Splits