image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 4:Dynamic self-correction with Many-Shot ICRL.The model’s actions are executed iteratively in the environment without any initial context or task identifier provided to the agent (i.e., cold start). Although it starts with a suboptimal policy, the model gradually improves through context-based self-correction. Results are aggregated over training tasks (ML45 split for Meta-World, ML20 split for Bi-DexHands and setpointsp∈[0,75]𝑝075p\in[0,75]italic_p ∈ [ 0 , 75 ]for Industrial-Benchmark) within each domain. | 2501.19400v1 | 4 |
|
Figure 5:Average domain-level demonstrator-normalized returns on training tasks: JAT vs. Vintix. JAT scores are aggregated over 100 trials per task, as originally reported by the authors, while for Vintix, we analyze 100 episodes after reaching inference-time convergence. | 2501.19400v1 | 5 |
|
Figure 6:Cold-start many-shot inference procedure on tasks with parameter variations compared to training tasks. In MuJoCo, variations include changes in viscosity (0.05 and 0.1 vs. the original 0) and gravity (±plus-or-minus\pm±10%). For the Industrial-Benchmark, we evaluate the model on previously unseen setpoint values,p∈[80,100]𝑝80100p\in[80,100]italic_p ∈ [ 80 , 100 ]. | 2501.19400v1 | 6 |
|
Figure 7:Inference-time performance on new tasks. One successful rollout and one failure case are reported for both Meta-World and Bi-DexHands. Inference is performed without an initial context in a task-agnostic manner. | 2501.19400v1 | 7 |
|
Figure 1:Comparison of Softmax and SSMax, illustrating the issue of attention fading and the effectiveness of SSMax in preventing it. As the input vector size increases, the maximum value of the output vector produced by Softmax decreases, demonstrating the problem of attention fading. In contrast, SSMax keeps the maximum value close to 1, regardless of the input size. The input vector consists of -2 for all elements except the last, which is set to +3. The scaling parameters𝑠sitalic_sof SSMax is set to 0.43. | 2501.19399v1 | 1 |
|
Figure 2:Relationship betweenpnsubscript𝑝𝑛p_{n}italic_p start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPTand the input vector sizen𝑛nitalic_n. The red dots represent the learned values ofpnsubscript𝑝𝑛p_{n}italic_p start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPTafter training, and the blue curve is a fitted logarithmic function of the formpn≈a1logn+a2subscript𝑝𝑛subscript𝑎1𝑛subscript𝑎2p_{n}\approx a_{1}\log n+a_{2}italic_p start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT ≈ italic_a start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT roman_log italic_n + italic_a start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT. This result suggests thatpnsubscript𝑝𝑛p_{n}italic_p start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPTdepends logarithmically onn𝑛nitalic_n, motivating the reformulation of Softmax inEquation4. | 2501.19399v1 | 2 |
|
Figure 4:Learning curves comparing the standard Transformer (a) and SSMax variants (b)–(d). All SSMax variants achieve consistently lower training loss compared to (a). Among them, the model with SSMax incorporating a bias parameter (d) exhibits the lowest loss throughout training. The results also indicate that removing the scaling parameter, as in (c), has little impact on the learning curve compared to (b). | 2501.19399v1 | 4 |
|
Figure 5:Per-position test loss across context sizes up to 20,000. The x-axis represents context size, and the y-axis represents test loss. RoPE’sθ𝜃\thetaitalic_θwas set to 50 times the training value, with no additional training after modification. The gray dotted line indicates the training sequence length of 1024. Results correspond to configurations (a)–(f). SSMax models (b) and (c) demonstrate improved long-context generalization compared to (a), while (d) exhibits degraded performance due to the bias parameter. Model (e), where Softmax was replaced with SSMax post-training, struggles with shorter contexts, whereas (f), which switched to SSMax during the final phase of pretraining, achieves performance somewhat close to (b), though not entirely equivalent. | 2501.19399v1 | 5 |
|
Figure 6:Needle-In-A-Haystack test results. The horizontal axis represents context size, while the vertical axis denotes the depth at which theneedleis embedded within the context. Colors indicate retrieval accuracy. RoPE’sθ𝜃\thetaitalic_θwas set to 500,000, a 50-fold increase from the pretraining value. The standard Transformer (a) fails to retrieve key information beyond short context sizes, while the SSMax model (b) maintains high retrieval accuracy even at context sizes approximately 10 times longer than in training. Models (c) and (d) show lower retrieval accuracy than (b), demonstrating that removing the scaling parameter or introducing a bias parameter degrades retrieval performance. Models where Softmax was replaced with SSMax after pretraining (e) and during pretraining (f) show partial improvements over (a) but remain far below (b). | 2501.19399v1 | 6 |
|
Figure 7:Needle score distribution across attention layers and heads. The horizontal axis represents attention heads ranked by needle score (highest to lowest), while the vertical axis shows the corresponding needle score. Note that only the top 25 heads are shown for clarity, rather than all 144 heads. RoPE’sθ𝜃\thetaitalic_θwas set to 500,000, a 50-fold increase from pretraining. The context size was 8000, with the needle sentence“The special magic Tokyo number is: 8106422.”inserted at a depth of 50%. The results demonstrate that the standard Transformer (a) fails to allocate significant attention to key tokens, whereas SSMax (b) effectively concentrates attention on them. Models (c), (d), (e), and (f) allocate more attention than (a) but fail to match the focus achieved by (b). Inference results indicate that (a) failed retrieval entirely, (b) and (c) successfully retrieved the correct number, and (d), (e), and (f) retrieved only the first digit but failed to recall the full number. | 2501.19399v1 | 7 |
|
Figure 8:Top needle score distribution across models. Each model was evaluated over 100 trials, and the highest needle score from each trial (corresponding to the leftmost value inFigure7) was recorded. The horizontal axis represents the rank of the top needle scores, sorted in descending order, while the vertical axis shows the corresponding score. Different markers indicate whether the retrieved number was fully correct (∙∙\bullet∙), incorrect but with the first digit correct (▲▲\blacktriangle▲), or completely incorrect (×\bm{\times}bold_×). RoPE’sθ𝜃\thetaitalic_θwas set to 500,000, a 50-fold increase from pretraining. Context size was fixed at 8000, with city names, numbers, and insertion depths randomly assigned. The results confirm that the standard Transformer (a) fails to focus attention on key tokens, whereas SSMax (b) exhibits strong concentration. Models (c), (d), (e), and (f) show partial improvements over (a) but fail to match (b)’s level of attention focus. | 2501.19399v1 | 8 |
|
Figure 1:A example gameplay for The Chameleon. In this example, the non-chameleons (blue players) correctly identify the chameleon (red player), but the chameleon wins the game in the second chance. | 2501.19398v1 | 1 |
|
Figure 2:Bounds on the winning probability of non-chameleons. The non-chameleons lose the game with high probability if they use revealing or concealing strategies: The chameleon correctly identifies the secret word for revealing strategies, and the non-chameleons misidentify the chameleon for concealing strategies. The non-chameleons can win the game with a probability that is𝒪(log(N))𝒪𝑁\mathcal{O}(\log(N))caligraphic_O ( roman_log ( italic_N ) )times than the trivial00-KL pairwise concealing strategy. | 2501.19398v1 | 2 |
|
Figure 1:Test-time scaling with s1-32B.We benchmarks1-32Bon reasoning-intensive tasks and vary test-time compute. | 2501.19393v2 | 1 |
|
Figure 2:s1K and s1-32B.(left)s1Kis a dataset of 1,000 high-quality, diverse, and difficult questions with reasoning traces.(right)s1-32B, a 32B parameter model finetuned ons1Kis on the sample-efficiency frontier. SeeLABEL:tab:perffor details on other models. | 2501.19393v2 | 2 |
|
Figure 3:Budget forcing with s1-32B.The model tries to stop after “…is 2.”, but we suppress the end-of-thinking token delimiter instead appending “Wait” leadings1-32Bto self-correct its answer. | 2501.19393v2 | 5 |
|
Figure 6:Rejection sampling on AIME24 with s1-32B.We sample with a temperature of 1 until all generations have less than (from left to right) 3500, 4000, 5000, 8000, and 16000 thinking tokens requiring an average of 655, 97, 8, 3, 2, and 1 tries per sample. | 2501.19393v2 | 10 |
|
Figure 7:Scaling further with parallel scaling methods.All metrics averaged over the 30 questions in AIME24. Average thinking tokens forREBASEdo not account for the additional compute from the reward model. For sequential scaling, we prompt the model to use up to (from left to right) 32, 64, 256, and 512 steps. ForREBASEand majority voting we generate 16 parallel trajectories to aggregate across. | 2501.19393v2 | 11 |
|
Figure 1:Comparison of AQUA-KV to alternative Key-Value Cache compression methods for Llama 3.x models in terms of average LongBench score on 14 english tasks (see Section4). | 2501.19392v1 | 1 |
|
Figure 2:Mean Explained Variance Ratios by linear probes from previous blocks (L), tokens (T) and role on Llama-3.2-3B. | 2501.19392v1 | 2 |
|
Figure 2:Mean Explained Variance Ratios by linear probes from previous blocks (L), tokens (T) and role on Llama-3.2-3B. | 2501.19392v1 | 3 |
|
Figure 3:An intuitive scheme of the AQUA-KV inference.Only the quantized residuals are saved for each block. | 2501.19392v1 | 4 |
|
Figure 4:Additional Mean Explained Variance Ratios by linear probes from previous blocks (L), tokens (T) and role on Llama-3.2-3B. | 2501.19392v1 | 6 |
|
Figure 5:Explained Variance Ratios per Transformer Block for chosen sets of linear probes on Llama-3.2-3B. | 2501.19392v1 | 7 |
|
Figure 1:An illustration of our proposed methodology where the server maintains a pair of global LoRA modules while the devices adaptively update submatrices of the global LoRA modules through sketching during each round. | 2501.19389v1 | 1 |
|
Figure 2:Convergence behavior of FSLoRA and baselines on the GLUE benchmark with the RoBERTa model. Testing accuracy is averaged over seven tasks. | 2501.19389v1 | 2 |
|
Figure 8:Comparison of top-k compression and its integration with sketching, evaluated on the commonsense reasoning benchmark using the LLaMA-3.2-3B model. The results show that combining these two orthogonal techniques significantly enhances performance, demonstrating the benefits of integrating sketching with top-k compression. | 2501.19389v1 | 7 |
|
Figure 10:Comparison of FSLoRA with and without sketching, with an upload budget400×400\times400 ×the global LoRA module size at each rank, evaluated on the LLaMA-3.2-3B model. The number of devices is set to50505050. The results are averaged over eight tasks from the commonsense reasoning benchmark. | 2501.19389v1 | 9 |
|
Figure 1:Process of blurring. (a) The ground-truth sharp imagex𝑥xitalic_x; (b) one simulated PSFkisubscript𝑘𝑖k_{i}italic_k start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT; (c) one ground-truth convolved imageki∗x∗subscript𝑘𝑖𝑥k_{i}\ast xitalic_k start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ∗ italic_x; (d) one blurred imageyi=ki∗x+nisubscript𝑦𝑖∗subscript𝑘𝑖𝑥subscript𝑛𝑖y_{i}=k_{i}\ast x+n_{i}italic_y start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT = italic_k start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ∗ italic_x + italic_n start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT. | 2501.19386v1 | 1 |
|
Figure 2:First row:12121212estimated blur kernels; second row: the associated real (simulated) blur kernels. | 2501.19386v1 | 2 |
|
Figure 3:First row:4444deconvolved images; second row: the associated enhanced deconvolved images. The baseline image is the ground-truth sharp imagex𝑥xitalic_x. The PSNR (peak signal-to-noise ratio) and SSIM (structural similarity) values of each showed image are calculated with respect to the ground-truth imagex𝑥xitalic_x. | 2501.19386v1 | 3 |
|
Figure 5:Process of denoising. (a) One blurred imageyisubscript𝑦𝑖y_{i}italic_y start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT; (b) the associated deconvolved imagex^isubscript^𝑥𝑖\widehat{x}_{i}over^ start_ARG italic_x end_ARG start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT; (c) the associated enhanced deconvolved imagex^i∗subscriptsuperscript^𝑥∗𝑖\widehat{x}^{\ast}_{i}over^ start_ARG italic_x end_ARG start_POSTSUPERSCRIPT ∗ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT; (d) the associated enhanced convolved imageyi~~subscript𝑦𝑖\widetilde{y_{i}}over~ start_ARG italic_y start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT end_ARG. | 2501.19386v1 | 5 |
|
Figure 8:Comparison of the effect ofr1subscript𝑟1r_{1}italic_r start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTon the latent image reconstruction quality, evaluated using two metrics: PSNR (peak signal-to-noise ratio) and SSIM (structural similarity):(a) Without prior (Equation 2.8). (b) With prior (Equation 2.9, proposed). | 2501.19386v1 | 8 |
|
Figure 1:Given any feature representationϕ(x)italic-ϕ𝑥\phi(x)italic_ϕ ( italic_x ), we attach a decoding-based head to output predictive distributionpθ(y|x)subscript𝑝𝜃conditional𝑦𝑥p_{\theta}(y|x)italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( italic_y | italic_x ). | 2501.19383v1 | 1 |
|
Figure 2:Lower (↓↓\downarrow↓) is better. Risk (theoretical and empirical) when varyingK𝐾Kitalic_KandN𝑁Nitalic_Nto fit a truncated𝒩[0,1](0.5,0.252)subscript𝒩010.5superscript0.252\mathcal{N}_{[0,1]}(0.5,0.25^{2})caligraphic_N start_POSTSUBSCRIPT [ 0 , 1 ] end_POSTSUBSCRIPT ( 0.5 , 0.25 start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT )distribution using binary tokenization. Results averaged across 10 runs each. | 2501.19383v1 | 2 |
|
Figure 3:Visualization of fitting a truncated Gaussian distribution. Each levelk𝑘kitalic_kof the binary tree represents the empirical fit usingk𝑘kitalic_kbits, and each bin gets subdivided into two. | 2501.19383v1 | 3 |
|
Figure 4:Fit to ground truth is better. Curve fitting plots for various 1D functions. Both models are trained over unlimited(x,y)𝑥𝑦(x,y)( italic_x , italic_y )points wherex𝑥xitalic_xis sampled from a bounded range. | 2501.19383v1 | 4 |
|
Figure 6:Higher (↑↑\uparrow↑) is better. Kendall-Tau regression scores over AMLB and OpenML-CTR23 tasks using up to 10K maximum training points. Each bar averaged over 10 runs. | 2501.19383v1 | 6 |
|
Figure 7:Lower (↓↓\downarrow↓) is better. Relative mean squared error (MSE) over selected AMLB tasks. Each method used a min-max linear scaling normalization ony𝑦yitalic_y-values. | 2501.19383v1 | 7 |
|
Figure 8:Fit to ground truth is better. Density estimation visualization over various shapes using an unnormalized decoder with vanilla temperature sampling. | 2501.19383v1 | 8 |
|
Figure 9:Lower (↓↓\downarrow↓) is better. NLL over UCI datasets, when varying different axis (layers, heads, units) from a fixed default of (3, 4, 128) respectively. | 2501.19383v1 | 9 |
|
Figure 10:Lower (↓↓\downarrow↓) is better. Relative MSE over selected AMLB tasks, when varying output repetitions. | 2501.19383v1 | 10 |
|
Figure 11:Lower (↓↓\downarrow↓) is better. Regression performance as a function of training data scaling between using the normalized decoder vs. Reimannian distribution as regression heads. Each point was averaged over 10 training runs over random combinations of datapoints from the original AMLB task’s training set. | 2501.19383v1 | 11 |
|
Figure 12:Higher (↑↑\uparrow↑) is better. Extended results from Table5in the main body. Regression performance as a function of input dimension over BBOB functions using Kendall-Tau correlation. Each point was averaged over 10 training runs, each with 100K training points(x,y)𝑥𝑦(x,y)( italic_x , italic_y )with eachx𝑥xitalic_xsampled uniformly from[−5,5]55[-5,5][ - 5 , 5 ]coordinate-wise.Note:Some functions such as RosenbrockRotated or GriewankRosenbrock are undefined when dimension is 1, so we skip those points. | 2501.19383v1 | 12 |
|
Figure 13:Lower (↓↓\downarrow↓) is better. Regression performance while vary sampling size fromy∼pθ(⋅|x)y\sim p_{\theta}(\cdot|x)italic_y ∼ italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( ⋅ | italic_x )using binary tree-based tokenization vs. Hamming representation on normalized decoder with mean aggregation. Each point was averaged over 10 training runs over random size-1000 combinations of the original AMLB task’s training data points. | 2501.19383v1 | 13 |
|
Figure 14:Visualizing density estimation ofp(y|x)𝑝conditional𝑦𝑥p(y|x)italic_p ( italic_y | italic_x )on 1D problems. We used an unnormalized decoder with(B=10,E=1,M=5)formulae-sequence𝐵10formulae-sequence𝐸1𝑀5(B=10,E=1,M=5)( italic_B = 10 , italic_E = 1 , italic_M = 5 ). | 2501.19383v1 | 14 |
|
Figure 1:The high-level workflow of the proposed semantic graph based loop closure system integrated into a SLAM framework. The proposed loop closure algorithm takes two semantically segmented point clouds as input, which are converted to semantic graphs. After that, semantic graph encoders are deployed to compress them into graph vectors. Finally, the graph comparison module predicts the similarity of the two loop candidates. When the similarity exceeds a specific threshold, a pose constraint is estimated using semantic registration, which is added to the pose graph for trajectory optimization. | 2501.19382v1 | 1 |
|
Figure 2:The architecture of the proposed semantic graph encoder. The semantic graphs created from input point clouds are passed through three GATs to extract contextual spatial, semantic and geometric features. These features are then concatenated and passed through a self-attention module to produce a node embeddingf𝑓fitalic_f. Another self-attention module operated on the node embeddingf𝑓fitalic_fto learn a global context vectorc𝑐citalic_c. We finally project the node embeddingf𝑓fitalic_finto the global context vectorc𝑐citalic_cto obtain corresponding node weights and use them to calculate the final graph vectore𝑒eitalic_e. | 2501.19382v1 | 2 |
|
Figure 3:Overview of the graph comparison module. We propose a relative difference vector (shown in orange color) as the absolute value of the difference between two graph vectors. The similarity vector is leart based from first-order and second-order difference vectors, and concatenated graph vectors. This similarity vector is then passed through fully connected layers to predict the similarity value between two input graph vectors. | 2501.19382v1 | 3 |
|
Figure 4:Precision-Recall curves of max F1 score on SemanticKITTI dataset. In the legend, AUC denotes the area under curve. Here we compare, Ours and Ours-RN with SGPR[19], SGPR-RN, ScanContext (SC)[14]and ISC[38]. It can be seen that Ours and Ours-RN outperform other methods on all sequences, and especially on Sequence 08, where there are many reverse loop closures. | 2501.19382v1 | 4 |
|
Figure 6:Trajectories of complete SLAM system based on our loop closure module, ISC-LOAM and the ground truth of sequence 02 from SemanticKITTI. | 2501.19382v1 | 6 |
|
Figure 1:The five images from left to right are examples of signal-present images associated with the considered MVNLumpy. Gaussian noise was added to simulate the measured image data. The rightmost image shows the Gaussian signal to be detected. | 2501.19381v1 | 1 |
|
Figure 4:From left to right: Five examples of signal-present images generated by use of VICTRE mammography ROIs and the spiculated mass signal to be detected. | 2501.19381v1 | 9 |
|
Figure 6:The first eight PLS channels (top) and L-grad channels (bottom) produced by use of 2000 images. The L-grad channels appear cleaner. | 2501.19381v1 | 13 |
|
Figure 1:Overview of the challenges and proposed solutions in this work. Tabular data is inherently structured, dense, concise, and numerical. Based on these characteristics, we identify four key challenges. To address them, we propose four targeted solutions. The gray arrows between the characteristics and challenges represent the potential causes of these challenges stemming from specific characteristics. Each proposed solution corresponds to the challenge presented on the left in the same row.TableMasteris a unified recipe developed based on these findings. | 2501.19378v2 | 1 |
|
Figure 2:Experimental analysis of challenges in table understanding with language models. (a) Impact of table size on task difficulty. (b) Effect of verbalized tables with enriched semantic context. (c) Performance comparison of different reasoning methods on calculation-required versus non-calculation questions. (d) Performance differences when processing normalized versus noisy tables. | 2501.19378v2 | 2 |
|
Figure 3:The framework ofTableMaster. It comprises three stages: (1) table structure understanding, where the table’s structure is analyzed, and a table-of-focus is constructed through row and column lookup; (2) table content understanding, where the table-of-focus is reconstructed based on the question, and its information is verbalized to enhance the semantic context; and (3) table reasoning for question answering, where an adaptive reasoning strategy determines whether to use textual reasoning or text-guided symbolic reasoning to derive the final answer. The dashed arrows indicate optional workflows, such as the table-of-focus re-construction and incorporating text-guided symbolic reasoning. | 2501.19378v2 | 3 |
|
Figure 4:An example (fetaqa-164) from the FetaQA dataset where the result is accurate, but the evaluation metric assigns a low score. | 2501.19378v2 | 4 |
|
Figure 5:Performance Comparison Across Table Sizes (Row Count, Column Count, Area Size, Token Count). | 2501.19378v2 | 5 |
|
Figure 6:The row count distribution in the WikiTQ dataset and the analysis of accuracy variation with different peek sizes. | 2501.19378v2 | 6 |
|
Figure 7:Changes in Table Condensation After Table-of-Focus Construction in Table Structure Understanding. | 2501.19378v2 | 9 |
|
Figure 10:Prompt for structure extraction inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model in extracting the table’s structure. | 2501.19378v2 | 12 |
|
Figure 11:Prompt for column ranking inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to rank the priority of all columns based on the given table, top headers, and related question. | 2501.19378v2 | 13 |
|
Figure 12:Prompt for column lookup inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to select relevant columns based on the given table, top headers, and related question. | 2501.19378v2 | 14 |
|
Figure 13:Prompt for SQL generation for row lookup inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to generate SQL for selecting relevant rows based on the given table and related question. | 2501.19378v2 | 15 |
|
Figure 14:Prompt for table verbalization inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to verbalize the given table by adding detailed descriptions and additional knowledge about the table. | 2501.19378v2 | 16 |
|
Figure 15:Prompt for information estimation inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to evaluate the given table’s content and determine whether it contains sufficient information to answer the provided question | 2501.19378v2 | 17 |
|
Figure 16:Prompt for reasoning strategy assessment inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to evaluate whether answering the given question requires direct information retrieval, counting, or mathematical calculations based on the table’s content. The response determines the subsequent reasoning strategy. | 2501.19378v2 | 18 |
|
Figure 17:Prompt for textual reasoning inTableMaster. Blue text represents placeholders for variables within the prompt, while the grey region indicates optional sections to adapt the prompt for question-answering or fact-verification tasks. The prompt guides the language model to answer the question step by step. | 2501.19378v2 | 19 |
|
Figure 18:Prompt for textual guidance generation inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to generate textual guidance that can be utilized for subsequent symbolic reasoning. | 2501.19378v2 | 20 |
|
Figure 19:Prompt for symbolic reasoning inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to generate Python code to answer the question. | 2501.19378v2 | 21 |
|
Figure 20:Prompt for answer formatting inTableMaster. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to format the final answer based on the given table, question, and reasoning process. | 2501.19378v2 | 22 |
|
Figure 21:Direct prompt for table understanding in analysis experiment. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to directly give the final answer based on the given table and question. | 2501.19378v2 | 23 |
|
Figure 22:Chain of thought prompt for table understanding in analysis experiment. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to give the answer step by step based on the given table and question. | 2501.19378v2 | 24 |
|
Figure 23:Program of thought prompt for table understanding in analysis experiment. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to generate code to derive the answer based on the given table and question. | 2501.19378v2 | 25 |
|
Figure 24:Prompt for table verbalization in analysis experiment. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to verbalize a table to add detailed description. | 2501.19378v2 | 26 |
|
Figure 25:Prompt for textual guidance generation in analysis experiment. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to generate textual guidance that used for symbolic reasoning. | 2501.19378v2 | 27 |
|
Figure 26:Prompt for reasoning strategy evaluation in analysis experiment. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to select the better reasoning process after table reasoning. | 2501.19378v2 | 28 |
|
Figure 27:Prompt for reasoning strategy evaluation in analysis experiment. Blue text indicates placeholders for variables within the prompt. The prompt guides the language model to select the better reasoning strategy before table reasoning. | 2501.19378v2 | 29 |
|
Figure 28:Prompt for classifying a question type based on whether calculation is required in the analysis experiment. Blue text indicates placeholders for variables within the prompt. | 2501.19378v2 | 30 |
|
Figure 29:Prompt for generating noised tables in the analysis experiment. Blue text represents placeholders for variables within the prompt. The prompt instructs the language model to add noise by altering the cell content format based on a given table. | 2501.19378v2 | 31 |
|
Figure 1:A typical pipeline for user input processing in virtual assistants (a) and our proposed approach (b). | 2501.19377v2 | 1 |
|
Figure 3:DET curves for VT detection and DDSD experiments from TableII. The False Accept Rate (FAR) represents either unintended queries or queries without the trigger phrase that were falsely classified as intended/containing the trigger phrase and the False Reject Rate (FRR) represents either intended queries or queries containing the trigger phrase that were falsely classified as unintended/not containing the trigger phrase.
The markers on each curve indicate the EER. Dotted lines represent baselines and solid lines represent our approach. | 2501.19377v2 | 3 |
|
Figure 1:10m wind speed and mean sea level pressure for winter storm Eunice, 18 Feb 2022 at 0h UTC. Top: HRES data at ¼° (ground truth), middle: 3.5d forecast produced by GraphCast, bottom: this work. This work produces an overall sharper forecast, with a better prediction of the winter storm’s strength. | 2501.19374v1 | 1 |
|
Figure 2:Amplitude ratio (solid) and coherence (dashed) for the spherical harmonic modes of temperature at 850hPa for 1° GraphCast during the training process. At top, values for 6h lead time during the single-step pre-training phase and at bottom, values for 6h–72h during the forecast rollout (batches 300,000–311,000, incrementing one step every 1,000 batches). | 2501.19374v1 | 2 |
|
Figure 3:10m wind speed and mean sea level pressure for Hurricane Ian, 28 Sept 2022 at 12h UTC. Top: HRES data at ¼°, middle: 5d forecast produced by the control GraphCast model, bottom: the model after 12-step fine-tuning with AMSE. | 2501.19374v1 | 3 |
|
Figure 4:Amplitude ratio (solid) and coherence (dashed) for all output variables and levels, weighted using the variable/level weights in the loss function, for the control model and this work after the 1-step training and after complete fine-tuning. Top: 6h lead time, middle: 120h (5d) lead time, bottom: 240h (10d) lead time. The dashed line is placed where a model would underrepresent the power spectral density by 25%. | 2501.19374v1 | 4 |
|
Figure 5:Lagged ensemble statistics for geopotential (z) at 500hPa, temperature (t) at 850hPa, specific humidity (q) at 700hPa, and 2-meter temperature (2t) from left to right. The statistics are the CRPS, root mean squared error of the ensemble mean, and spread-error ratio, from top to bottom. | 2501.19374v1 | 5 |
|
Figure 6:Predictions of tropical cyclone intensity ((a), mean maximum surface wind speed; (b) mean minimum central pressure) and mean absolute position error (c) for forecasts initialized 20 June–19 September 2022. Orange squares show statistically significant differences between the AMSE AR12 and control predictions. | 2501.19374v1 | 6 |
|
Figure 8:Quantile-quantile plots of 10m wind speed at surface station locations for the North American domain. At left, 1 Jan–30 March 2022 (boreal winter), and at right 20 June–19 September 2022 (boreal summer). The control and AMSE AR12 points show model evaluations for 5-day forecasts. The shaded region denotes confidence interval based on the Kolmogorov-Smirnov test. | 2501.19374v1 | 8 |
|
Figure 9:As in figure8, for 2m temperature. Low percentiles (extreme cold) are shown for the Northern Hemisphere winter, and high percentiles (extreme heat) are shown for the Northern Hemisphere summer. | 2501.19374v1 | 9 |
|
Figure 10:CRPS skill score (% improvement), measured as the relative difference between the CRPS of the 12-step fine-tuned model and the CRPS of the control model, for a selection of variables and lead times. Orange up-arrows show where the fine-tuned model performs better, blue down-arrows show where the control model performs better. Hollow arrows represent a difference of less than 1%, and differences of 2% or larger are marked. Hollow circles mark values that are not statistically significant at the 90% level. | 2501.19374v1 | 10 |
|
Figure 1:The objective is to map missing values back to their origin by leveraging a consistency model with conditional information. Given a Probability Flow ODE (PF-ODE) that progressively adds noise to the missing values, the model learns to reconstruct points along the ODE trajectory, allowing accurate imputations without the need for a full reverse path. | 2501.19364v1 | 1 |
|
Figure 2:Graphical representation of the architecture implemented in CoSTI. The figure illustrates the two information channels (primary and conditional), the internal structure of the U-Net, and the details of the STFEM and NEM blocks. | 2501.19364v1 | 2 |
|
Figure 1.LLMs slightly outperform humans on the AUT and DAT, but humans slightly outperform LLMs on FF. | 2501.19361v1 | 1 |
|
Figure 2.LLM responses exhibit far less variability than human responses, as measured by cosine distance between embedded responses. | 2501.19361v1 | 2 |
|
Figure 3.LLM responses cluster together in feature space more than do human responses.K-means clustering of TSNE of AUT sentence embeddings. | 2501.19361v1 | 3 |
|
Figure 4.LLM responses have far more words in common than do human responses.We look at word overlaps between “full” responses from LLMs and humans—e.g. all uses from the AUT, all words in the FF, etc. This corresponds to the sentence embedding method of population originality measurement, and explains why the difference between LLMs and humans is more pronounced in this setting. | 2501.19361v1 | 4 |
|
Figure 5.Effect of different AUT prompt wordings on length of LLM AUT responses.We use prompt verison 3 in most experiments in this paper, since LLM responses to this prompt most closely match the human distribution of response lengths. | 2501.19361v1 | 5 |
|
Figure 6.Even when considering only one-word responses to control for response structure, LLM AUT responses have lower population-level variability (left plot) and are closer in feature space (right plot) than human responses. LLM responses are generated with prompt version 3. We create sentence embeddings from only single-word uses provided by AUTs and humans, ignoring all longer responses. | 2501.19361v1 | 6 |
Subsets and Splits