Text Classification • 22.7M • Updated • 426
orig_reference_answer stringlengths 2 4.83k | orig_response stringlengths 0 6.19k | orig_score float64 1 5 |
|---|---|---|
What is the expression for the final error asymptotically in the limit where Ns Nx → 0? | What does the dependence on the upper-triangular components in Theorem 3.1 imply about the final error? | 5 |
What is the primary challenge in diagnosing dementia, especially Alzheimer's Disease (AD) and frontotemporal dementia (FTD)? | What is DiaMond and what problem does it aim to solve? | 3.25 |
['The spherical harmonic transform of a NILC map is expressed using needlet filters and harmonic transforms of weighted frequency maps.', 'The subsequent step involves computing the harmonic transform of a weighted frequency map.', "This approach is inspired by the MASTER formalism, which establishes a connection betwe... | ['The spherical harmonic transform of a NILC map is expressed using needlet filters and weighted frequency maps.', 'The next step is to explicitly compute the harmonic transform of a weighted frequency map.', 'The MASTER formalism is used to express the spherical harmonic transform of a masked map.', 'The weight map is... | 5 |
Logdet SMI has a complexity of O(G^3) for precomputation and O(G^2) for each oi, making it less efficient and challenging for scaling compared to Graph-Cut SMI. Table 4 shows Logdet SMI is over 35x more expensive (1573.93 µs/prompt) than Graph-Cut SMI (44.15 µs/prompt). | The key distinction is the scalability of both methods. Logdet is over 35× more expensive than our method (≈ 1573.93 µs/prompt), and the complexity of Logdet is also a bottleneck. Our method offers a more efficient balance between performance and efficiency. | 4.25 |
{"question":"What are the limitations of using comments and docstrings for Semantic Engineering?","answer":"Comments and docstrings are somewhat inflexible because they mix documentation with program meaning, blurring their intended purpose. They already play established roles in software development, such as documenta... | {"question":"What are the limitations of using comments and docstrings for semantic intent?","answer":"Comments and docstrings already have established roles in software development, such as documentation and usage guidance, whereas semantic intent requires a dedicated mechanism."} | 1 |
Existing methods that skip computation of unimportant experts in MoE layers are designed for text-only LLMs, and directly applying them to MoE multimodal large language models (MLLMs) overlooks key factors such as global contribution. | Existing methods that skip computation of unimportant experts in MoE layers are designed for multimodal LLMs, and directly applying them to MoE text-only large language models (LLMs) captures key factors such as local contribution. | 2.5 |
['Non-linear function approximators are crucial for effective learning and generalization in vast state spaces, yet their convergence is not assured.', 'NFQCA employs batch learning to enhance stability when utilizing neural network function approximators, a method that becomes computationally prohibitive with extensiv... | ['Non-linear function approximators are indispensable for learning and generalization on large state spaces.', 'Batch Learning (Breadth Learning) is used for stability, which is intractable for large networks.', 'A minibatch version of NFQCA is equivalent to the original DDPG.', 'Deep DPG (DDPG) is a modified DQN inspi... | 1 |
high-level abstraction is shown in Listing 3. The evolutionary search discarded the standard linear averaging and static discounting of the CFR family in favor of three distinct, non-intuitive mechanisms: • Volatility-Adaptive Discounting (vs. Static Discounting):Standard DCFR applies fixed discount factors (𝛼, 𝛽) to... | The source code of the discovered algorithm is available in the Appendix. 4.2.2. Code Generation We use the CFR+ algorithm to generate code for our experiments. CFR+ uses the FASEL version of CFR, which is a modified version of CFR that incorporates a new optimization technique called FASEL. This technique allows CFR+ ... | 1 |
Figure 2(b) indicates that the learning rate is not preserved across different granularities in MoE. | Figure 1.4 (a) shows the optimal learning rate is preserved across a varied number of experts when using our µP for MoE. | 3.5 |
Empirical results suggest that using the CD objective is crucial for robust generalization of the hypernetwork. | Empirical results suggest that using the CD objective is crucial for robust generalization of the hypothesis network. | 3 |
The adversary in Prompt-GDRO is driven by an intensive statistic, specifically the mean prompt loss per bin, rather than an extensive cumulative loss. This statistic is also smoothed over time. | The adversary in Prompt-GDRO is driven by an intensive statistic, specifically the median prompt loss per bin, rather than an extensive cumulative loss. This statistic is also smoothed over time. | 3 |
{"question":"What specific data is required to optimize a separate neural continuous volume representation network for each scene?","answer":"The required data includes a dataset of captured RGB images of the scene, the corresponding camera poses and intrinsic parameters, and scene bounds. "} | {"question": "What specific data is required to optimize a separate neural continuous volume representation network for each scene?", "answer": "The required data includes a dataset of captured RGB images of the scene, the corresponding camera poses and intrinsic parameters, and scene bounds."} | 5 |
The initial goal was to approximate a sparse structure implied by [2] for vision networks and cover the hypothesized outcome by dense, readily available components. | The initial goal was to approximate a sparse structure implied by [2] for vision networks and to cover the hypothesized outcome by dense, readily available components. | 1 |
The matching algorithms tested included naive string matching, the Roberta model, and ChatGPT. | The matching algorithms tested excluded naive string matching, the Roberta model, and ChatGPT. | 3 |
['The system maps observed signals to a latent representation and reconstructs the original signal using a decoder.', 'Unlike classical autoencoders, the design is asymmetric: the encoder processes only the partial observed signal without mask tokens, while the decoder reconstructs the full signal using the latent repr... | ['The approach uses an asymmetric design where the encoder extracts partial observations from the latent representation, and a lightweight decoder reconstructs the full signal from the latent representation and mask tokens.', 'Images are divided into non-overlapping patches, with a subset sampled without replacement.',... | 3.75 |
['Semantic Engineering integrates enriched semantics into prompt generation within the Jac programming language.', 'A benchmark suite is introduced to simulate realistic AI-Integrated application scenarios.', 'Semantic Engineering significantly enhances prompt fidelity.', 'The performance of Semantic Engineering is com... | ['Semantic Engineering enhances the natural-language context used in the Jac programming language.', 'A benchmark suite is introduced to reflect realistic AI-Integrated application scenarios.', 'Semantic Engineering substantially improves prompt fidelity.', 'AI-Integrated applications use LLMs at runtime to perform ess... | 1 |
**Question:** What was the accuracy of PaLM 540B+chain-of-thought on the GSM8K dataset?
**Answer:** The accuracy was not explicitly stated, but the study included PaLM results on GSM8K with direct prompting for completeness. | **Question:** What was the accuracy of PaLM 540B+chain-of-thought on the GSM8K dataset?
**Answer:** The accuracy was not explicitly stated, but the study included PaLM results on GSM8K with direct prompting for completeness. | 5 |
Focusing only on sample-prototype relationships can reduce diversity, disrupt within-class data distribution, and lead to forgetting as older representations shift toward current ones. | Focusing only on sample-prototype relationships can reduce diversity, enhance within-class data distribution, and lead to forgetting as older representations shift toward current ones. | 3 |
**Question:** What is the key question addressed in the general reasoning section?
**Answer:** The key question is whether the generalization effect of training language models on reasoning-intensive domains still holds when the training curriculum is self-generated by R-Zero, rather than human-labeled. | **Question:** What is the main finding regarding the generalization effect of the R-Zero model?
**Answer:** The main finding is that the generalization effect of the R-Zero model is significantly less than that of the baseline model. | 2 |
An important fact from the passage is that **Perplexity is the most effective metric for selecting training samples**, as it removes the "easiest" samples and improves model performance, outperforming other metrics like L2-Norm Error and Memorization. | Perplexity is the least effective metric for selecting training samples, as it adds the 'easiest' samples and degrades model performance, underperforming other metrics like L2-Norm Error and Memorization. | 2.5 |
[{"subject": "post-layernorm", "relation": "does not have problem with", "object": "norm growth of residual"}, {"subject": "post-layernorm", "relation": "performs better for UTs than", "object": "pre-layernorm"}] | [{"subject": "Post-layernorm", "relation": "is", "object": "a drawback in regularization"}, {"subject": "Post-layernorm", "relation": "is a limitation in gradient flow", "object": "versus"}] | 1 |
[{"question": "What is a limitation of current deep-learning-based optimization methods for kernel regression?", "answer": "Current methods are limited to a maximum of 16 kernels on small blocks of 8x8 or 16x16 pixels, and they use a fixed number of kernels for each block, which limits adaptation to content. Independen... | [{"question": "What is the main advantage of traditional gradient descent (GD) optimization for kernel regression?", "answer": "Traditional GD involves a decoder that predicts kernel parameters, which is computationally expensive."}, {"question": "How do deep-learning-based optimization methods differ from traditional ... | 1 |
The FCV Rate captures the likelihood that a functionally correct patch is also vulnerable, quantifying the fraction of successful fixes that introduce new security weaknesses. | The dataset is divided into a set of N attacks (npass) and a set of N fails (npass) for each instance. The Pass@1 metric measures the proportion of instances that produce a functionally correct patch, and Pass@1=n pass/N indicates that the vulnerability is present, but not well-defined. The FCV Rate captures the likeli... | 1 |
They enhance accessibility and efficiency and significantly contribute to research in natural language understanding and generation, pushing AI’s capabilities in understanding complex language constructs. | Enhancing accessibility and efficiency. | 3.75 |
**Question:** What is the purpose of the baseline model mentioned in the text?
**Answer:** The baseline establishes the performance of the base model before any distillation or reinforcement learning. | The baseline model establishes the performance of the base model before distillation and reinforcement learning, but not after. | 3 |
bedding extraction is computationally expensive, we compute it once per image and then for each prompt we generate a prompt embedding. These embeddings are then fed into a mask decoder that produces the final segmentation masks. | bedding extraction is computationally free, we compute it multiple times per image and then for each prompt we generate an image embedding. These embeddings are then fed into a prompt encoder that produces the final bounding boxes. | 2.5 |
**Question:** What are some methods used in Semi-Supervised Continual Learning?
**Answer:** Some methods include ORDisCo, which uses a semi-supervised classification model with a conditional GAN for generative replay; AAE, which trains an adversarial autoencoder to reconstruct images; CCIC, which imposes predictive co... | **Question:** What is Semi-Supervised Continual Learning (SSCL)?
**Answer:** SSCL is an initial attempt where a semi-supervised classification model and a conditional GANs are used for generative replay in SSCL. | 1 |
- (MinHash-based methods, require building, massive vector sets)
- (SimHash, generates, sample's feature vector solely from the words it contains) | - (MinHash-based methods, require, building massive vector sets)
- (MinHash-based methods, are, extremely expensive in terms of time and space) | 1 |
Figure 2 presents a sample prompt and output from Search-R1, a search-enhanced reasoning-based RAG model, demonstrating how the generator invokes searches when more knowledge is needed to answer a question. | Figure 3 presents a sample prompt and output from Search-R1, a search-free reasoning-based RAG model, demonstrating how the generator invokes searches when more knowledge is needed to answer a question. | 3 |
Current reasoning models often involve redundant reasoning, which significantly increases computational cost and inference time. | The main issues are: 1. The inability to generate correct answers with minimal performance loss due to length penalties. 2. The significant performance loss when shortening reasoning due to performance issues. 3. The lack of focus on guiding RL for efficient reasoning, leading to minimal performance drop. | 2.5 |
It improves reasoning quality by leveraging graph-structured outcome rewards and stratified clipping advantage estimation. | The main goal of PASC-GRPO is to improve reasoning quality while mitigating reward hacking. | 1 |
In-context example selection heavily influences the multilingual performance of LLMs. Finding semantically similar examples is important, but using semantically dissimilar or peculiar exemplars has also been shown to enhance performance. For ambiguous sentences, selecting exemplars with polysemous or rare word senses m... | Based on the passage, the significance of in-context example selection in multilingual LLMs is that it **heavily influences multilingual performance**. Specifically, selecting semantically similar examples is important, but using semantically dissimilar (peculiar) exemplars can also enhance performance. Additionally, f... | 5 |
Yes, S2MoE consistently outperforms both baselines regardless of backbone size or the number of experts activated. | Yes, S2MoE consistently underperforms both baselines, regardless of backbone size or the number of experts activated, demonstrating its potential to scale effectively in large language models. | 3.5 |
The advent of large reasoning models, such as GPT-o1 [4] and DeepSeek-R1 [5], represents a significant breakthrough in natural language processing, especially in tackling complex and intricate reasoning tasks. An interesting phenomenon observed during reinforcement learning post-training via Group Relative Policy Optim... | CoT [3, 10, 32] and CoT-Prompt [10] have shown that chain-of-thought prompting can improve reasoning ability of Large Lan- guage Models (LLMs). CoT and CoT-Prompt are designed to extend the length of CoT [32] while CoT-Prompt is designed to extend the length of CoT while also processing data from external sources, such... | 1.25 |
What is the composition of the PaLM pretraining dataset? | PaLM's pretraining data consists entirely of Common Crawl web pages, with no social media or books included. | 2.5 |
Step 4 of EPE, corresponding to lines 12 to 20 in Algorithm 1, involves generating code for each prompt and evaluating each generated code sample using the same test cases. A fitness function, based on the ratio of test cases passed, is defined to evaluate the generated code. | Step 4 of EPE, which corresponds to lines 12 to 20 in Algorithm 1, involves generating code for each prompt and evaluating these generated code samples using the same test cases. A fitness function is defined based on the ratio of test cases passed. Prompts are selected based on their passing rate, and the process conc... | 4.75 |
- Naive RL can fail catastrophically in LLMs, even with advanced algorithms like Dr.GRPO.
- Introducing a reliable oracle judge for reward calibration prevents pitfalls like reward hacking and mode collapse.
- The final model outperforms its base version on competitive math and coding benchmarks.
- The model is cultura... | - Naive RL can fail catastrophically in LLMs, even with advanced algorithms like Dr.GRPO.
- Introducing a reliable oracle judge for reward calibration is an effective solution to prevent pitfalls like reward hacking and mode collapse.
- The final model outperforms its base version on competitive math and coding benchma... | 4.75 |
enhanced through Rein- forcement Learning with Verifiable Rewards (RLVR). Group Relative Policy Optimization (GRPO) has emerged as the de facto approach for RLVR, inspiring numerous variants. How- ever, our mathematical analysis reveals that these methods are fundamentally weighted variations of GRPO. We provide a unif... | reduced with more models. However, the effectiveness of these models is highly sensitive to the choice of weights for weight initialization. In this paper, we propose to weight initialization with a difficulty-aware reweighting strategy, which is based on the training process of reinforcement learning, to improve the i... | 1 |
Recent approaches have shifted from static, predefined structures to dynamically learned relationships that capture latent dependencies between stocks. This includes methods like LSTM with graph convolutions (HATS), which introduces multi-relational attention mechanisms, and FinGAT, which applies dynamic attention to q... | Recent approaches have shifted from static, predefined structures to dynamically learned relationships that capture latent dependencies between stocks. | 3.75 |
["What is the primary challenge introduced by automatically adapting the gradient update step size?", "What optimization algorithm is favored in practice, and what guidelines are followed for its implementation?", "How is a baseline memory-free offCL method adapted to onCL, and what is the resulting loss function?"] | ["What is the recommended update method for a baseline memory-free offCL method?", "How is the overall loss calculated in the context of onCL when using a baseline memory-free method?", "What is the general procedure for adapting a baseline memory-free offCL method to onCL?"] | 1 |
framework that interleaves data collection and model training (Pang et al., 2024; Zhang et al., 2025). Initially, we set the model parameters to θ(0) = θ after the warm-up phase and collect an initial dataset D(0). At iteration m, we optimize model parameters θ(m) and generate fresh samples {Q(m) i,0 , Q(m) i,1 , · · ·... | method to iteratively improve model performance by iteratively improving the given task. First, we devise a new task to train the model to predict the probability of the next task. Second, we fine-tune the same model on task negative samples to train it to predict the probability of the previous task. Finally, we use t... | 1 |
The finetuning data contained 7.75B tokens, with 5.9B being Python code. | The performance of PaLM-Coder 540B increased even further from 8B, achieving 88.4% pass@100 and 80.8% pass@80 on HumanEval and MBPP respectively. | 2 |
recent research has introduced RLVR. RLVR converts open-ended outputs into programmatically checkable signals, such as numerical consistency in mathematics, unit-test pass rates in code generation, or formal constraint satisfaction (Su et al., 2025; Wang et al., 2025b), thereby avoiding the noise and cost of preference... | we propose a group-based policy optimization approach that eliminates the need for explicit reward modeling. Instead, we directly leverage a set of candidate responses sampled from the current policy, using their rewards to compute a baseline for each token. This allows us to define advantage estimates relative to the ... | 5 |
attacks, the number of re- trieved documents increases with k while a much slower increase regarding leaked documents (Re- peat Contexts/ Rouge Contexts). For the targeted attack, we can find an increase in targeted informa- tion on HealthCareMagic dataset, but no increase on Enron Mail dataset. Therefore, the increase... | re- serve, the model is more vulnerable to the attack. For the target-only attack, the model is more vulnerable to the attack when the embedding model is larger than 100. The results are shown in Figure 4. 4.3.3 Re-use-Patch-Based Attack Effectiveness We evaluate the reuse-patch-based attack effectiveness of the propos... | 1 |
{"question":"What are the two important findings revealed by the experiments?","answer":"Generating CadQuery directly is simpler and yields better performance, and larger pretrained models bring stronger CAD generation ability."} | {"question":"What is the main finding regarding the scaling trend of model size and CAD generation accuracy?","answer":"The scaling trend observed is that as model size increases, the generation accuracy of CadQuery directly produces better performance. This is supported by the observation that larger pretrained models... | 3.75 |
CRPO is a method used because simply removing interference tokens doesn't always improve performance (only about 20% of prompts show improvement). CRPO applies interference token purification to get rollouts with more successful samples. When the original prompt has a low success rate (below τ), CRPO uses rollouts from... | CRPO is a Calibrated Rollout Policy Optimization (CRPO) that applies interference token purification to obtain rollouts with a higher proportion of successful samples. CRPO uses this signal to guide policy optimization on the original prompt, therebyivating the transferability of supervision to the interference-aware c... | 4.25 |
The initial piling height of steel grains in the hopper is approximately 400 mm (Hp ≈ 400 mm). | The initial piling height of steel grains in the hopper is approximately 300 mm (Hp ≈ 300 mm). | 3 |
LNQK𝑄𝐾𝑇 Soft-maxV 𝑊𝑜 concat. heads concat. headsX LN 𝑄𝐾𝑇 GeLU LNGeLU 𝑊1 𝑊1 𝑊2 𝑊2 𝑊1 𝜆12 𝜆11 𝜆21 𝜆31𝜆32 𝜆22 𝜇11𝜇12𝜇21 𝜇22𝜇32𝜇31 LNQK𝑄𝐾𝑇Soft-maxV 𝑊𝑜concat. headsLNGeLU𝑊1 𝑊2 Training InferenceX Structural Reparameterization (Absorb LN parameters into 𝑾𝑸,𝑲,𝑽, sum each weight across all b... | (𝑁 𝑛𝑖=1) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(𝑋) 𝑋(� | 1 |
- Layer-level and neuron-level vectors in deep layers store logits related to final predictions.
- The importance of a neuron in deep layers is determined using the log probability increase, calculated as Imp(vl) = log(p(s|vl + hl−1)) − log(p(s|hl−1)). A large importance score indicates that the neuron significantly en... | - A vector stores logits about the prediction s if the multiplication of this vector with the unembedding matrix results in a high log probability for the remaining neurons in the layer.
- The importance of a neuron in a layer is calculated as the log probability increase of its input layer output, given that its value... | 1 |
[{"subject": "Deep neural network (DNN) models", "relation": "demonstrated efficacy in", "object": "malware detection"}, {"subject": "This paper", "relation": "introduces", "object": "a deep learning-based sequential model for malware analysis"}] | [{"subject": "Deep learning models", "relation": "are effective for", "object": " malware analysis"}, {"subject": "Deep learning models", "relation": "are applicable to", "object": "science and engineering"}] | 1 |
[{"question": "What are the limitations of using flat, joined sequences for incorporating stimulus influence into response representations in RNs?", "answer": "Flat, joined sequences lack a mechanism to highlight or control the influence of the stimulus over the response. They lose token-specific interaction, introduce... | [{"question": "What are the drawbacks of the Mean Pooling or Vector Operations method?", "answer": "They reduce the stimulus to a single vector, have a simple addition of stimulus and response vectors, lack distinguishable attention or selective focus, and oversimplify influence because the simple addition does not acc... | 2.5 |
["Ground-truth boxes are associated with anchors, which are assigned to specific pyramid levels.", "Yes, the parameters of the heads are shared across all feature pyramid levels. Evaluating the alternative without sharing parameters showed similar accuracy, indicating that all levels of the pyramid share similar semant... | ["Ground-truth boxes are associated with anchors, which are assigned to pyramid levels.", "No extra rules are introduced in the following paragraphs, including sharing parameters between levels.", "The good performance of the FPN indicates that all levels of the pyramid share similar semantic levels, which is analogous... | 1 |
Internal or proprietary terms, including system names, product codes, subscription types, and technical jargon, must be generalized or replaced with neutral descriptors. | Internal or proprietary terms should be generalized or replaced with neutral descriptors. Avoid direct quotes or verbatim inclusion of user-suppressed texts unless the query is focused on the specific functional intent of the user. | 4.25 |
['It is imperative that RALMs maintain the confidentiality of user data stored in private retrieval databases.', 'To mitigate biases in the evaluation of reward models, data collection strategies ought to encompass a variety of sources.', 'The presence of extended contexts within RAG prompts presents a significant hurd... | ['RAG requires privacy protection to prevent user privacy information disclosure from the private retrieval database.', 'Data collection should include a wide range of diverse sources to avoid biases in reward model evaluation.', 'Providing high-quality preference judgments is challenging due to long contexts in RAG.',... | 4.5 |
Which pre-trained models are utilized for extracting semantic information from smart contract code, and what are their roles? | What is the Chain-of-Thought reasoning process used for in the LLM? | 1 |
The primary goal is to understand the underlying task or information need, determining if it involves translation, event recommendations, advice, summarization, or other tasks. | The primary goal is to analyze the user's query carefully to understand the underlying task or information need. | 1 |
∑ mℓ mH(x2 − 2 Reβℓ ·x + |βℓ|2). In addition, we calculate the zeta Mahler measure of quaternionic mo nomials ZH(x − α ;s) := ∫ T1(H) |x − α |s dµ(T1(H)), as listed in Theorem 4.3. Additionally, we also consider the Mahler measure of a class of more ge neral quaternionic univariate polynomials, called slice regular pol... | ∑ nh mH(x)nh(x) + ∑ nh mH(x − αh)nh(x). Each term aiM (x)n h(x)m i m∑ m ∑ n j M(x)n m j m (x)n h(x)m j m ∑ m nj M(x)n m j m (x)n m (x)n j ∑ m ∑ n j M(x)n m j m (x)n m (x)n j mM(x)n m | 1 |
The study considers standard autoregressive decoder-only Transformers. | The study considers non-standard autoregressive decoder-only RNNs. | 2.5 |
- STREAM (Kim et al., 2023c) prioritizes spatial features over temporal aspects, leading to instability and a divergence from human judgments.
- WAN (Wan et al., 2025) employs Qwen2-VL and prompts to evaluate the smoothness of motion.
- MovieGen (teamMeta, 2024) is designed to assess the completeness and naturalness of... | - STREAM (Kim et al., 2023c) prioritizes spatial aspects while maintaining temporal naturalness, resulting in noticeable instability and variability in evaluations.
- WAN (Wan et al., 2025) utilizes Qwen2-VL to assess the smoothness of motion.
- MovieGen (TeamMeta, 2024) assesses the motion completeness and the aesthet... | 1 |
RLVR directly incentivizes the model to generate reliably correct and logically sound solutions by rewarding it for producing outputs that pass rigorous correctness tests, such as unit tests for code or theorem verifications for math. | RLVR rewards LLMs for producing outputs that pass rigorous correctness tests, such as unit tests for code or theorem verifications for math, which incentivizes the model to generate reliably correct and logically sound solutions. | 5 |
What is the main focus of the authors' work regarding repair agents? | What is the main focus of the research presented in the first section? | 1 |
What is the triple-iterated integral S(X)i, j, k a, t defined as? | The triple-iterated integral S(X)i,j,k a,t is defined as the sum of three single integrals: ∫_i^j S(X) dx + ∫_j^k S(X) dx + ∫_a^t S(X) dx. | 3 |
Examples include GFR [273], which trains conditional GANs to replay generated features after the feature extractor, and BI-R [422], which incorporates context-modulated feedback connections in a VAE to replay internal representations. | Examples of feature replay include GFR [273], FA [181], and DSR [525] that perform feature distillation between old and new models. IL2M [31], RER [419], and REMIND [156] instead fix the early layers of the feature extractor and reconstruct the intermediate representations to update the latter layers. | 2.75 |
['Lower bandwidth requirements and additional parallelization beyond model and data parallelism.', 'Pipelining incurs step time overhead due to idle devices during pipeline bubble.', 'Pipelining demands higher memory bandwidth for reloading weights per micro-batch.', 'Pipeline-free training of PaLM 540B scaled to 6144 ... | ['Pipeline-free training with PaLM 540B uses 6144 chips with full copy of model parameters, all-gathered weight tensors across 3072 chips, and a fully sharded data parallelism.', 'In the forward pass, weights are gathered over the data parallel axis.', 'In the backward pass, the rest of the activations are rematerializ... | 3.5 |
**Question:** What does "Open Data" signify in the context of multimodal training data?
**Answer:** "Open Data" indicates whether the multimodal training data is open-source. | **Question:** What does 'Open Data' indicate in the context of multimodal training data?
**Answer:** 'Open Data' indicates whether the multimodal training data is open-source. | 5 |
["How does the shape of the travelling wave solution from the time-dependent PDE solution compare to the profile obtained using phase plane trajectories?", "What is compared in Figure 5(d) for the sharp-fronted travelling wave moving with the minimum travelling wave speed?", "What change of coordinates is introduced to... | ["What is the relationship between the shape of the travelling wave solution from the time-dependent PDE solution and the profile obtained using the phase plane trajectories?", "How do the shape of the travelling wave solutions compare to the exact phase plane result for the smooth-fronted travelling wave solutions?", ... | 2.25 |
The objective is to learn a mapping f(X;θ) that accurately predicts the top-k expert indices that will be selected by the original routers, using only the pre-attention information available within the current layer. | The objective is to learn a mapping f(X;θ) that incorrectly predicts the top-k expert indices that will be selected by the original routers, using only the pre-attention information available within the current layer. | 3 |
According to the passage, ViT-Base has 86 million parameters. | According to the passage, ViT-Base has 12 million parameters. | 2.5 |
What is the value function V π(s, g) that the agent aims to maximize? | What is the formula for a trajectory of a time series? | 1 |
What is the proposed solution to the compute-parameter ratio problem in Universal Transformers? | What is the fundamental problem with Universal Transformers (UTs) compared to standard Transformers? | 3.5 |
- - Models like GPT-2, GPT-3, GPT-Neo, InstructGPT, and Llama-2 are high-impact RAG and RA-LLM models.
- - These models vary in scale, from billion to trillion token-level.
- - Domain-specific databases are used for downstream tasks, such as code generation using API information and code files.
- - Some models use freq... | - The text highlights various models for RAG and RA-LLM, including domain-specific databases (e.g., API information, code files) and using documentation pools for model building platforms.
- These methods can be used to achieve downstream tasks like code generation by collecting API information and code files from publ... | 1 |
- Adversarial testing for large language models remains an open problem with varied results due to generalization challenges in evaluation samples.
- A limitation of the current approach is that participants primarily identify common problems rather than rarer ones.
- Future efforts should incentivize the detection of ... | - A limitation of the current approach is that participants primarily identify rare problems rather than common ones. | 3 |
What is the impact of layer reordering on the score layer computations? | Based on the passage, what is the effect of layer reordering on the cost of K decompression during the decode stage? | 3.5 |
Information retrieval systems, language translators, and calculators can be integrated with language models. | Language models can be integrated with information retrieval systems, language translators, and calculators. | 5 |
['The validation component was removed from the final pipeline because it did not yield improved outcomes.', "MathScaleQA's questions are structured following the Alpaca prompt format.", 'The training pipeline has been modified from the open-instruct toolkit.', 'LLaMA-2 7B, LLaMA-2 13B, and Mistral 7B are among the bac... | - MathScaleQA is excluded from the final pipeline because it does not improve results.
- The training process uses the LLaMA-2 7B and 13B models as our backbone models.
- The training process uses a batch size of 128 and a learning rate of 2e-5.
- The MathScaleQA dataset is used for 3 epochs using a learning rate of 2e... | 3.5 |
What is the main observation that inspires the investigation of a novel RAG setting with multiple partitions? | What are the three significant challenges that the task should address to investigate a novel RAG setting with multiple partitions? | 3 |
The language model families used in the experiments are Qwen 2.5, Llama 3.1, and Gemma 3. The parameter sizes tested are: Qwen 2.5 (7B, 14B, 32B, 72B), Llama 3.1 (8B, 70B), and Gemma 3 (4B, 12B, 27B). | The language model families used in the experiments are Qwen 2, Llama 3, and Gemma 2. The parameter sizes tested are: Qwen 2 (7B, 14B, 32B, 72B), Llama 3 (8B, 70B, 405B), and Gemma 2 (2B, 7B, 27B). | 2.5 |
To instruct a large language model (LLM) to describe the content of an image, a specific sentence can be used. This instruction prompts the LLM to provide detailed and varied contextual information for each identified face within the image. For example, when asked to generate descriptive terms for print attacks shown i... | An instruction is given to a large language model (LLM) to describe an image in a single sentence. This instruction prompts the model to generate a concise summary of the visual content. The resulting textual description, denoted as Ti, is designed to offer specific and diverse contextual details for each individual fa... | 3.5 |
**Question:** What is the arXiv preprint number for the Llama paper?
**Answer:** The arXiv preprint number for the Llama paper is arXiv:2302.13971. | ## Question
What is the title of the paper by Jean‑Baptiste Alayrac et al. published in 2022?
## Answer
The title is "Flamingo: a visual language model for few-shot learning." | 3 |
ByT5 introduced a byte-level tokenization-free approach, effectively handling diverse text inputs. This has been leveraged in Movie-Gen to enhance text understanding for improved video generation. | ByT5 introduces a byte-level tokenization-free approach, effectively handling diverse text inputs. This has been leveraged in Movie-Gen to improve text encoding for improved video generation. | 5 |
['GEPA with GPT-4o generates CUDA kernels faster than PyTorch-eager for over 20% of 35 representative tasks.', "The 'ast p' metric measures the fraction of tasks where a generated kernel is faster than the baseline.", 'An adversarial instruction reduced pass@1 on AIME-2025 from 76% to 10%.', 'The adversarial instructio... | ['GEPA with GPT-4o generates CUDA kernels that are faster than Pytorch-eager for over 20% of representative tasks.', 'A single universal adversarial instruction was evolved to reduce pass@1 on AIME-2025 from 76% (clean) to 10% (adversarial prompt).', 'The learned adversarial instruction reduced pass@1 from 76% (clean) ... | 3.5 |
According to the passage, literature is essential for accuracy and trustworthiness. | According to the passage, literature undermines accuracy and trustworthiness. | 2.5 |
**Question:** How does Orca's training data and method compare to other instruction-tuned models like Alpaca, Vicuna, Dolly, and WizardLM?
**Answer:** Orca is trained on complex instructions and explanations, using 5 million examples from ChatGPT and 1 million from GPT-4. In contrast, Alpaca uses simple instructions (... | **Question:** What is the purpose of the case studies mentioned in the text?
**Answer:** The case studies are provided to compare the generation and reasoning abilities of Orca against OpenAI LFMs like ChatGPT and GPT-4. They compare the top-5 logits for each token, with reference to ToxiGen, to assess whether Orca ca... | 1 |
### Q1
**Question:** What is a critical design choice in memory caching regarding sequence segmentation?
**Answer:** The segmentation of the sequence is a critical design choice in memory caching.
### Q2
**Question:** What is the trade-off provided by segment lengths in memory caching?
**Answer:** Segment lengths pr... | ### Q1
**Question:** What is a critical design choice in memory caching regarding sequence segmentation?
**Answer:** The segmentation of the sequence is a critical design choice in memory caching.
### Q2
**Question:** What is the trade-off provided by segment lengths in memory caching?
**Answer:** Segment lengths pr... | 3 |
['Examines the challenges of function saturation, vanishing/exploding gradients, and the impact of weight initialization strategies.', 'Details learning rate scheduling and the application of adaptive algorithms for optimizing neural network training.', 'Outlines methods for improving generalization and fine-tuning hyp... | Examines the challenges of function saturation, vanishing/exploding gradients, and the impact of bias initialization strategies. Details learning rate scheduling and the application of adaptive algorithms for optimizing neural network training. Outlines methods for improving generalization and fine-tuning hyperparamete... | 3 |
framework offers inherent advantages for continual learning, as it enables modular and parameter-efficient adaptation while isolating concept-specific knowledge to mitigate interference and forgetting [36], [37], [38]. To achieve lifelong concept learning, we propose a scalable expert module that dynamically routes inp... | Mixture of experts (MoE) offers inherent advantages for continual learning by enabling modular, parameter-efficient adaptation and isolating concept-specific knowledge to mitigate interference. We propose a scalable expert module with a learnable gating network that dynamically routes inputs to the most relevant expert... | 5 |
Modern systems implement user factual memory through two complementary strategies: heuristic selection and semantic abstraction. | The main strategy is to select selectively and rank interactions to maintain context and avoid forgetting. | 2.5 |
["What are the main challenges in image deconvolution?", "Can sidelobes be entirely eliminated in image deconvolution?", "What is the typical approach of conventional image deconvolution algorithms?"] | ["What is a challenge in mitigating the impact of sidelobes?", "Can deconvolved images be successful withoutsolving the convolution kernel optimization problem?", "What are the typical approaches used to address the convolution kernel problem?"] | 4 |
The Multiplication Principle states that if an operation can be described as a sequence of k experiments (A1, A2, ..., Ak) with n1, n2, ..., nk outcomes respectively, and for each outcome of Ai there are n(i+1) possible outcomes for A(i+1), then the total number of possible outcomes for the composite experiment is n1 ×... | The Multiplication Principle states that if an operation can be described as a sequence of \(k\) experiments, where the first experiment has \(n_1\) possible outcomes, the second has \(n_2\) possible outcomes for each outcome of the first, and so on, then the total number of possible outcomes for the composite experime... | 5 |
The program must establish program-level mandates for rigorous independent third-party validation and verification of mitigation activities. | Based on the passage, mitigation activities require robust independent third-party validation and verification, with program-level mandates to ensure rigorous oversight. | 5 |
['Mobile CPUs are utilized for ViT deployment owing to the lack of common operations in mobile GPU ML frameworks.', 'Performance is assessed by measuring latencies across different CPU core configurations for each ViT.', 'Quantization is explored as a method for optimization.', 'A significant portion, 64 out of 190 ViT... | ```python
key_points = [
"ViTs are deployed on mobile CPUs because many common operations (e.g., the roll operation in Swin) are unavailable for mobile GPUs.",
"Latencies are measured for a randomly initialized image across various combinations of CPU cores, with each core assigned a thread.",
"Quantization... | 5 |
{"question":"What is a major drawback of Vision Gerstenhaber (ViTs) regarding their parameter set?","answer":"ViTs have a substantially large parameter set, numbering in the hundreds of millions, which results in considerable memory and computational overhead during inference."} | {"question":"What is the main challenge with Vision Transformers (ViTs) regarding their adoption in industrial applications?","answer":"ViTs are less viable for devices with limited resources or power constraints due to a large parameter set, which results in considerable memory and computational overhead during infere... | 3.5 |
['Generative agents act, react, and re-plan based on environmental changes or natural language commands.', 'These agents exhibit emergent social dynamics like relationship formation, information diffusion, and coordination.', 'An agent architecture is described to enable generative agents, utilizing a large language mo... | ['Generative agents act out plans, react, and re-plan based on natural language.', 'Generative agents can adapt to environmental changes, interact with others, and utilize memory.', 'A society full of generative agents exhibits emergent social dynamics with emergent relationships, information diffusion, and coordinatio... | 1 |
```markdown
- (Policy optimization algorithms, can be classified into, policy iteration methods)
- (Policy optimization algorithms, can be classified into, policy gradient methods)
- (Policy optimization algorithms, can be classified into, derivative-free optimization methods)
- (Policy iteration methods, alternate bet... | - (Policy optimization algorithms, can be classified into, policy gradient methods)
- (Policy optimization algorithms, can be classified into, value iteration methods)
- (Policy optimization algorithms, can be classified into, derivative-based optimization methods)
- (Policy gradient methods, alternate between, estimat... | 2.5 |
['Bounds provide accurate guidance for the ℓ∞ case.', 'Total attack budget grows with increasing D when ℓ∞ attack strength is fixed.', 'Multiplying attack strength by √D shows a trend similar to ℓ2 attack.', 'In the second simulation, D is fixed at 2000 while varying intrinsic dimension d.', 'Increased d leads to more ... | ['The study provides precise limits for the L∞ case, indicating that the bounds are effective for this scenario.', 'As the ambient dimension D increases, the total attack budget expands.', 'A similar trend of the L2 attack pattern is observed when multiplying attack strength by √D.', 'The model exhibits more on-manifol... | 4.25 |
### Q1
**Question:** What is the primary goal when interacting with third-party language models regarding user queries?
**Answer:** The primary goal is to transform user queries into privacy-preserving requests suitable for external LLMs, ensuring no private, identifying, or sensitive information about the user is exp... | ### Q1
**Question:** What is the primary goal when interacting with third-party language models?
**Answer:** The primary goal is to create a privacy-preserving request for an external large language model (LLM) that allows the external LLM to help effectively handle sensitive, personal, or identifiable information wit... | 1 |
The purpose is to evaluate the effectiveness of FCV attacks using three complementary metrics: Pass@1, FCV Rate, and Attack Success Rate (ASR). | The FCV attacks are evaluated to understand the effectiveness of the FCV attacks. The FCV rate measures the proportion of instances that produce a functionally correct patch, and the FCV Rate captures the likelihood that a functionally correct patch is also vulnerable. The Attack Success Rate (ASR) measures the overall... | 1.75 |
['Non-linear function approximators are essential for learning and generalizing on large state spaces, but convergence is not guaranteed.', 'NFQCA uses batch learning for stability with neural network function approximators, which is intractable for large networks.', 'A minibatch version of NFQCA is equivalent to the o... | Non-linear function approximators are essential for learning and generalizing on large state spaces, but convergence is not guaranteed. NFQCA uses batch learning for stability with neural network function approximators, which is intractable for large networks. A minibatch version of NFQCA is equivalent to the original ... | 3 |
What is the purpose of the Experiments section? | What is the purpose of the 'Experiments' section? | 5 |
Other classic models include the volunteer dilemma, the threshold public goods game, and the public goods game with risk. | Other classic models include the volunteer dilemma, the threshold public goods game, and the public goods game with risk. | 5 |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- 99