A naukri.com initiative
ML News
Arxiv
1w
96
Image Credit: Arxiv
DashCLIP: Leveraging multimodal models for generating semantic embeddings for DoorDash
- Researchers introduce a joint training framework for product and user queries by aligning uni-modal and multi-modal encoders through contrastive learning on image-text data.
- The framework eliminates the reliance on engagement history and trains a query encoder using an LLM-curated relevance dataset.
- The generated embeddings demonstrate strong generalization capabilities and improve performance in applications such as product categorization and relevance prediction.
- The deployment of the framework shows a significant uplift in click-through rate and conversion rate for personalized ads recommendation.
Read Full Article
5 Likes
Arxiv
1w
365
Image Credit: Arxiv
SolRPDS: A Dataset for Analyzing Rug Pulls in Solana Decentralized Finance
- SolRPDS is the first public rug pull dataset derived from Solana's transactions.
- The dataset covers approximately four years of DeFi data (2021-2024) and consists of 62,895 suspicious liquidity pools.
- Preliminary analysis of the dataset reveals clear distinctions between legitimate and fraudulent liquidity pools.
- A total of 22,195 tokens in the dataset exhibit rug pull patterns during the examined period.
Read Full Article
21 Likes
Arxiv
1w
369
Image Credit: Arxiv
Can SGD Select Good Fishermen? Local Convergence under Self-Selection Biases and Beyond
- Researchers have revisited the problem of estimating multiple linear regressors with self-selection bias.
- They have introduced a new algorithm that resolves the main open question from a previous study on self selection.
- The algorithm reduces self-selection to a statistical problem called coarsening, allowing for significant improvements in running time.
- The findings have implications for linear regression and coarse Gaussian mean estimation.
Read Full Article
22 Likes
Arxiv
1w
222
Image Credit: Arxiv
PLM-eXplain: Divide and Conquer the Protein Embedding Space
- Protein language models (PLMs) have been successful in generating powerful sequence representations for computational biology.
- A new approach called PLM-eXplain (PLM-X) is introduced, which makes PLMs more interpretable and facilitates actionable insights.
- PLM-X factors PLM embeddings into two components: an interpretable subspace based on established biochemical features and a residual subspace preserving predictive power.
- The effectiveness of PLM-X is demonstrated through protein-level classification tasks, maintaining high performance while enabling biological interpretation.
Read Full Article
13 Likes
Arxiv
1w
226
Image Credit: Arxiv
GAAPO: Genetic Algorithmic Applied to Prompt Optimization
- Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, with their performance heavily dependent on the quality of input prompts.
- GAAPO (Genetic Algorithm Applied to Prompt Optimization) is a hybrid optimization framework that leverages genetic algorithm principles to evolve prompts through successive generations.
- GAAPO integrates multiple specialized prompt generation strategies within its evolutionary framework.
- Experimental analysis reveals insights into the tradeoff between population size and the number of generations, the effect of selection methods on stability, and the ability of different LLMs to automatically generate prompts.
Read Full Article
13 Likes
Arxiv
1w
345
Image Credit: Arxiv
R2E-Gym: Procedural Environments and Hybrid Verifiers for Scaling Open-Weights SWE Agents
- Improving open-source models on real-world SWE tasks (solving GITHUB issues) faces challenges in scalable curation of execution environments and optimal test-time compute scaling.
- AgentGym is introduced as the largest procedurally-curated executable gym environment for training real-world SWE-agents, with over 8.7K tasks.
- SYNGEN, a synthetic data curation recipe, is used to enable scalable curation of executable environments, leading to improved training performance.
- Hybrid Test-time Scaling is employed, showcasing the complementary strengths and limitations of execution-based and execution-free verifiers.
Read Full Article
20 Likes
Arxiv
1w
54
Image Credit: Arxiv
HypoEval: Hypothesis-Guided Evaluation for Natural Language Generation
- Large language models (LLMs) have demonstrated great potential for automating the evaluation of natural language generation.
- Previous frameworks of LLM-as-a-judge fall short in two ways: they either use zero-shot setting without consulting any human input, which leads to low alignment, or fine-tune LLMs on labeled data, which requires a non-trivial number of samples.
- In this paper, the authors propose HypoEval, a Hypothesis-guided Evaluation framework, which incorporates a checklist-like approach to combine LLM's assigned scores on each decomposed dimension to acquire overall scores.
- With only 30 human evaluations, HypoEval achieves state-of-the-art performance in alignment with both human rankings and human scores, outperforming previous methods.
Read Full Article
3 Likes
Arxiv
1w
368
Image Credit: Arxiv
Evolutionary algorithms meet self-supervised learning: a comprehensive survey
- The combination of Evolutionary Machine Learning and self-supervised learning has been a growing trend in recent years.
- Evolutionary Machine Learning can automate the design of machine learning algorithms and provide more reliable solutions.
- Self-supervised learning is effective for learning useful features when labeled data is limited.
- This survey provides an overview of the studies in this area, introduces Evolutionary Self-Supervised Learning as a new sub-area, and suggests directions for future research.
Read Full Article
21 Likes
Arxiv
1w
67
Image Credit: Arxiv
Reservoir Computing with a Single Oscillating Gas Bubble: Emphasizing the Chaotic Regime
- Researchers propose a reservoir computing system based on a single bubble trapped within a liquid.
- The system utilizes an external acoustic pressure wave to encode input information and excite complex nonlinear dynamics.
- The single-bubble reservoir computing system demonstrates the ability to forecast time series and perform classification tasks with high accuracy.
- A chaotic physical regime of bubble oscillation is found to be most effective for these computations.
Read Full Article
4 Likes
Arxiv
1w
214
Image Credit: Arxiv
Earth-like planet predictor: A machine learning approach
- A machine learning approach has been developed to predict which stars are most likely to host an Earth-like planet.
- The aim is to avoid blind searches and maximize the number of detections by focusing observation time on the most promising systems.
- The Random Forest model achieved precision scores of up to 0.99, correctly identifying systems with Earth-like planets in 99% of cases.
- 44 observed systems have been highlighted as having a high probability of hosting an Earth-like planet.
Read Full Article
12 Likes
Arxiv
1w
63
Image Credit: Arxiv
Better Decisions through the Right Causal World Model
- Reinforcement learning agents often exploit spurious correlations in training data, resulting in brittle behaviors that fail to generalize to new environments.
- The Causal Object-centric Model Extraction Tool (COMET) is an algorithm designed to learn interpretable causal world models (CWMs).
- COMET extracts object-centric state descriptions from observations and models object-centric transitions using symbolic regression.
- COMET constructs CWMs that align with the true causal structure of the environment, enabling better planning and decision-making in reinforcement learning.
Read Full Article
3 Likes
Arxiv
1w
306
Image Credit: Arxiv
Evaluating Parameter-Based Training Performance of Neural Networks and Variational Quantum Circuits
- Neural networks (NNs) have driven significant advances in machine learning, but often require large numbers of trainable parameters.
- Variational quantum circuits (VQCs) offer a promising alternative as they require fewer parameters and leverage quantum mechanics.
- A study evaluated NNs and VQCs on simple supervised and reinforcement learning tasks with different parameter sizes.
- Results showed that VQCs can match NNs in performance, despite longer training durations.
Read Full Article
18 Likes
Arxiv
1w
46
Image Credit: Arxiv
Objaverse++: Curated 3D Object Dataset with Quality Annotations
- Objaverse++ is a curated subset of Objaverse with detailed attribute annotations by human experts.
- 10,000 3D objects in Objaverse++ are manually annotated with attributes such as aesthetic quality scores, texture color classifications, multi-object composition flags, and transparency characteristics.
- A neural network trained on Objaverse++ achieved better performance in image-to-3D generation tasks compared to models trained on the larger Objaverse dataset.
- Approximately 500,000 curated 3D models are released as the enhanced dataset, offering a more efficient path to develop 3D generative models.
Read Full Article
2 Likes
Arxiv
1w
163
Image Credit: Arxiv
Learning to erase quantum states: thermodynamic implications of quantum learning theory
- Learning algorithms can acquire knowledge of quantum states to erase them at optimal energy cost.
- Learning can be made fully reversible and has no fundamental energy cost itself.
- The energy cost of erasing quantum states is related to their complexity, entanglement, and magic.
- Efficient work extraction can be achieved based on learning.
Read Full Article
9 Likes
Arxiv
1w
189
Image Credit: Arxiv
Throughput-Optimal Scheduling Algorithms for LLM Inference and AI Agents
- Optimizing systems for efficient Large Language Model (LLM) inference and AI agent workloads becomes critical as their demand is rapidly growing.
- A new study bridges the gap between queuing theory and LLM system communities to develop queuing fundamentals for LLM inference.
- The study proves that 'work-conserving' scheduling algorithms can achieve maximum throughput for individual requests and AI agent workloads.
- Evaluations of real-world systems show that Orca and Sarathi-serve are throughput-optimal, while FastTransformer and vanilla vLLM are not maximally stable and should be used with caution.
Read Full Article
11 Likes
For uninterrupted reading, download the app