beam search
PulseAugur coverage of beam search — every cluster mentioning beam search across labs, papers, and developer communities, ranked by signal.
-
Research: Larger models outperform increased inference compute for text-to-SQL
A new research paper explores the trade-offs between model size and inference compute for grammar-constrained text-to-SQL tasks. The study found that increasing model size generally yields better accuracy than increasin…
-
New method increases machine translation benchmark difficulty
Researchers have developed a new method called Adversarial Translation Optimization (ATO) to create more challenging machine translation benchmarks. By combining adversarial optimization with a differentiable difficulty…
-
New method increases machine translation benchmark difficulty
Researchers have developed a new method called Adversarial Translation Optimization (ATO) to create more challenging benchmarks for machine translation models. ATO uses gradients from a difficulty and fluency objective …
-
LLM decoding strategies: Greedy, Beam Search, Sampling, Top-K, and Top-P explained
Language models generate text by turning probability distributions into sequences of tokens, with different decoding strategies leading to varied outputs. Greedy decoding selects the most probable token at each step, wh…
-
KV-PRM paper introduces efficient reward modeling for multi-agent LLMs
Researchers have introduced KV-PRM, a novel method for improving the efficiency of Process Reward Models (PRMs) used in multi-agent systems. Unlike existing text-based PRMs that re-encode entire trajectories, KV-PRM dir…
-
NextCrystal framework uses LLMs for advanced crystal structure prediction
Researchers have developed NextCrystal, a novel generative framework for crystal structure prediction that leverages large language models and a diffusion backbone. This approach directly generates Wyckoff site patterns…
-
New framework automates compression of low-power AI vision models
Researchers have developed AQ4SViT, an automated framework designed to compress Spiking Vision Transformers (SViTs) for use in resource-constrained embedded AI systems. This new framework addresses the scalability issue…
-
ARC-AGI solver success predicted by structural grid descriptors
Researchers have developed a method using structural grid descriptors to predict the success of symbolic solvers on ARC-AGI tasks. Across numerous runs and distinct solver architectures, these descriptors, measured at 5…