Yifan Zhang
PulseAugur coverage of Yifan Zhang — every cluster mentioning Yifan Zhang across labs, papers, and developer communities, ranked by signal.
1 day(s) with sentiment data
-
FlashSampling integrates exact sampling into LM head for faster token generation
Researchers have developed FlashSampling, a novel method for exact sampling in large-vocabulary language models that integrates sampling directly into the language model head's matrix multiplication. This approach avoid…
-
Deep Reinforcement Learning applied to Bus Fleets and High-Speed Flows
Two recent arXiv papers explore the application of deep reinforcement learning (DRL) in complex control scenarios. The first paper details a single-agent RL framework for bus fleet control, which reformulates the proble…
-
Princeton researcher proposes Recurrent Looped Transformer with unbounded temporal depth
A Princeton researcher, Yifan Zhang, has proposed a novel architecture called the Recurrent Looped Transformer (RLT). This design aims to enhance the temporal depth of decoder-only language models by carrying the comple…
-
New Comparables XAI method enhances AI decision explanations
Researchers have introduced Comparables XAI, a novel method for generating faithful example-based explanations of AI decisions. This approach draws inspiration from real estate valuation, where property values are estim…
-
Deep Delta Learning introduces targeted residual updates for Transformers
Researchers have introduced Deep Delta Learning (DDL), a novel structured residual update for Transformer models. DDL enables targeted edits to the residual state by explicitly parameterizing reading, comparison, and re…
-
SynthFix framework enhances automated code vulnerability repair
Researchers have developed SynthFix, a novel neuro-symbolic framework designed to improve the automated repair of code vulnerabilities. This system integrates supervised fine-tuning with compiler-informed feedback, allo…
-
New EyeMulator method enhances Code Language Models by mimicking human visual attention
Researchers have developed EyeMulator, a novel method designed to enhance the performance of Code Language Models (CodeLLMs). This technique injects human visual attention priors into the fine-tuning process without req…
-
Discrete Transformer extracts algorithms from model weights
Researchers have developed a "Discrete Transformer" architecture designed to extract interpretable algorithms from trained models. This approach addresses the challenge of representation entanglement in standard Transfo…
-
New benchmark tests LLMs for organic chemistry reasoning capabilities
Researchers have introduced oMeBench, a new benchmark designed to evaluate the organic mechanism reasoning capabilities of large language models. The benchmark includes over 10,000 annotated mechanistic steps and a dyna…