DecryptPrompt

Aggregates peer-reviewed LLM research papers from arXiv, conferences, and preprint servers, organizing them into a hierarchical taxonomy covering 20+ research areas (RLHF, CoT, RAG, agents, alignment, etc.). Uses a curated folder structure with PDF storage and README-based indexing to enable semantic navigation across interconnected topics like chain-of-thought reasoning, instruction tuning, and multi-agent systems without requiring a database backend.

Maintains a curated collection of prompting methodologies including chain-of-thought (CoT), few-shot learning, zero-shot learning, in-context learning, and instruction tuning, with associated research papers and implementation patterns. Organizes prompting techniques into discrete categories with explanations of when and how to apply each approach, enabling practitioners to understand the theoretical foundations and empirical trade-offs between techniques.

Maintains a series of 51+ educational blog posts explaining LLM concepts, techniques, and research findings in accessible language. Covers topics from fundamentals (tokenization, attention mechanisms) to advanced techniques (RLHF, multi-agent systems), with explanations designed for practitioners and researchers new to specific areas. Blog posts serve as entry points to deeper research papers and provide conceptual foundations for understanding complex LLM methodologies.

Catalogs research on post-training techniques including SFT vs. RL trade-offs, test-time scaling, reasoning enhancement through inference-time computation, and optimization strategies for improving model performance after pre-training. Documents how different post-training approaches (supervised fine-tuning, reinforcement learning, constitutional AI) affect model capabilities and generalization, with papers on inference-time scaling that show how additional computation at inference time can improve reasoning quality.

Catalogs research on LLM agents including tool-augmented LLMs, agent planning and reasoning, multi-agent systems, and agent-environment interaction patterns. Documents how agents decompose tasks, select tools, handle failures, and coordinate with other agents, with references to foundational papers on ReAct, chain-of-thought agents, and tool-use frameworks that enable LLMs to interact with external APIs and knowledge sources.

Aggregates research on RAG systems, document retrieval methods, knowledge base augmentation, and table/chart understanding, documenting how LLMs can be enhanced with external knowledge sources. Covers retrieval strategies (dense retrieval, sparse retrieval, hybrid), knowledge base construction, and integration patterns that enable LLMs to ground responses in factual information and reduce hallucination through knowledge-augmented inference.

Catalogs research on alignment techniques including RLHF (Reinforcement Learning from Human Feedback), constitutional AI, preference modeling, self-critique mechanisms, and LLM critics. Documents the alignment pipeline from supervised fine-tuning (SFT) through reward modeling and RL training, with papers on how to make LLMs more helpful, harmless, and honest through preference optimization and principle-driven alignment approaches.

Aggregates research on chain-of-thought (CoT) prompting, implicit vs. explicit reasoning, test-time scaling, and reasoning enhancement techniques that enable LLMs to solve complex problems through step-by-step inference. Documents how CoT improves performance on reasoning tasks, the relationship between reasoning depth and accuracy, and techniques for eliciting and verifying intermediate reasoning steps.

Catalogs research on instruction tuning, supervised fine-tuning (SFT), and how to adapt pre-trained LLMs to follow instructions and perform specific tasks. Documents the relationship between instruction tuning and in-context learning, the role of instruction diversity in generalization, and techniques for constructing high-quality instruction datasets that improve model performance across diverse downstream tasks.

Maintains a curated index of open-source LLM models (LLaMA, Mistral, Qwen, etc.), inference frameworks (vLLM, TensorRT-LLM, etc.), fine-tuning tools (LoRA, QLoRA, etc.), and evaluation leaderboards (MMLU, HumanEval, etc.). Provides links to model repositories, framework documentation, and evaluation benchmarks, enabling practitioners to navigate the rapidly evolving open-source LLM ecosystem and make informed decisions about model selection and deployment.

Catalogs research on domain-specific LLM variants (biomedical, legal, code, etc.), domain adaptation techniques, and how to specialize pre-trained models for specific industries or knowledge domains. Documents approaches for incorporating domain knowledge through continued pre-training, instruction tuning on domain data, and retrieval-augmented generation with domain-specific knowledge bases.

Aggregates research on improving LLM reliability through hallucination reduction, fact verification, interpretable reasoning, refusal capabilities, and evaluation frameworks. Documents techniques for detecting and mitigating hallucinations (knowledge grounding, self-verification), making LLM reasoning more interpretable, and building guardrails that enable models to refuse unsafe or out-of-scope requests.