<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.>

Dynamic Cheatsheet (DC) tackles LM inference limitations by providing a persistent, evolving memory. This lightweight framework allows black-box language models to store and reuse insights across queries without parameter modification. DC enhances LM problem-solving on-the-fly, mimicking human cumulative learning and reducing routine errors, thereby boosting performance on diverse tasks.

How It Works

DC equips LMs with a growing, self-curated knowledge base accessible during inference. It focuses on concise, transferable snippets, enabling compatibility with any LM without parameter access and offering zero-shot improvements. Two variants exist: DC-Cumulative builds a single memory across queries, while DC-Retrieval & Synthesis uses similarity retrieval to generate query-specific cheatsheets, suitable for large-scale applications. This experience-driven learning bridges isolated inference with cumulative knowledge.

Quick Start & Requirements

Usage involves importing LanguageModel from dynamic_cheatsheet.language_model and initializing it with a model name (supporting OpenAI, Anthropic, DeepSeek, Llama, Gemini, etc.). Custom prompts can be defined or loaded. The advanced_generate method produces results and an updated cheatsheet. An ExampleUsage.ipynb notebook demonstrates capabilities, and run_benchmark.py facilitates result reproduction with configurable tasks, models, and output paths.

Highlighted Details

• Demonstrates significant performance gains: Claude 3.5 Sonnet's AIME math accuracy more than doubled; GPT-4o's Game of 24 success rate rose from ~10% to 99%; near-perfect accuracy on arithmetic equation balancing; 9% improvement on GPQA-Diamond and 8% on MMLU-Pro for knowledge-intensive tasks.
• Key features include: Persistent Memory, Self-Curated Storage, Black-Box Compatibility, Zero-Shot Learning, and Experience-Driven Learning.

Maintenance & Community

The README provides no details on maintainers, community channels, sponsorships, or a public roadmap.

Licensing & Compatibility

The provided README omits crucial software license information, hindering assessment for commercial use or derivative works.

Limitations & Caveats

No explicit limitations are listed. Effectiveness may vary based on the underlying LM. Performance claims are benchmark-specific. The absence of license details is a significant adoption blocker.