The idea in one paragraph.
Stop hand-writing prompts. Express an LM pipeline as declarative modules with typed signatures, then compile it against a metric, letting an optimiser choose the prompts.
THE MACHINERY
Signatures
Declare what a step does as typed input and output fields, not as prose.
Modules
Composable, parameterised strategies (Predict, ChainOfThought, ReAct) over a signature.
Teleprompters / optimisers
Compile a program by bootstrapping demonstrations and searching prompts against a metric.
Bootstrapped demonstrations
Generate few-shot examples from the program's own successful traces.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Within minutes of compiling, a few lines of DSPy let GPT-3.5 and llama2-13b-chat self-bootstrap pipelines that beat standard few-shot prompting by over 25% and 65% respectively.
Inspect a source excerpt
a few lines of DSPy allow GPT-3.5 and llama2-13b-chat to self-bootstrap pipelines that outperform standard few-shot prompting (generally by over…Read the passage in context on arXiv ↗
They also beat pipelines with expert-created demonstrations by 5-46% and 16-40% respectively.
Programs compiled to small open models (770M T5, llama2-13b-chat) are competitive with expert-written prompt chains for GPT-3.5.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Requires a metric and a train set - unavailable for most open-ended agent work.
- Optimises the prompt only; tools, memory, and control flow stay fixed.
Why it belongs in this atlas.
The first rung where the agent writes its own instructions. The talk's framing is exact: this gives you CRUD over the system prompt. Note the ceiling - only the prompt moves.