PPDL turns LLM flows into probabilistic programs for trustworthy AI
New ICML 2026 language propagates uncertainty across multi-call LLM pipelines without extra code.
Louis Mandel, Guillaume Baudart, Mandana Vaziri, and Martin Hirzel introduce PPDL, a probabilistic language for programming LLM-based flows, in a paper published at ICML 2026. PPDL lets developers quantify and propagate uncertainty through flows involving multiple calls to LLMs and other tools, and experiment with different inference scaling techniques without adding code beyond the flow's logic. An experimental study demonstrates this capability, and a case study builds a theorem-proving agent for the Rocq theorem prover.
- PPDL is a probabilistic programming language that models LLM flows as probability distributions over outputs
- It quantifies and propagates uncertainty across chained LLM and tool calls, enabling confidence-aware results
- Inference scaling experiments can be run without changing flow code; case study built a Rocq theorem-proving agent
- Published at ICML 2026 with authors from IBM Research (Mandel, Baudart, Vaziri, Hirzel)
Why It Matters
PPDL gives developers a principled way to measure and control uncertainty in LLM pipelines, unlocking safer deployment for critical applications.