AINEWS
IBM Research's ICML 2026 paper (a top machine learning research conference) introduces PDL (Prompt Declaration Language), a layer that gives each step of a chained AI workflow an explicit uncertainty estimate, so the final answer carries its own
A chained AI workflow confidently returns a wrong answer. Not because any one model call was bad, but because each step's uncertainty compounded into a sure-sounding result. That compounding is the structural problem a new ICML 2026 paper from IBM Research sets out to fix.
In practice, an "AI workflow" means stitching several large language model calls together with code, search, and tool use: a planner that calls a search tool, a summarizer that calls the planner, a verifier that calls the summarizer. The single-call accuracy race is one thing; when a product chains those calls, the wrong answer from step one compounds through the chain and inflates the final output's confidence. The result is a workflow that sounds more certain than its evidence supports, the failure mode most likely to reach a paying user.
The paper, by IBM Research engineers Louis Mandel, Guillaume Baudart, Mandana Vaziri, and Martin Hirzel, introduces PDL, or Prompt Declaration Language, a probabilistic-programming layer for LLM-based flows. The mechanism is older than LLMs: a probabilistic program is code that, instead of returning a single value, returns a distribution of possible values, with an explicit way to update that distribution as new evidence arrives. PDL applies that idea to a chain of model calls. Instead of asking the model for a single best answer, the developer declares the flow in YAML, and the language attaches an explicit uncertainty estimate to each call and propagates that estimate through the rest of the chain. A final answer can then come with a calibrated confidence, and downstream steps can branch on that confidence rather than on a brittle string match.
The mechanism matters because it reframes the AI trust problem. Most fixes target a single model's accuracy: bigger model, better fine-tune, sharper prompt. PDL targets the chain. The paper argues that for products people rely on, the bottleneck is not one bad model call but the way a chain amplifies small uncertainties into a confident mistake. Treating the flow as a first-class probabilistic object, the paper claims, lets developers experiment with different inference-scaling techniques (asking the model to think longer, sampling multiple candidates and voting, running a critic) without adding a single line of code beyond the flow's logic.
IBM/prompt-declaration-language is a declarative YAML-based language with a pip-installable package, a command-line interface, a GUI, and an AutoPDL prompt optimizer. The paper's case study is a theorem-proving agent built for the Rocq theorem prover, a formal proof assistant more familiar to type theorists than to most product teams. The deployment target, however, is the broader class of multi-step LLM workflows that any modern AI product already runs.
The limit: this is a research direction at a top venue, not a published standard and not an industry adoption signal. The IBM repo and the ICML 2026 acceptance give it concrete, time-stamped credibility this week, but no shipping product yet calls itself PDL-compliant. Independent validation of the PGM-framing claim, that flow-level uncertainty propagation actually improves reliability, still has to come from outside IBM.
Two things would move PDL from a credible research direction to a layer the rest of the stack has to learn: AutoPDL, the project's prompt optimizer, landing in the language's reference distribution before the next ICML cycle, or any cloud LLM platform exposing a comparable chain-level confidence API. Neither has happened yet.