agentic-systems · 2026-06-06 · Tier 2

AdaPlanBench: Evaluating Adaptive Planning under World and User Constraints

AdaPlanBench: Evaluating Adaptive Planning under World and User Constraints

TL;DR. Real planning problems do not hand you all the constraints up front; they leak out as you go ("actually it can't go in the dishwasher", "I need this before 6pm"). AdaPlanBench is a dynamic, interactive benchmark that reveals world and user constraints only when the agent proposes a plan that violates them, forcing iterative re-planning under accumulating feedback. Built on 307 household tasks. The best of ten leading LLMs reaches only 67.75%, performance degrades as constraints pile up, and user constraints are harder than world constraints, with failures tracing to weak physical grounding.

Source: HuggingFace Daily Papers · arxiv 2606.05622

What it is

A testbed for adaptive planning: agents interact with an environment over multiple turns, propose plans, and discover hidden constraints reactively when a plan trips one. They must infer and track the accumulating constraint set while re-planning effectively. A scalable construction pipeline augments each of 307 household tasks with dual (world + user) constraints.

Key results

  • Best model: 67.75% accuracy; the task is hard for every frontier model tested.
  • Performance degrades as more constraints accumulate, the realistic long-horizon regime.
  • User constraints pose a larger challenge than world constraints; failures often stem from weaker physical grounding and reduced re-planning effectiveness.

How it relates to prior wiki knowledge

AdaPlanBench extends the wiki's benchmark-skepticism / trajectory-grounded-evaluation thread (see agent-benchmarks.md): like GTA-2 (04-20, top models 14.39% on open-ended workflows) and AgentLens (05-14, pass rate overstates real competence), it shows agents look much weaker once the evaluation grades the adaptive process rather than a single static plan. It is the planning-side analogue of today's SABER (grade the final environment state, not the turn): both move the unit of evaluation from response to interaction-with-consequences.

Gaps

Household tasks are a narrow, simulated slice; whether the progressive-disclosure protocol predicts performance on real interactive workflows (software, ops) is untested. The constraint-construction pipeline is automated, so constraint realism and difficulty calibration depend on its quality, which the abstract does not characterize.

Related pages

Raw source: raw/huggingface/2026-06-06-adaplanbench-evaluating-adaptive-planning-in-large-language.md