Planning a sequence of actions before acting

Claims

• observationsingle paper↑1↓0
  On 102 IPC Blocksworld instances with a 180-second budget, a classical planner (Fast Downward) solved 85%, a model planning directly solved 64%, and the same model choosing one action at a time against a PDDL simulator with observe, reset and retry solved 67%, a consistent but small three-point gain at about 5.7 times the token cost, so interactive simulation helps a little and classical search still wins where a formal model exists.
• observationsingle paper↑1↓0
  When asked to generate a full procedural execution graph for an unseen task (recipe) in a specific creator's style, few-shot LLMs assign ingredients to steps better than a statistical two-stage pipeline but predict the number of steps worse, and swapping in a larger LLM backbone does not close that step-count gap.
• observationsingle paper↑1↓0
  In hours-long ML development runs on Kaggle-style competitions, agent scaffolds each lock into one strategy policy — the single-loop CLI agent pivots far less than humans (9% vs 25% of transitions) and the evolutionary search agent far more (58%) with pivot gains cancelling out — and neither reopens an approach it had abandoned, even though it recovers scores after setbacks as well as top humans.
• observationsingle paper↑1↓0
  When a language model must judge whether a specific vehicle can perform a described mission, it is overconfident about partially-capable configurations — declaring the mission fully possible despite missing hardware — and supplying the required-capability list plus the vehicle's inferred capabilities from an ontology is what corrects that specific error, not the harder fully/impossible cases which the models already get mostly right.
• mechanismmechanism reasoning↑1↓0
  For LLM agents acting in partially observable environments, replacing the raw action-observation log in the prompt with an externally maintained Bayesian posterior over latent states makes the agent's action choice a function of the belief alone, removing surface-form sensitivity in which two histories with identical posteriors but different serialized text elicit different actions.
• mechanismsingle paper↑1↓0
  For LLM agents that must compose and then execute domain tool workflows, training separate adapters for the two roles splits the gains cleanly: supervised fine-tuning of the planner on reference tool sequences mainly improves workflow-order fidelity, while node-level RL on argument-match rewards mainly improves tool-argument grounding and final-answer quality, and combining both beats either on a 9B backbone.
• observationsingle paper↑1↓0
  For LLM-generated multi-day travel itineraries, high qualitative coherence scores (spatial, persona, meal, ordering) coexist with near-zero robustness to realistic disruption: across GPT-5, Qwen3, Mistral-7B and Phi-4-mini, transport-delay buffer scores stay under 12% and buffer-adequacy scores under ~41%, so coherence and static constraint satisfaction do not indicate that a plan survives empirical transit delays or crowd peaks.
• mechanismsingle paper↑1↓0
  For long-horizon embodied task planning where a symbolic model defines action preconditions and effects, correcting a violated plan by edits the action model itself implies (inserting the missing navigate or open) raises plan success more per LLM call than returning the violation to the LLM for replanning, and the gain persists for frontier models as well as compact ones.
• mechanismsingle paper↑1↓0
  For multi-modal robot motion planning stated in natural language, the hard part for an LLM is synthesizing the cross-modal coupling (meeting points, contact locations, shared intermediate states), not the single-modality formulation: giving an LLM fixed expert single-modal MIP modules and asking it only to predict the interface variables gives 100% single-modal success but 55.6-63.2% multi-modal, below a model fine-tuned with solver-verified rewards to write the joint formulation (80.9%).
• mechanismsingle paper↑1↓0
  For multi-step assembly task planning from an action history and a workstation image, prompting a multimodal LLM to emit a structured (JSON) state representation and propagate it through each predicted action — then truncating the plan at the predicted state closest to the inferred goal — raises action executability and lowers final-state error relative to predicting actions alone, at the cost of shorter plans (lower overlap with recorded human action sequences) and roughly eight times the latency.
• observationsingle paper↑1↓0
  Prompting with a Task-Method-Knowledge model, which spells out what to do, how, and why in a hierarchical structure, lifted o1 on the obfuscated Random Blocksworld set of PlanBench from about 32% to 97%, and o1-mini from about 9% to 27%, gains far beyond the roughly 1% that zero-shot chain-of-thought gave on PlanBench in earlier work; the smaller model did not benefit on the Mystery variant.
• observationsingle paper↑1↓0
  Translating natural-language planning problems into PDDL can be learned from solver feedback alone, with one model acting as actor, judge and editor: on PlanBench average success rose from 35.5% (LLM+P) to 70.8%, with 66.3% of successes judged faithful to the described task and semantic drift, where a solvable but wrong problem is produced, cut to 6.4%.
• mechanismsingle paper↑1↓0
  In VLM-driven human-aware robot task planning over scene graphs, the bottleneck is inferring and grounding the symbolic goal, not the planning: once both human and robot goals correctly cover the ground truth, a formal multi-agent planner produces a valid, conflict-free joint plan in 99.6% of trials, while end-to-end success with VLM-inferred human goals is 69.7%.

Techniques

• Actor-judge-editor RL on solver feedback for NL-to-PDDLtraining
  Have the model propose a PDDL formalization, judge it with a solver-calibrated score, and edit it under bounded diagnostic feedback, training all three roles with reinforcement learning from solver outcomes rather than annotations.
• Distil human planning practice into a re-injected planning pprompting
  Adds a ~1000-token prompt block at run start and every 30 minutes that forbids mono-loop moves (single hold-out, repeated tweaks) and prescribes human practices (K-fold from version one, early ensembling, cached out-of-fold predictions, per
• Generate the action skeleton from creator transition priors,process
  Decode the step sequence with beam search over a creator's learned bigram/positional action priors and a deterministically predicted step count, then route ingredients to those steps by a learned ingredient-action prior instead of an LLM.
• Ground mission feasibility in an ontology of inferred capabiretrieval
  Store vehicle components, mission types and required capabilities in an ontology with SWRL rules that infer capabilities from hardware/software, then put both the mission's required capabilities and the vehicle's inferred capabilities into
• Predict human intention as a symbolic goal, then plan jointlprocess
  A VLM infers the human's latent intention from third-person images plus a scene graph as symbolic goal predicates rather than an action sequence, and those goals are grounded with the robot's goal into a multi-agent PDDL problem solved by a
• Propagate a structured state and truncate at closest-to-goalprompting
  Prompts a multimodal LLM to estimate the workstation state as a JSON object, propagate it forward through each predicted action, compute a numeric distance to the inferred goal state, and cut the action plan at the state with the lowest dis
• Repair plans from the world model, ask the LLM only for semaprocess
  Roll an LLM-generated action plan forward on an explicit graph world model of object relations, preconditions, effects and safety constraints, apply edits whose corrections follow from the action semantics, and return only unresolved semant
• Route routine calls to a smaller modelprocess
  Send the short, structured, routine calls in an agent pipeline to a small cheap model and reserve a frontier model for the steps that need deep planning or sustained constraint tracking.
• Score itineraries for buffer adequacy against empirical delaprocess
  Evaluate a generated plan by whether each activity gap and transport segment allocates buffer inside a risk-profile-scaled range derived from historical delay distributions and hourly crowd density, rather than by binary constraint satisfac
• Show the LLM a Bayesian belief state, not the historyarchitecture
  An external module runs a two-step Bayes filter over a specified POMDP and passes only the normalized posterior over latent states to the LLM, which selects an action from that posterior with no access to the action-observation log.
• Task-Method-Knowledge promptingprompting
  Give the model an explicit hierarchical model of the task: the goals, the methods that achieve each, and the knowledge of why each step is taken, drawn from the TMK framework in educational science, before asking it to plan.
• Train an LLM to write solver-checked MIP programs from task training
  Supervised fine-tuning on single-modal MIP program demonstrations followed by GRPO with rewards for output format, discretization-API coverage, successful solver execution, and solver-verified trajectory correctness.
• Train planner and executor as separate role adapterstraining
  Fit one LoRA adapter by sequence-level SFT on reference tool sequences for workflow planning and a second by node-level GRPO with argument-match rewards for runtime argument grounding, so workflow errors and argument errors get separate cre

Sources

Every paper, post or observation behind this capability. Ones cited by a claim show the claim and its stance; ones filed directly are about the capability but have no claim drafted from them yet. Open one to add support or contest it.

• GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planning2026-09-16
  ↑ supports a claim here
• M$^3$P-R1: Reinforcement Learning for Large Language Model Guided Multi-Modal Motion Planning via MIP Code Generation2026-09-16
  ↑ supports a claim here
• HINT-Plan: Human Intention-Aware Robot Task Planning in Context-Rich Environments using Vision Language Models2026-09-15
  ↑ supports a claim here
• UDAV: Uncertainty-Driven Adaptive VLM Waypoint Planner2026-09-14
  about this capability · no claim drafted yet
• Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents2026-09-11
  ↑ supports a claim here
• Bridging Thought and Action: Taming Long-Horizon Instability in Open-Source LLM Agents with a MetaTool-Enhanced ROS Framework2026-09-11
  about this capability · no claim drafted yet
• MAPLE: Memory-Augmented Planning with Language and Evolution2026-09-10
  about this capability · no claim drafted yet
• Belief-State Engine: Augmenting LLMs for Principled Planning Under Partial Observability2026-09-09
  ↑ supports a claim here
• UTP-Bench: Uncertainty-aware Travel Planning Benchmark2026-09-02
  ↑ supports a claim here
• A Cognitive Architecture for Shared Autonomy in AUV Operations2026-08-29
  ↑ supports a claim here
• STEP: State-Aware Task Estimation and Planning with Multi-Modal LLMs for Human-Robot Collaboration2026-08-27
  ↑ supports a claim here
• TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development2026-08-26
  ↑ supports a claim here
• Do Recipes Have Personas? Characterizing and Generating Creator Style in Attributed Procedural Graphs2026-08-25
  ↑ supports a claim here
• From Solver Feedback to Faithful Plans: Multi-Role Reinforcement Learning for Symbolic Planning2026-08-22
  ↑ supports a claim here
• Agentic LLM Planning via Step-Wise PDDL Simulation: An Empirical Characterisation2026-03-06
  ↑ supports a claim here
• Knowledge Model Prompting Increases LLM Performance on Planning Tasks2026-02-03
  ↑ supports a claim here