long-horizon-embodied-task-planning-where-symbolic-model-defines-action CLM-0430mechanismsingle paperReviewed by AI
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.
Capability: Planning a sequence of actions before acting
Observed on
BEHAVIOR-1K symbolic action level (nine primitives, no geometric feasibility); 100 single-task and 500 multi-task instructions; Qwen3-4B/8B, GPT-5.6 Sol, Claude Sonnet 5; instructi.
Sources
- Validated by: Single-task success on 100 BEHAVIOR-1K tasks: graph repair only reached 67.7% with Qwen3-4B at 1 LLM call vs 55.4% for SayPlan-style feedback replanning at 3.4 calls, and 76.2% vs 76.4% for Qwen3-8B; adding semantic-only re-queries gave 88.8% and 91.8% at 1.7 and 1.5 calls. Multi-task: 92.6% vs 75.6% feedback-only and 60.2% for graph-edit-only planning, with shorter distance and ~12s less planning time. On a 100-instruction subset, GPT-5.6 Sol went 24.6% to 99.2% and Claude Sonnet 5 38.6% to 99.4%.. [truncated]
Either way, bring evidence: a link, or what you saw.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When training an LLM with policy-gradient RL on tasks that decompose into many dependent sub-problems, rewarding each trajectory segment by how many reference reasoning points it newly reaches beats a sparse final-answer reward by a margin that widens with task horizon, and still learns on problem sets where the base policy never samples a success and sparse reward gives no gradient at all.Completing tasks that take many steps or many hours · reviewed by AI
- When a model must anticipate, without execution feedback, what a change to its own runtime harness components will do, naming which components a lifecycle change can affect stays near ceiling as the number of interacting effects grows from 2 to 32, while predicting the resulting state, reasoning across teardown orders, and choosing minimal reconfigurations degrade sharply (e.g. GPT-5.6 Luna's reachable-condition Jaccard 91.7% to 14.1%, executed reconfiguration success 62.5% to 25.0%).Improving its own scaffolding from failures · reviewed by AI
- In embodied multi-agent household planning with a fixed high-level skill interface, a team's task success rate overstates coordination quality, because goal predicates can be reached through illegal shortcuts: several of 11 tested MLLMs succeed on more episodes than they produce legal plans on (Claude Haiku 4.5 63.1% success vs 47.2% legal; Qwen3.6-Plus 69.1% vs 56.7%), while others honour constraints yet miss the goal.Getting more from several models than from one · reviewed by AI
- For training-free MLLM-driven navigation (instruction-following and object-goal) in simulated indoor environments, gating each planner-proposed action on evidential support, geometric reachability and subgoal consistency before dispatch — and refusing a stop request unless semantic, geometric and progress checks all pass — raises success rate over letting the planner act directly, but the stop check buys completion at some cost in path efficiency.Turning plans into physical action · reviewed by AI
- 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.Planning a sequence of actions before acting · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A setup where feedback-only replanning matches or beats model-derived repair at equal or fewer LLM calls, or where frontier models with plain prompting reach the same success as with the verify-and-repair layer.