{"id":"4a84653f-9ad1-4371-b4fa-53f18a99b0d7","arxiv_id":"2502.02066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using LLM-predicted future chores as joint goals for a PDDL planner reduces simulated household task execution time by about 31%.","lead":"This paper combines large language model task anticipation with classical planning to make household robots do several chores in one trip. In a simulated home, the approach cut planned execution time by about 31% compared with planning each chore separately.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Execution-time reduction is an artifact of hand-assigned action costs; no VirtualHome execution validates the 31% figure.","rationale":"The reader's weakest_assumption correctly identifies that plan cost is treated as execution time without actual simulation, and that this undermines the 31% figure. I agree; this is the most load-bearing weakness because the paper's core contribution is the efficiency gain, and the empirical support for that gain rests entirely on hand-assigned PDDL costs. The paper's internal reasoning (positive goal interaction can reduce plan length) is plausible, and the plan-length reduction is a more robust signal. However, the execution-time headline is not grounded. The proposal to re-run with actual VirtualHome execution or a cost-sensitivity analysis directly tests whether the reduction is an artifact. Since the reader already reached CONDITIONAL and this concern reinforces that, no verdict change is needed. The paper is a useful engineering contribution, but the quantitative claim should be treated as provisional until the cost model is validated.","tokens_in":10874,"tokens_out":3807,"duration_ms":36994,"concrete_test":"Re-run the paired H3 trials in VirtualHome, actually executing the plans (or using VirtualHome's action-duration annotations) and measuring wall-clock time for the myopic and joint conditions. If the measured reduction is materially smaller than 31% (e.g., below 15%), the hand-assigned cost model was the source of the claimed gain. Alternatively, perturb the action costs over a plausible range (e.g., move costs 2-20, manipulation 5-30) and recompute the paired trials; if the mean reduction across cost samples falls below the reported 31%, the claim is cost-model dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is the 31% reduction in execution time, but Section IV-B2 equates execution time with the total PDDL plan cost: \"Since the cost of the actions (in the domain description) was based on the execution time, we used the total cost of any executed plan as the execution time (in seconds) of the plan.\" These per-action costs are hand-assigned (e.g., dusting has a fixed cost of 10, Figure 3), with no basis in measured durations or VirtualHome's own action time model. The plans are never actually executed in VirtualHome for H3; only the sum of costs is reported. Consequently, the 31% reduction is a property of the chosen cost weights. If the relative cost of moving vs. manipulation actions were different, the observed reduction could change substantially. The plan-length reduction (12%) is more robust to cost choices, but the headline claim is specifically about execution time. Without grounding the cost model in real execution durations, the claim that joint planning yields a 31% execution-time speedup is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework that combines LLM-based high-level task anticipation with classical PDDL planning for household robot tasks. Given a partially specified routine of high-level tasks, an LLM (e.g., GPT-4) predicts likely future tasks; these anticipated tasks are then encoded as joint goals for the Fast Downward planner, which computes a fine-grained action sequence that achieves all goals together. The authors evaluate task-anticipation accuracy of several LLMs against a Markov baseline, and they evaluate planning efficiency by comparing plan length and plan cost ('execution time') under myopic versus multi-task joint planning. They report a 31% reduction in execution time and a 12% reduction in plan length when six tasks are anticipated, concluding that anticipating future tasks and planning jointly improves efficiency. The paper also includes a qualitative demonstration of replanning after user interruptions (H4).","tokens_in":10965,"tokens_out":4653,"duration_ms":47025,"significance":"If the reported efficiency gains held up under real execution, the idea of using LLMs for high-level anticipation and classical planners for fine-grained joint execution would be a practical and timely contribution. The framework is lightweight in that it uses off-the-shelf LLMs and a standard planner with no fitted parameters, and the anticipation evaluation (H1/H2) is reasonably thorough, with multiple LLMs, multiple metrics, and a baseline. The pairing of trials and the use of ratios in Figure 7 are appropriate ways to handle varied initial states. However, the central quantitative claim about execution-time reduction is currently not well supported because it rests on hand-assigned action costs rather than measured execution times, and the headline improvement is taken from a single planner configuration while other configurations show much smaller gains. The plan-length reduction is a real cost-free quantity, but it is also smaller and less consistent across planner configurations.","major_comments":[{"comment":"The headline 31% execution-time reduction is not a measured execution time. The paper states that 'Since the cost of the actions (in the domain description) was based on the execution time, we used the total cost of any executed plan as the execution time (in seconds) of the plan,' but the action costs are hand-assigned (e.g., dusting is assigned a fixed cost of 10 in Figure 3) with no validation against VirtualHome's actual action durations or any other source. The plans are never executed in VirtualHome for the H3 experiments. Consequently, the reported 31% reduction is a property of the chosen cost weights; different but equally plausible relative costs for movement versus manipulation actions could change or eliminate the observed reduction. The plan-length reduction is robust to cost choices, but it is not the headline claim. Please either run the plans in VirtualHome to obtain real execution times, or explicitly reframe the claim as a reduction in total plan cost rather than execution time.","section":"Section IV-B2, Table IV, Figure 7"},{"comment":"The experimental setup introduces a confound: 'As the number of anticipated tasks increased, the search time limit provided to the planner was increased by units of 30 seconds.' This means the planner had more time to find lower-cost plans in the multi-task conditions, so the observed cost reduction could be due to additional search time rather than to the joint consideration of goals. To support the claim that anticipation and joint planning cause the improvement, the myopic baseline should be run with the same increased search time budgets, or the search time limit should be held constant across conditions. Without this control, the 31% reduction cannot be attributed to the proposed mechanism.","section":"Section IV-B2, paragraph on search time limits"},{"comment":"The reported 31% execution-time reduction and 12% plan-length reduction are taken from the seq-sat-fd-autotune-1 configuration only, as stated for Figure 7. Table IV shows that the other two planner configurations exhibit substantially smaller improvements: for LAMA, plan execution cost decreases from 1835 to 1599 (about 13%) and plan length from 65.7 to 61.2 (about 6.8%); for AT-2 the corresponding reductions are about 25% and 10%. Presenting the best configuration's numbers as the paper's headline result in the abstract and introduction is misleading. Please report results for all configurations (or an aggregate with variance) and avoid cherry-picking the most favorable configuration.","section":"Table IV and Figure 7"}],"minor_comments":[{"comment":"There is a typo in the equation for total cost: 'aciton' should be 'action'.","section":"Section III-B"},{"comment":"The phrase 'during training or execution' is confusing because the LLM is not trained in this work; consider 'during prompting or execution'.","section":"Section IV-A.1"},{"comment":"For the LAMA configuration, the execution time for 3 anticipated tasks (1613) is very close to that for 6 anticipated tasks (1599), and the reduction from myopic is far smaller than for AT-1. This pattern deserves discussion, as it suggests the benefit of anticipation is configuration-dependent.","section":"Section IV-B.2, LAMA row"},{"comment":"The evaluation of H4 is qualitative and does not report any quantitative metrics or confirmation that the plan was actually executed in VirtualHome. If this is intended as a demonstration, please state explicitly that it is illustrative rather than a full experimental validation.","section":"Section IV-B.3"},{"comment":"The sentence 'positive interaction between goals was unlikely in our experiments' is difficult to reconcile with the observed reduction in plan length, since interleaving actions across goals is itself a form of positive interaction. Please clarify what kind of interaction is meant and how it relates to the observed results.","section":"Section IV-B.2, paragraph on positive interaction"},{"comment":"No error bars, standard deviations, or significance tests are reported for the 10 repetitions in Table IV and Figure 7. Adding these would help the reader assess the reliability of the claimed reductions.","section":"General"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing here is the integration: LLM-based high-level task anticipation feeding multiple goals into a classical planner, evaluated in a household domain with 33 actions and multiple rooms. Each piece is known—LLM+P, anticipatory planning, etc.—but the combination is a legitimate extension and the paper does a decent job showing it works. The task anticipation experiments are actually strong: with contextual prompts, GPT-4 is near-perfect on a closed task set, and the Markov baseline comparison is fair.\n\nThe problem is the headline. The 31% execution-time reduction is not measured; it is the sum of hand-assigned PDDL action costs. The authors state this explicitly (\"Since the cost of the actions... was based on the execution time, we used the total cost of any executed plan as the execution time\"), and the dusting action has a fixed cost of 10. No plan is run in VirtualHome for H3. So the 31% is a property of their chosen cost weights. If moving cost is high relative to manipulation, joint planning looks better because it saves trips; different weights would change the number. The plan-length reduction (12%) is more robust, but it is not the abstract's claim.\n\nOther soft spots are minor by comparison: no error bars or significance tests; only 10 repetitions; H3 does not verify that the LLM-anticipated tasks are the actual tasks that will be requested, so the benefit assumes correct anticipation; the search time limit grows with the number of anticipated tasks, which could confound plan quality; no code or data released.\n\nThat said, the core idea is sound and the paper is honest about its cost assumption. The direction of the benefit is supported by planning theory—joint goals can share actions. The weak spot is the quantitative claim, not the architecture. This deserves a serious referee, but the authors should either run the plans in VirtualHome to measure elapsed time or rebuild action costs from VirtualHome's own durations, and report variance. As it stands, I would treat the 31% as illustrative, not established.\n\nFor a reading group, it is a reasonable choice if you are interested in LLM-plus-planning pipelines. Otherwise, my verdict is conditional: the integration is worth publishing after the execution-time claim is reworked.","headline":"Sensible integration of LLM task anticipation with PDDL planning, but the 31% execution-time claim is an artifact of hand-assigned costs, not measured execution.","tokens_in":11616,"tokens_out":2055,"would_cite":false,"duration_ms":20466,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Anticipating the next few household tasks and planning for them jointly cuts the planner's execution cost by about 31% compared with handling one task at a time, the authors report.","keywords":["task anticipation","large language models","classical planning","PDDL","Fast Downward","VirtualHome","household robotics","assistive agents"],"falsifier":"Run the two plan types in VirtualHome with real action durations measured from execution, compare wall-clock time for myopic versus joint-goal plans on the same initial states and task sequences; if the 31% reduction does not appear, the result is an artifact of the cost assignment.","tokens_in":10606,"feed_emoji":"🤖","tokens_out":5816,"duration_ms":53002,"temperature":0.7,"pith_summary":"What if a household robot could peek at the next few chores on a person's schedule and weave them into one efficient action sequence? This paper claims yes: by asking a large language model to anticipate which high-level tasks come next, and then feeding those tasks as joint goals to a classical planner, an agent can reduce the total cost of executing the sequence by about 31% and its plan length by about 12% compared with tackling one task at a time. The result, demonstrated in the VirtualHome simulation environment, argues that the generic world knowledge of an LLM and the guarantee-providing search of a PDDL-based planner are complementary rather than competing. If the cost model holds up, the same recipe could make assistive robots in homes and workplaces noticeably more efficient without needing large labelled datasets.","feed_headline":"Anticipating tasks cuts household robot time by 31 percent","feed_subtitle":"Pairing an LLM that forecasts upcoming chores with joint planning beats one-at-a-time plans in VirtualHome.","key_machinery":"The central mechanism is the joint-goal planning formulation. High-level tasks anticipated by the LLM are mapped onto ground literals in a PDDL domain description—a formal action language with typed objects, preconditions, effects, and per-action costs—and the Fast Downward heuristic planner searches for a sequence of actions that satisfies all these goals simultaneously while minimizing total cost. The action costs are hand-assigned to stand for execution time in seconds. This setup lets the planner discover and exploit positive interactions between goals, eliminating redundant trips and actions that a myopic one-task-at-a-time approach would take.","core_discovery":"The paper's central claim is that treating LLM-anticipated tasks as joint goals for a classical planner produces shorter, cheaper action sequences than planning each task independently. In their experiments, the authors prompt an LLM (GPT-4, with optional contextual examples) with a partial routine of about twenty high-level household tasks, ask it to predict the next one to six tasks, translate those predictions into PDDL goal literals, and let the Fast Downward planner compute a single plan that accomplishes all goals at once. Averaged over paired trials with identical initial states and task sequences, planning for six anticipated tasks cut execution cost by roughly 31% and plan length by about 12% relative to the myopic baseline. This is the evidence for the claim; the speedup comes from the planner interleaving subactions that serve multiple goals, such as fetching breakfast ingredients while already fetching milk for coffee.","pith_inferences":["The 31% figure is a planner-cost reduction; whether it translates into a real-time 31% speedup on a robot depends on how faithfully the hand-assigned action costs reflect physical execution durations, which the paper does not measure.","The benefit of joint-goal planning likely grows with overlap between tasks (shared objects, locations, and subactions); a direct test would run paired trials with controlled amounts of task overlap.","Because the LLM only outputs tasks from the fixed set in the prompt, the approach inherits that set's coverage and silently ignores off-list tasks; a fallback mechanism would be needed for truly novel requests.","The one-prompt adaptation to schedule constraints (e.g., an urgent morning meeting) suggests the same machinery could personalise anticipation from a user's past routines, but that extension is beyond the paper's experiments."],"forward_implications":["Anticipating six tasks instead of zero cuts execution cost by about 31% and plan length by about 12% in the paper's paired trials.","With contextual examples, GPT-4 anticipates task order perfectly in the paper's 500 experiments (KRCC 1.0, miss ratio 0.0006), well above the Markovian baseline.","The gain is not limited to one planner configuration: all three Fast Downward configurations (AT1, AT2, LAMA) show lower plan length and mostly lower execution cost with more anticipated tasks.","When a plan is interrupted by a new user prompt, the framework can revise the anticipated routine and replan, undo reversible actions, and proceed with the new tasks."],"supporting_citations":[{"why":"Defines PDDL, the action language used to encode the household domain and goal states.","marker":"[6]"},{"why":"Fast Downward, the heuristic planner that computes the joint-goal action sequences.","marker":"[7]"},{"why":"VirtualHome, the simulation environment used to set up and evaluate the household scenarios.","marker":"[8]"},{"why":"GPT-4, the LLM used as the default task anticipator for the planning experiments.","marker":"[9]"},{"why":"PaLM, an LLM evaluated for task anticipation with and without context.","marker":"[10]"},{"why":"GPT-3.5, an LLM evaluated for task anticipation with and without context.","marker":"[28]"},{"why":"Generalized planning in PDDL with pretrained LLMs; its prompting style is adapted for contextual examples.","marker":"[18]"}],"fun_headline_variants":["LLM foresight + joint planning: 31% faster household chores","Household robots save 31% time by anticipating next tasks","Anticipating tasks with LLM cuts robot execution time 31%","Predict next tasks with LLM, plan together: 31% speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed speedup treats the planner's total hand-assigned action cost as execution time, and the resulting plans are not actually run in the simulator to measure elapsed seconds.","fun_headline_variants_meta":{"raw":{"variants":["LLM foresight + joint planning: 31% faster household chores","Household robots save 31% time by anticipating next tasks","Anticipating tasks with LLM cuts robot execution time 31%","Predict next tasks with LLM, plan together: 31% speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001123,"raw_usage":{"total_tokens":4638,"prompt_tokens":875,"completion_tokens":3763,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":3685}},"tokens_in":491,"tokens_out":3763,"duration_ms":27735,"temperature":1.0,"reasoning_tokens":3685,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:27:19.696285+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two plan types in VirtualHome with real action durations measured from execution, compare wall-clock time for myopic versus joint-goal plans on the same initial states and task sequences; if the 31% reduction does not appear, the result is an artifact of the cost assignment.","supporting_citations":[{"cited_title":"The fast downward planning system,","cited_arxiv_id":null,"evidence_quote":"Fast Downward, the heuristic planner that computes the joint-goal action sequences."},{"cited_title":"Virtualhome: Simulating household activities via programs,","cited_arxiv_id":null,"evidence_quote":"VirtualHome, the simulation environment used to set up and evaluate the household scenarios."},{"cited_title":"Palm: Scaling language modeling with pathways,","cited_arxiv_id":null,"evidence_quote":"PaLM, an LLM evaluated for task anticipation with and without context."},{"cited_title":"Language models are few-shot learners,","cited_arxiv_id":null,"evidence_quote":"GPT-3.5, an LLM evaluated for task anticipation with and without context."},{"cited_title":"Generalized planning in pddl domains with pretrained large language models,","cited_arxiv_id":null,"evidence_quote":"Generalized planning in PDDL with pretrained LLMs; its prompting style is adapted for contextual examples."}],"review_version":1}