{"id":"fdc48c5f-e5e7-4134-89db-85eba111c3e8","arxiv_id":"2602.12244","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A reinforcement-trained LLM that decomposes abstract household requests into PDDL subgoals and solves them with a symbolic planner outperforms prompting and end-to-end planning baselines on long-horizon tasks.","lead":"AHAT trains a 7B language model to turn abstract household instructions and scene graphs into PDDL subgoals, which a symbolic planner then solves; the training method TGPO injects externally corrected reasoning traces into GRPO. The paper reports large success-rate gains over GPT-5, SayPlan, and GRPO-style baselines on synthetic and public planning benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline success-rate gains rest on a circular judge: the same Qwen3-235B reviewer supplies the completion score in both the TGPO reward (§V-A) and the evaluation metric (§VI-A2), with no human or simulator ground truth. A biased/lenient reviewer can inflate all reported gains.","rationale":"The strongest claim is the empirical success-rate advantage, so the metric is the load-bearing component. The reader's weakest-assumption identification is exactly right: the same Qwen3-235B reviewer is used in the TGPO reward and the reported success metric, with no independent ground truth. This is not merely a philosophical concern; it is a concrete correctness risk because RL can exploit a fixed judge, and the reported gains are large enough (e.g., 89.0% vs. 20.1% on abstract tasks) that even modest reviewer bias would change the conclusion. The manuscript's own limitations section does not flag this shared-reviewer issue, and the public-benchmark results still use the same proxy, so there is no external validation. The method does have independent support: PDDL feasibility is objectively checked, and the symbolic planner is a real component. But the completion score is subjective and unvalidated. A human or simulator-based evaluation on a random sample would settle whether the gains are real. Since this is addressable and the reader already conditioned acceptance on it, the verdict remains CONDITIONAL rather than being upgraded or downgraded.","tokens_in":12534,"tokens_out":3577,"duration_ms":42681,"concrete_test":"Select a stratified random sample (e.g., 60 easy/complex/abstract from the AHAT benchmark, 30 human tasks, 60 from PARTNR/Behavior-1K) of plans produced by AHAT-TGPO and the strongest baselines (GPT-5, SayPlan, DELTA). Have independent human annotators, blind to system, judge whether each plan fulfills the instruction given the scene graph; optionally also execute the plans in a simulator with benchmark-defined task success. Recompute Tables I and II with this human/execution-based SR in place of R_comp. If AHAT-TGPO's margin shrinks or the ordering changes, the circular-reviewer concern lands and the empirical claim must be re-quantified. Also release the reviewer prompts and sampled judgments for audit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—AHAT's large success-rate advantage in Tables I and II—depends on the definition of success in §VI-A2: SR = (1/N) Σ I_feas(o_i)·R_comp(o_i). R_comp is produced by Qwen3-235B, the same model used as the completion reviewer in TGPO's training reward R(o)=I_feas·R_comp (§V-A). The policy is therefore explicitly optimized against this specific judge: the reviewer also provides trace-improvement feedback and filters SFT annotations, so AHAT is shaped to satisfy its preferences. Since R_comp is unvalidated, any systematic leniency or stylistic bias toward AHAT's output inflates every reported success number, including results on PARTNR and Behavior-1K, where the same proxy is used instead of simulator/ground-truth success. The concern is not that the method is unsound or that PDDL feasibility is fake—I_feas is objective—but that the task-completion component, which dominates the headline gaps, is measured by the very model that was used as the reward signal. This is load-bearing because all conclusions about scalability, ablations, and superiority over baselines trace to this single proxy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AHAT, a long-horizon household task planner built on a Qwen2.5-7B LLM that first generates a natural-language decomposition trace and then grounds each subtask into PDDL subgoals, which are solved by an off-the-shelf PDDL planner. The model is trained with SFT on a large synthetic dataset and then with TGPO, an extension of GRPO that uses an external 'trace improver' to correct failed decomposition traces and regenerates subgoals under the corrected trace. The authors claim large gains over prompting-based, general-purpose LLM, and learning-based baselines on a held-out synthetic benchmark, human-authored tasks, and two public benchmarks (PARTNR and Behavior-1K), and report scalability in environment size, plan length, abstractness, and constraints. The paper's central quantitative claims rest on an evaluation metric SR = (1/N) Σ I_feas·R_comp, where R_comp is produced by the same Qwen3-235B 'reviewer' that supplies the completion reward in TGPO training. This identity raises serious concerns about the validity of the reported success-rate figures.","tokens_in":12925,"tokens_out":12424,"duration_ms":131346,"significance":"If the reported numbers are trustworthy, the paper would be a significant advance: AHAT's decompose-then-solve architecture cleanly combines LLM-based intention understanding with symbolic planning, and TGPO is an interesting RL variant that incorporates external trace correction. The 50k-task dataset with persona-driven instruction diversity is a valuable resource, and the paper is transparent about its PDDL feasibility checking and planning pipeline, which are machine-checkable and reproducible in principle. However, the evaluation's dependence on the same judge used for training is a major threat to the internal validity of all reported gains. The paper's strengths in method design and dataset construction cannot be fully credited until an independent success criterion—human judgments or simulator-based ground truth—is provided.","major_comments":[{"comment":"The training reward (§V-A) and evaluation metric (§VI-A2) are both R(o)=I_feas·R_comp with the same Qwen3-235B 'reviewer' producing R_comp. Since TGPO optimizes the policy against this judge and the SFT data are filtered/reviewed by the same model (§IV-B), the reported success rate is a measure of fit to this judge, not of task completion. Feasibility is objective, but completion dominates on abstract tasks (Table I: 89.0% vs 20.1% baseline). No human or simulator ground truth is reported for any benchmark. This affects every claimed gain (Tables I–IV, Fig. 3). The authors should add human-rated subsets or use the official success criteria of PARTNR/Behavior-1K.","section":"§V-A, §VI-A2"},{"comment":"The 'Human Tasks' test set is not human-evaluated: the instructions are authored by humans, but the completion score is calculated with the same Qwen3-235B reviewer used in training. The 82.2% success rate therefore reflects the reviewer's opinion, not human satisfaction. This label is misleading and should be changed (e.g., to 'human-authored tasks'), and a genuine human evaluation should be provided to support the claim of performance on real user tasks.","section":"§VI-A1"}],"minor_comments":[{"comment":"The name 'E2E-TRPO' appears in Table III, while the text and Fig. 3 refer to 'E2E-TGPO'. Please fix this inconsistency.","section":"Table III"},{"comment":"The trace improver model is not identified. Please state which model is used and whether it is the same as the reviewer, as this affects the interpretation of the training signal.","section":"§V-B"},{"comment":"The paper does not report how the baseline methods are evaluated with respect to the reviewer; e.g., whether the same PDDL problem construction and reviewer prompt are used for SayPlan and DELTA. Please provide details for reproducibility.","section":"§VI-A2"},{"comment":"The x-axis labels 'Task Abstractness' and 'Constraint Complexity' are vague; please define the precise metrics used to compute these axes.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the circularity of the evaluation: the same Qwen3-235B reviewer provides both the training reward and the reported success metric. This is a load-bearing threat to every quantitative claim. If the authors can provide an independent evaluation (even a small human study or official simulator-based success criteria for PARTNR/Behavior-1K), the paper would be substantially stronger. The rest of the methodology is well-structured and the dataset is a valuable contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: AHAT is a real methodological step — training an LLM with GRPO to emit PDDL subgoals, with an external trace-correction loop (TGPO) that constrains intermediate reasoning tokens while letting the policy generate the subgoals. That combination is new as far as I can tell. The paper also ships a large synthetic data pipeline (50k tasks, 308 scene graphs, 1.6k personas) that looks careful on the surface. Credit where due: the feasibility check via an actual PDDL planner is objective, and the ablations show the subgoal decomposition and the trace guidance each matter. The scalability plots are suggestive.\n\nNow the soft spot, and it's a big one. The reported success rate is not an independent measure of task completion. In §VI-A2 the SR is (1/N) Σ I_feas·R_comp, and R_comp is a discrete label from Qwen3-235B. The TGPO training reward in §V-A is exactly the same formula with the same reviewer. So the policy is optimized to maximize this judge's opinion, and then evaluated with the same judge's opinion. Any systematic lenience in the reviewer toward AHAT's output style inflates every number in Tables I and II. This applies to the PARTNR and Behavior-1K results too: they use the same proxy, not a simulator, not human raters. The feasibility part is solid, but the completion part is the part that carries the big head-to-head gaps (e.g., 89% vs 20%).\n\nThis is not a knock on the algorithm. The method is coherent and the writing is clear. The limitations section is honest about scene-graph dependence and full observability, but it doesn't mention that the judge used for evaluation is the same judge used for training. For a systems paper, that's a load-bearing omission. And the held-out AHAT benchmark is sampled from the same synthetic distribution as the training set, so in-domain generalization is less impressive than it first looks. The human-task set is small (50 tasks).\n\nWho benefits from reading this: anyone working on LLM+PDDL planning or on RL for task decomposition. The TGPO idea is worth trying in other settings. But don't take the headline numbers at face value.\n\nRecommendation: send it out to serious reviewers, but the authors need to either (a) re-evaluate with an impartial judge (e.g., a different model or human raters, or simulator execution where available) or at least report agreement with the current reviewer, and (b) release code/data so the numbers can be checked. Without an independent judge, the empirical claim is not supportable. With it, this could be a worthwhile systems contribution.","headline":"Solid new training recipe for LLM+PDDL, but the headline success rates are computed with the same LLM judge that supplies the training reward, so the numbers are not trustworthy as reported.","tokens_in":13397,"tokens_out":2780,"would_cite":true,"duration_ms":29879,"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":"A trained planner turns terse household instructions into symbolic subgoals and reaches 89% success on abstract tasks, sharply ahead of prior methods.","keywords":["long-horizon planning","household robotics","PDDL","subgoal decomposition","reinforcement learning","abstract instructions","scene graphs","GRPO"],"falsifier":"Run AHAT on a held-out set of abstract instructions and have human annotators judge whether each generated plan would satisfy the request, then compare those human labels with the reviewer model's scores; a large mismatch would indicate that the reported success rates reflect reviewer leniency rather than genuine task understanding.","tokens_in":12465,"feed_emoji":"🤖","tokens_out":3581,"duration_ms":38624,"temperature":0.7,"pith_summary":"The paper argues that long-horizon household planning fails when language models try to generate entire action plans directly, because intention understanding, constraint satisfaction, and long-horizon reasoning get entangled. Instead, the authors propose AHAT, a model trained via reinforcement learning to decompose an abstract instruction into a sequence of PDDL subgoals, which are then solved by an off-the-shelf symbolic planner. The central claim is that this split, plus a new RL algorithm called TGPO that externally corrects the intermediate decomposition trace, lets planning scale to large scenes, long plans, and highly ambiguous instructions. Reported success rates on an abstract task suite are 89.0% versus 20.1% for the strongest prompting baseline. If true, it means a relatively small trained model plus a symbolic planner can handle human-level housework requests without end-to-end action generation.","feed_headline":"RL-trained planner solves 89% of abstract household requests","feed_subtitle":"It turns terse instructions into symbolic subgoals, beating much larger prompted models.","key_machinery":"The central object is TGPO, a two-pass reinforcement learning algorithm built on GRPO. In the first pass, the policy generates candidate decomposition traces and PDDL subgoals; failed candidates are sent to an external trace improver that rewrites the natural-language decomposition. In the second pass, the corrected trace is forced into the rollout via token-level constrained sampling while the policy continues to generate subgoals autoregressively. Group-relative advantages are computed over the augmented candidate group. This gives the policy credit for producing subgoals that the improved trace makes solvable, while preserving on-policy optimization. The second key piece is the PDDL subgo","core_discovery":"AHAT establishes that a language model trained to emit PDDL subgoals, rather than low-level actions, can plan household tasks from terse instructions with high success. The core discovery is that reinforcement learning benefits materially from externally fixing the intermediate decomposition trace while letting the policy propose the subgoals: TGPO's two-pass design yields 89.0% success on abstract tasks versus 20.1% for the strongest prompting baseline. The paper further reports that performance degrades only modestly as scene graphs grow, plans lengthen, instructions become more abstract, and constraints become denser, whereas baselines degrade substantially along one or more of these axes","pith_inferences":["The reported success rates rely on the same LLM reviewer that provides training rewards; if that reviewer is lenient toward the policy's output style, the absolute numbers may overstate real-world performance, and an independent human evaluation would be needed to confirm the gains.","The method assumes a closed, predefined PDDL domain and full observability, as the paper acknowledges; a natural extension is to learn domain constraints from raw sensor data, which could open partially observable settings.","TGPO's trace correction could be applied beyond planning: any task where an intermediate chain-of-thought can be improved by an external checker (for example, proof steps or code reasoning) might benefit from the same constrained-sampling RL loop."],"forward_implications":["If AHAT's central claim holds, household robots can handle natural, terse instructions by combining a small trained LLM with a symbolic planner, reducing the need for large end-to-end action-generating models.","Robustness to plan lengths over 30 actions and scene graphs of roughly 150 nodes suggests the decompose-then-solve design avoids the error accumulation that plagues monolithic action-sequence generation.","Because the policy only outputs subgoals, the same framework could transfer to new PDDL domains with additional domain-specific annotations and retraining.","TGPO provides a template for injecting external feedback into the reasoning traces of any LLM whose outputs are evaluated by a verifier, which may extend beyond planning to other structured reasoning tasks."],"fun_headline_variants":["RL planner hits 89% on abstract household tasks","Corrected reasoning boosts RL planning to 89%","Household planner: 89% on terse instructions","Scalable planner beats large models on abstract tasks","RL with corrected traces solves household tasks at 89%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported success rates are scored by the same LLM reviewer that provides the reward during training; if that reviewer is biased toward the policy's style or overly lenient, the numbers are not a reliable measure of how well a real user's request is fulfilled.","fun_headline_variants_meta":{"raw":{"variants":["RL planner hits 89% on abstract household tasks","Corrected reasoning boosts RL planning to 89%","Household planner: 89% on terse instructions","Scalable planner beats large models on abstract tasks","RL with corrected traces solves household tasks at 89%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000586,"raw_usage":{"total_tokens":2580,"prompt_tokens":720,"completion_tokens":1860,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":1783}},"tokens_in":464,"tokens_out":1860,"duration_ms":13287,"temperature":1.0,"reasoning_tokens":1783,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T06:00:13.705273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AHAT on a held-out set of abstract instructions and have human annotators judge whether each generated plan would satisfy the request, then compare those human labels with the reviewer model's scores; a large mismatch would indicate that the reported success rates reflect reviewer leniency rather than genuine task understanding.","supporting_citations":[],"review_version":1}