{"id":"040d0c94-0e7e-4df4-bc0a-50f680a127f3","arxiv_id":"2607.06501","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HUME lets robots generate, plan over, and actively verify object-centric hypotheses from foundation models so incomplete symbolic models become usable for open-world household tasks.","lead":"Service robots can expand incomplete world models by treating foundation-model guesses as uncertain hypotheses and planning actions that both verify them and finish the task. This closes a gap between brittle symbolic planners and hallucinating LLMs for household open-world settings.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Central claim rests on foundation-model reliability for hypothesis generation, verification conditions, and VLM evaluation; experiments do not isolate or stress this assumption.","rationale":"The reader correctly isolates Assumption (3) as the weakest link and rates the paper CONDITIONAL with medium correctness risk. My stress test confirms that this assumption is load-bearing for the strongest claim: without reliable generation + verification, the uncertainty-aware loop does not deliver the reported gains over deterministic expansion. The paper already surfaces related failures (misclassification, grasp occlusion) and optimistic determinization, so the concern is not a hidden flaw but the unquantified sensitivity of the main result. The proposed noise-injection experiment is a direct, concrete check that would settle whether the advantage survives imperfect foundation-model feedback. No stronger internal inconsistency or derivation error appears; novelty and multi-domain ablations remain solid. Hence the verdict stays CONDITIONAL (pending that robustness evidence and the public code the reader already requested), with full agreement on the identified soft spot.","tokens_in":35701,"tokens_out":631,"duration_ms":7309,"concrete_test":"Re-run the five real-world tasks (T1–T5) and the AI2-THOR suite under a controlled verification-noise ablation: inject a fixed false-positive/false-negative rate (e.g., 20–30%) into the VLM attribute/existence answers while keeping hypothesis generation and the planner unchanged. If success rate / SPL of the uncertain-expansion PDDL planner falls to or below the deterministic-expansion baseline, the headline gains are not robust to realistic verification error and the central claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is that uncertainty-aware hypothesis expansion (generate → plan with verification interleaved → update) yields autonomous knowledge expansion and substantially higher success than static or deterministic-expansion baselines for both PDDL and LLM planners. For the iterative loop (Alg. 1, §IV) to close the model gap, Assumption (3) in §IV.F must hold: the LM generates correct hypotheses within a bounded number of attempts, proposes sufficient verification_conditions, and the VLM correctly evaluates them from raw observations. The paper acknowledges this and the optimistic all-outcomes determinization that drops a_h-verify (§IV.D). Real-world failure analysis (App. A.II-B) already attributes residual errors to attribute misclassification by the verification module, and grasp pose can hide discriminative features (App. A.II-C). The reported gains (Figs. 6, 8, 10) therefore confound the value of the planning representation with the quality of the particular gpt-4.1 / VLM stack used; if hypothesis generation or verification systematically fails on harder attribute/effect cases, the loop cannot restore solvability and the claimed advantage over deterministic expansion collapses. This is the single most load-bearing soft spot for the central empirical claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes HUME, an open-world planning framework in which incomplete symbolic models are expanded by object-centric hypotheses (existence, attributes, action effects) generated by foundation models. Hypotheses are treated as uncertain latent variables; classical or LLM planners produce plans that interleave task actions with verification actions via all-outcomes determinization and cost bias, then update the model from VLM/perception feedback and replan (Alg. 1, §IV). Experiments in Block Processing World, AI2-THOR mobile manipulation, real Fetch household tasks, and a microwave appliance demo show that uncertainty-aware expansion substantially improves success rate and SPL over no-expansion and deterministic-expansion ablations for both PDDL and LLM planners.","tokens_in":36052,"tokens_out":1272,"duration_ms":20558,"significance":"If the empirical claims hold under the stated assumptions, the work offers a practical bridge between structured automated planning and unstructured foundation-model knowledge for household robots: model expansion is made explicit, uncertainty-aware, and actively verifiable inside goal-directed planning rather than treated as passive domain generation. Strengths include a clean six-way ablation (representation × inference), consistent gains across simulation and real hardware, dual support for formal and LLM planners, and an honest limitations discussion (§VI, App. A.IV). The hypothesis structure and verification-in-planning design are reusable engineering contributions for open-world service robotics.","major_comments":[{"comment":"§IV.F Assumption (3) and App. A.II-B: The iterative loop (Alg. 1) closes the model gap only if the LM generates correct hypotheses within I_max attempts, proposes sufficient verification_conditions, and the VLM correctly evaluates them. Residual real-world failures are already attributed to attribute misclassification and grasp-dependent visibility (App. A.II-C). The main results (Figs. 6, 8, 10) do not report hypothesis-generation success rates, verification accuracy, or how often regeneration was required. Without these metrics, the claimed advantage of uncertainty-aware expansion over deterministic expansion cannot be fully separated from the reliability of the particular gpt-4.1/VLM stack. Please add quantitative failure breakdowns for generation and verification, or qualify the autonomy claims accordingly.","section":"§IV.F, App. A.II-B/C, Figs. 6/8/10"},{"comment":"§V.D Real-world protocol: Execution noise is mitigated by retrying skills until success and manually resetting invalid states (e.g., drops). This is reasonable for isolating planning, but the abstract and conclusion claim “autonomous knowledge expansion” and “effective operation” without reporting how often retries/resets occurred or how many trials would have failed without intervention. Please report intervention rates per task/planner or explicitly scope the real-world claims to planning performance under idealized low-level execution.","section":"§V.D, Abstract, §VII"},{"comment":"§IV.D Determinization: Verification actions are all-outcomes-determinized and a_h−_verify is excluded so the planner is optimistic; negative outcomes are handled only by rejection and replanning. The paper frames the setting as Bayes-adaptive (§I, §III.B), yet ternary beliefs and branch-cut optimism do not reason about graded risk or irreversible verification side-effects (acknowledged in §VI.a and App. A.IV-D). This is acceptable as a design choice, but the Bayes-adaptive framing should be softened or the optimistic bias stated as a first-class limitation of the planning objective, not only of future work.","section":"§IV.D, §III.B, §VI.a"}],"minor_comments":[{"comment":"Fig. 3 and hypothesis JSON examples: “object_existance” / “existance” should be “existence”; “Trigged” in domain listings should be “Triggered” for consistency with prose.","section":"Fig. 3, §IV.B, App. A.III-A"},{"comment":"Fig. 4’s six-category taxonomy is useful but the caption and body (§V.A) could more explicitly map each bar group in Figs. 6/8/10 to the six named approaches to avoid reader cross-referencing.","section":"Fig. 4, §V.A"},{"comment":"Free parameters (verification penalty c ≫ 0, I_max, LLM sampling) are listed only implicitly; a short sensitivity note or fixed values in the appendix would aid reproducibility.","section":"§IV.D, App. A.III"},{"comment":"Related work §II.B: Tru-POMDP [58] and Seeing-is-Believing [75] are close; a one-sentence contrast on whether verification actions are planned vs. manually specified would sharpen novelty.","section":"§II.B"},{"comment":"Typo: “preconditiaons” in §VI; “deks1” in Fig. 3 plan snippet; “wallmountedcontrolpanel1” style names are fine but ensure PDDL listings match figure labels.","section":"§VI, Fig. 3"}],"recommendation":"minor_revision","confidential_remarks":"Solid empirical systems paper for a robotics venue; the core idea is clear and the ablations are the right ones. The skeptic’s concern about FM reliability is real but partially mitigated because expansion vs. uncertain-expansion share the same hypothesis generator—so the planning-representation claim is still supported. I would not reject on that basis; requiring generation/verification metrics and clearer real-world intervention reporting is enough. Fit for a journal that values embodied planning systems with real-robot validation."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: if you expand a PDDL model with LLM guesses and then treat those guesses as facts, you still fail a lot; if you keep them as hypotheses, force verification into the plan (optimistic all-outcomes determinization + cost bias), and replan on refutation, both PDDL and LLM planners get large gains in Block Processing World, AI2-THOR, and on a Fetch in a real kitchen. That loop (HUME) is the actual contribution—not a new theory of open-world planning, but a concrete, reproducible way to make incomplete symbolic models usable with foundation models.\n\nWhat they do well: the factorization into object-centric hypotheses (existence / attribute / action effect) with explicit verification conditions is clean; the ablations are the right ones (no expansion / expansion-as-fact / uncertainty-aware); and they show the same pattern for formal and LLM planners, which is more informative than another pure-LLM demo. Real-world traces (zero-sugar drink, blue-floral bowl + light switch, microwave defrost) make the behavior readable. Assumptions are stated up front (reliable skills, goals that name the missing concepts, LM can generate and check hypotheses in a bounded number of tries). No circular math; metrics are external success and SPL.\n\nSoft spots, in proportion: the load-bearing assumption is exactly the one the stress note flags—gpt-4.1 / VLM quality for hypothesis generation, verification conditions, and attribute checks. Residual real-world failures are already attribute misclassifications and grasp poses that hide features; they do not isolate the planning representation from the particular model stack. Optimistic branch-cutting of negative verification outcomes is pragmatic but not risk-aware. Real runs retry/reset low-level failures, so “full autonomy” is softened. Free parameters (verification cost, I_max) are ordinary engineering knobs. None of this sinks the central empirical claim under the stated scope; it just bounds how far you can extrapolate.\n\nWho it is for: people building LLM+symbolic household planners who need a working pattern for incomplete models. Worth a serious referee. I would cite the framework when discussing uncertainty-aware domain expansion, and I would bring it to reading group if we are talking open-world TAMP or service robots. Send to peer review.","headline":"Solid systems paper: treating LLM model expansion as uncertain hypotheses and interleaving verification with classical planning is a clean, useful engineering move, with real-robot evidence that the ablations matter.","tokens_in":36598,"tokens_out":560,"would_cite":true,"duration_ms":7942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Robots can expand incomplete world models by treating missing facts as uncertain hypotheses and verifying them while planning toward goals.","keywords":["open-world planning","hypothesis-driven model expansion","uncertainty-aware planning","service robots","foundation models","PDDL","mobile manipulation"],"falsifier":"Run the same open-world household tasks with a language model that systematically invents wrong locations or attributes and mislabels verification images: if success collapses to the no-expansion baseline, the central claim fails.","tokens_in":36642,"feed_emoji":"🤖","tokens_out":583,"duration_ms":6359,"temperature":0.7,"pith_summary":"Service robots in homes face incomplete knowledge: objects may be hidden, attributes unknown, and action effects underspecified. Closed-world planners fail when the model is incomplete, and language models alone can invent facts that look plausible but are wrong. This paper claims that the right move is not to replace planning with a language model, but to use the language model to propose object-centric hypotheses about missing locations, attributes, and action effects, keep those hypotheses uncertain, and plan so that verification and task progress happen together. The resulting system, HUME, iteratively generates hypotheses, plans under their uncertainty, executes verification actions, and updates the model from foundation-model feedback. Experiments in block-processing worlds, household mobile manipulation, real Fetch-robot kitchens, and microwave operation show large gains in success and path efficiency when uncertainty is made explicit, for both classical and language-model planners. A sympathetic reader cares because this is a concrete path from static pre-programmed knowledge to autonomous knowledge expansion in real open environments.","feed_headline":"Robots expand world models by verifying uncertain hypotheses","feed_subtitle":"Planning that treats missing facts as testable guesses beats static models in open homes","key_machinery":"HUME (Hypothesis-driven Uncertainty-aware Model Expansion): foundation models generate factorized object-centric hypotheses; all-outcomes determinization plus cost penalties produce plans that interleave verification with task actions; verification feedback updates the symbolic model and triggers re-generation.","core_discovery":"Explicitly representing model expansion as a set of uncertain object-centric hypotheses, and integrating hypothesis verification into goal-reaching planning, restores solvability and substantially raises success rates under incomplete knowledge, for both formal PDDL planners and language-model planners, in simulation and on real robots.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Robots expand models by planning tests of their own uncertain guesses","Hypothesis checks inside planning restore solvability under missing knowledge","Open-world robots turn incomplete facts into verifiable object hypotheses","Uncertainty-aware expansion plus goal planning lifts robot success rates","Verifying model hypotheses while planning lets service robots handle unknowns"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The language model must generate correct hypotheses in a bounded number of tries, propose workable verification conditions, and judge them correctly from sensory observations; if it cannot, the expansion loop cannot close the model gap.","fun_headline_variants_meta":{"raw":{"variants":["Robots expand models by planning tests of their own uncertain guesses","Hypothesis checks inside planning restore solvability under missing knowledge","Open-world robots turn incomplete facts into verifiable object hypotheses","Uncertainty-aware expansion plus goal planning lifts robot success rates","Verifying model hypotheses while planning lets service robots handle unknowns"]},"model":"grok-4.5","effort":"low","cost_usd":0.00455,"raw_usage":{"total_tokens":1297,"prompt_tokens":712,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":45500000,"prompt_tokens_details":{"text_tokens":712,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":501,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":712,"tokens_out":84,"duration_ms":5827,"temperature":1.0,"reasoning_tokens":501,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T00:17:10.688152+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same open-world household tasks with a language model that systematically invents wrong locations or attributes and mislabels verification images: if success collapses to the no-expansion baseline, the central claim fails.","supporting_citations":[],"review_version":2}