{"id":"345f9154-3c34-4eef-8fc5-13d81872ef3a","arxiv_id":"2411.14995","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SIFT learns lifted STRIPS action models (predicates and action schemas) from action traces alone, with formal guarantees and scalability to state graphs of about 500,000 edges.","lead":"A new algorithm named SIFT learns the predicates, preconditions, and effects of a planning domain from action sequences alone, without seeing the states. It combines the scalability of the LOCM heuristic with formal soundness and completeness guarantees, and is evaluated on standard puzzle and logistics domains with hundreds of thousands of transitions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Formal guarantees require Definition 14 completeness (all traces from the initial state, connected, covering every predicate), but the paper's plain-trace protocol violates it — 4 of 5 traces start from non-initial states — so Theorems 15 and 20 do not cover the headline experiments, which fail on…","rationale":"The paper's contribution is substantial and its intended claim is clear: SIFT learns a lifted STRIPS domain equivalent to the hidden one, from action traces or extended traces, with a scalable, sound-and-complete method (Theorems 15 and 20). Independent support is real: a crisp feature-consistency formulation (Definitions 1-8), a linear-time consistency algorithm (Theorem 9), an internally coherent proof structure for Theorem 15, public code, and scalable experiments on full state graphs (n-puzzle, roughly 500k edges). For the central claim to be true, the training traces must satisfy Definition 14's completeness: all traces from the initial state, connected via state equalities, affecting every predicate, with the initial state recoverable as A(D,T). The least secure point is not inside the proofs but in the relation between that condition and the paper's own experimental protocol: the plain-trace training set takes only the first of five traces from s0, the rest from arbitrary reachable states, with no state equalities, so Theorem 12's propagation — and hence Theorem 15's non-applicability transfer — requires a connectedness this protocol lacks. The empirical fingerprints match: plain traces fail exactly on Grid (0%) and Sokoban (4%), the two domains the paper argues need state equalities, and the authors concede Grid is not learnable from plain traces at all, so the abstract's unqualified sound-and-complete claim is only substantiated for extended traces. I also examined whether a different gap is more severe: the type-based feature pruning in the Implementation section is not proved to preserve the completeness of the feature set that Theorems 15 and 20 quantify over, and Theorem 20's separating-trace existence is asserted in a one-sentence appendix proof, with the Generalization section itself flagging that proving it may even be undecidable in general. These are genuine soft spots, but they compound the conditionality rather than replacing it, since the extended-trace experiments achieve 100% verification on every domain and the completeness condition appears attainable in practice. The proposed test settles the operative question: whether enforcing Definition 14 in the training protocol is what separates the 0%/4% failures from 100% success. I therefore agree with the reader's identification of the weakest assumption, and the CONDITIONAL verdict stands: the paper should either restrict the strong guarantee to extended traces, prove completeness for its plain-trace protocol and for the typed enumeration, or adopt a truth-based rather than not-false verification criterion where determinable.","tokens_in":18227,"tokens_out":38399,"duration_ms":364432,"concrete_test":"Use the released SIFT code to compute, for each Table-1 domain, whether the training traces satisfy Definition 14: (i) every trace starts from s0, (ii) every hidden predicate changes value on some action in T, and (iii) Init agrees with A(D,T) up to signs. Then re-run plain-trace learning under a fully Def-14-compliant protocol — all traces from s0, long enough to flip every predicate, with the shared initial state provided as an extended-trace equality — and compare verification rates against Table 1. If Grid and Sokoban rise from 0%/4% to 100% while the other domains stay at 100%, Definition 14's completeness is exactly the load-bearing condition, confirming the reader's diagnosis and showing the abstract's claim must be restricted to extended traces or complete trace sets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that SIFT soundly and completely learns an equivalent lifted STRIPS domain from action traces alone — is conditional on Definition 14, and that condition is both unverifiable and violated by the paper's own protocol. Definition 14 requires that all training traces start from the initial state, that the trace set be connected (so truth values propagate per Theorem 12), that every hidden predicate be affected by some action in T, and that the initial state match A(D,T) up to signs. The learner cannot verify any of this from plain traces. More sharply, the plain-trace protocol in the Experiments section does not satisfy the definition: of the n = 5 traces, only the first starts at s0; the other four start at random states reachable in m ∈ [2L, 5L] steps, and plain traces contain no state equalities, so the set is not connected. The proof of Theorem 15 depends on Theorem 12's propagation, which requires exactly that connectedness, and Definition 10's intersection-based precondition learning needs the resulting determinacy. Hence Theorems 15 and 20 do not apply to the headline 'traces' columns of Table 1. The predicted failure appears precisely there: Grid verifies at 0% and Sokoban at 4%, and the authors argue Grid is not learnable from plain traces at all. The unqualified abstract claim — 'from action traces alone … sound and complete' — is therefore only substantiated for extended traces. Two compounding gaps point the same direction: the implementation prunes features by trace-derived types without a proof that the typed feature set still contains every valid feature (Theorems 15 and 20 quantify over the full feature set F(T)), and Theorem 20's separating-trace existence is non-constructive, with the Generalization section noting that proving it may even be undecidable in general. Each gap is a place where the advertised guarantee exceeds what is established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SIFT, a method that learns a lifted STRIPS domain — including predicate symbols and arities — from action traces, extended traces (with state equalities), or labeled state graphs, with no prior knowledge of the domain predicates. The central device is a 'feature' f = ⟨k, B⟩, an assumption that a k-ary predicate is affected exactly by the action patterns in B. Features are tested for consistency with the traces through equality/inequality constraints over pattern signs induced by consecutive and forking occurrences; the test reduces to 2-CNF satisfiability (Theorem 9). Consistent ('admissible') features define the learned domain's predicates and effects, while preconditions are derived by intersecting literals true in all trace nodes where the action is applied (Definition 10). The main results are: truth-value determinacy over connected traces (Theorem 12); soundness and completeness of the learned instance for complete trace sets (Theorem 15, Definition 14); validity of hidden-domain features (Theorems 17 and 19); and an existential statement that some finite extended-trace set makes consistency coincide with validity, so the learned domain is equivalent to the hidden one (Theorem 20). Experiments on 13 domains, including full state graphs with up to about 500,000 transitions (n-puzzle, Logistics), achieve 100% verification on full and partial graphs across all domains, while plain traces fail on Grid (0%) and Sokoban (4%).","tokens_in":18571,"tokens_out":24870,"duration_ms":235960,"significance":"If the results hold, this is a material advance over both LOCM (scalable but heuristic, limited arity handling) and SAT-based approaches (sound and complete but graph-only and small-scale): the consistency test is elegant and scalable, the learned features are interpretable and largely match the predicted maximal description Dmax, and the experiments validate on instances larger than the training instances, with public code and data. The paper is unusually honest about the cases where the method fails (Grid, Sokoban on plain traces) and about the non-constructive nature of Theorem 20. The main caveats are: (i) the formal guarantees require trace completeness in the sense of Definition 14, a condition that the paper's own plain-trace protocol violates and that the learner cannot verify; (ii) the type-based feature pruning in the implementation is not covered by the theorems; and (iii) the completeness claim is existential. These caveats mean the abstract's unqualified 'sound and complete' phrasing is stronger than what is established, but they are addressable through sharper claim-scoping rather than through reworking the core mechanism, which I found sound.","major_comments":[{"comment":"The formal guarantees are stated only for trace sets that are complete in the sense of Definition 14: all traces drawn from the initial state of P (hence connected), every hidden predicate affected by some action in T, and I agreeing with A(D,T) up to signs. The plain-trace protocol described in the Experiments section violates the connectedness requirement by construction: the first trace is sampled from s0 but traces 2–5 start from states reached in m random steps with 2L ≤ m ≤ 5L, and plain traces carry no state equalities, so the trace set shares no common state. Theorem 12's truth-value propagation, on which the proof of Theorem 15 rests, requires exactly that connectedness, and Definition 10's intersection-based precondition learning requires the resulting determinacy. Consequently, Theorems 15 and 20 do not apply to the plain-trace columns of Table 1, and the abstract's unqualified claim of being 'sound and complete' when learning 'from action traces alone' is substantiated only for complete extended traces. The observed failures (Grid Verif 0%, Sokoban Verif 4%) are consistent with this gap rather than being explained as isolated sampling artifacts. I recommend stating the scope of the guarantee explicitly (complete, connected trace sets from the initial state) and presenting the plain-trace results as empirical findings outside the theorem coverage; notably, the extended-trace inputs (full and partial graphs) do satisfy the connectedness requirement, and SIFT achieves 100% verification there, which is consistent with the theory's intended scope.","section":"Experiments (Training Data); Definition 14; Theorem 15"},{"comment":"The feature space is reduced by the type-extraction and type-merging heuristic before consistency testing, but no lemma establishes that this pruning preserves the set of valid or admissible features. The completeness guarantees (Theorem 15, part 2, and Theorem 20) are stated for the full set F(T), so the implemented system is not literally covered by the theory. Moreover, preservation is not automatically true: a valid feature f = ⟨k, B⟩ can be lost when the objects bound to an argument of f through different action patterns in B never coincide on the merged positions, so that the feature type for f is never enumerated — for example, if f is affected by patterns a[1] and b[1] and the objects appearing at position 1 of a are disjoint from those appearing at position 1 of b in the given traces. Please provide a preservation lemma with its conditions, or explicitly mark the typing step as a heuristic whose safety is currently established only empirically by the 100% verification rates.","section":"Implementation (Features)"},{"comment":"Theorem 20 is purely existential: it asserts the existence of a finite extended-trace set that separates consistent from valid features, without constructing it or bounding its size, and its appendix proof is a three-sentence sketch. The paper itself concedes in the Generalization section that identifying such a set or proving validity 'is not simple and may even be undecidable in general.' Combined with Definition 14 — whose conditions (connectedness from the initial state, predicate coverage, initial-state agreement with A(D,T)) are properties of the hidden world that the learner cannot check — this means the completeness half of the 'sound and complete' claim is not an algorithmic guarantee. I do not see a circularity here, only an existentiality gap; admissibility is a genuine necessary condition for validity (Theorem 17), and the missing direction is the non-constructive one. I suggest that the abstract and introduction be qualified accordingly, and that the proofs of Theorems 17, 19, and 20 (and the existence argument for the maximal domain Dmax) be expanded well beyond their current one-line form.","section":"Generalization (Theorem 20); Appendix proofs"},{"comment":"The statement that 'one can show that Grid cannot be learned correctly from plain traces alone' is load-bearing for the paper's interpretation of the Grid 0% verification as a learnability boundary rather than a limitation of SIFT, but no proof or even proof sketch is provided. Please include the argument in the appendix — for example, an information-theoretic obstruction or two Grid instances with the same set of plain traces but different hidden domains — or soften the claim to a conjecture. The Grid-Lock and Sokoban-Pull results provide indirect support for the sampling interpretation, but the direct 'one can show' claim is currently unsupported.","section":"Experiments (Grid and Sokoban paragraph)"}],"minor_comments":[{"comment":"Typo: 'the the elimination of redundant features' should read 'the elimination of redundant features.'","section":"Discussion"},{"comment":"The don't-care symbol used in the motivational questions renders as an empty glyph in this version; choose an explicit token (e.g., '·' or '_') and use it consistently.","section":"Preview"},{"comment":"The feature tables contain several typos and ambiguities ('corrdinate,' 'has has,' 'black' for 'blank'); also, the left/right pattern convention (positive vs. negative signs) is only explained implicitly through the meanings column and should be stated in the table captions.","section":"Tables 5–8"},{"comment":"Please specify how the per-action static atoms pa(o) are instantiated in the larger verification instances; as written they are true iff the ground action occurs in the trace set, which renders the static part of the verification vacuous, so the claim that learned domains are 'verified on larger instances' should be stated as covering the dynamic predicates.","section":"Static Predicates"},{"comment":"The informal use of 'SIFT, on the other hand, is complete' in the footnote is stronger than what is proved; tie it to the formal statements (Theorems 15 and 20) to avoid confusion about the scope of the guarantee.","section":"Related Work, footnote 1"},{"comment":"The two border conditions (a pattern instance acting as a fork with itself, and consecutive patterns across sub-traces connecting common states) are described only in prose; a precise formalization or a small example would clarify the edge cases of the consistency test.","section":"Definitions 5–7"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern lands in its essentials: the plain-trace protocol indeed violates Definition 14, so the headline plain-trace columns of Table 1 are outside the theorem coverage, and the failures in Grid and Sokoban confirm rather than contradict the gap. I do not see genuine circularity — admissibility is a necessary condition for validity (Theorem 17), and the paper is honest about the existentiality of Theorem 20. The paper is a strong candidate after claim-scoping revisions: the core consistency test is elegant, the scale (n-puzzle full graph with 483,840 edges) is impressive, and public code and data are provided. The main risk is overclaiming in the abstract and introduction; the required fixes are local and within scope. Fit with the journal is good, and I have no citation or novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jonas et al. have a genuinely new idea here. The feature-as-assumption formulation, the 2-SAT consistency test, and the linear-time coloring algorithm are a real advance over both LOCM's heuristics and the unscalable SAT approaches. The experiments on full graphs with 500k edges, and the public code, show the method scales. The paper is also honest about the main limitation: the soundness and completeness theorems require Definition 14 'complete' traces, which are unverifiable, and the plain-trace protocol used in the headline experiments violates that definition—only one of five traces starts at the initial state, and the others start at random reachable states. So Theorems 15 and 20 do not cover the 'Traces' columns of Table 1. The predicted failure shows up exactly there: Grid verifies at 0% and Sokoban at 4%. The claim 'from action traces alone ... sound and complete' is therefore only established for extended traces, not for the plain-trace setting the abstract advertises.\n\nTwo other gaps matter. The implementation prunes features by trace-derived types, but there is no proof that the typed feature set still contains every valid feature; the theorems quantify over the full feature set. And Theorem 20's separating traces are non-constructive; the paper itself notes that proving validity may be undecidable. These don't sink the contribution, but they do limit how far the guarantees reach.\n\nI want to be clear: the central mechanism—testing a feature by checking sign constraints—is solid. The true positives are real: valid features always pass, and the learned domains on extended traces match the hidden predicates plus sensible redundant ones. The verification metric is permissive ('not false' rather than 'true'), but for incomplete traces that is a reasonable choice.\n\nThis paper deserves a serious referee. The right outcome is probably acceptance after the authors either restrict the claims to extended traces with the stated completeness conditions, or prove that the type pruning preserves completeness for the trace families they use. I would cite it for the feature-consistency test alone.","headline":"SIFT's feature-consistency test is a real advance, but the 'sound and complete from action traces alone' claim only holds for extended traces that satisfy an unverifiable completeness condition—which the paper's own plain-trace protocol violates.","tokens_in":19163,"tokens_out":2559,"would_cite":true,"duration_ms":23414,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a hidden lifted STRIPS planning domain can be reconstructed from action traces alone, with the learned domain provably equivalent to the hidden one whenever the traces are complete, and it implements this claim in a…","keywords":["lifted STRIPS learning","action model acquisition","action traces","feature consistency","action patterns","2-CNF satisfiability","classical planning","domain learning"],"falsifier":"Take a hand-built STRIPS domain with a predicate of arity 3 whose only effect is through one action pattern set, enumerate all extended traces up to a modest bound from a complete-data protocol, and run SIFT on the full state graph; if any feature outside Dmax survives the consistency test on all complete traces, the equivalence guarantee of Theorem 20 is false. A practical version is the paper's own Grid domain: if plain traces sampled from the initial state ever reach 100% verification on larger instances, the claim that Grid is not learnable from plain traces alone is falsified.","tokens_in":18031,"feed_emoji":"🧩","tokens_out":6595,"duration_ms":61667,"temperature":0.7,"pith_summary":"The paper claims that a STRIPS planning domain can be learned from action traces alone—sequences of action names with no state observations—without being told the predicates, their arities, or the action schemas. The method, SIFT, hypothesizes candidate features (a guessed predicate together with the action patterns that may change it) and tests each hypothesis against the traces by solving a small system of sign-equality and sign-inequality constraints that reduces to 2-CNF satisfiability. The paper proves that on complete trace sets the learned domain reproduces the hidden one: every trace stays applicable, and every action that is inapplicable in the hidden instance is also inapplicable in the learned instance. A further theorem states that a finite set of extended traces (traces with state equalities) forces the learned domain to be equivalent to the hidden domain. If these guarantees hold in practice, action-model learning becomes a tractable, assumption-light alternative to approaches that need labeled state graphs or image-level state information.","feed_headline":"SIFT learns hidden planning rules from action traces alone","feed_subtitle":"Scalable like LOCM yet sound and complete at 500k-edge scale, verified at 100% on benchmark domains.","key_machinery":"The central object is the feature f = ⟨k, B⟩: a guessed predicate of arity k together with a non-empty set B of action patterns a[t], where a is an action name and t is a tuple of argument positions binding the atom's variables to the action's arguments. The mechanism that carries the argument is the consistency test: from action traces one extracts pattern constraints (inequalities for patterns that appear consecutively while affecting the same ground atom, equalities for fork patterns around a shared state) and asks whether the signs of the patterns can be assigned 0/1 to satisfy them. The test reduces to 2-CNF satisfiability and runs in time linear in the number of patterns, which is what lets SIFT scale to state graphs with half a million edges.","core_discovery":"On its own terms, the paper's central discovery is that the hard part of learning a lifted STRIPS domain—finding the predicates and which actions affect them—can be reduced to a consistency test on features. A feature f = ⟨k, B⟩ is the hypothesis that the hidden domain contains an atom f(x1,...,xk) whose truth value is changed by exactly the action patterns in B (an action name plus a tuple of argument positions). From the traces one reads off two kinds of constraint: consecutive occurrences of patterns that both affect the same ground atom must have opposite signs (one adds, one deletes), while fork patterns—patterns whose ground actions diverge from or converge to a common state—must have the same sign. A feature is admissible when these constraints are satisfiable, and the admissible features become the predicates of the learned domain once preconditions are filled in from the observed truth values. The paper shows that every feature of the hidden domain is admissible over any traces from it, that invalid features are ruled out by some finite set of extended traces, and hence that the learned domain can be equivalent to the hidden domain, with experiments reporting 100% verification on all benchmark domains when the inputs are full or partial state graphs.","pith_inferences":["Theorems 15 and 20 suggest that state equalities are the real currency of trace-based learning; adding even sparse pairwise state equalities to plain traces may rescue domains like Grid without requiring a full state graph, a testable extension of the partial-graph experiments.","The redundant predicates SIFT learns (Dmax features such as an undirected 'on' relation in Blocksworld) are exactly the derived predicates that can be tracked by action effects, so the learned domain could double as a library of useful abstractions for planning rather than merely a ground-truth replica.","The paper mentions a robust variant that prunes a feature only after it is inconsistent with k traces; observing how verification degrades as trace length or trace count shrinks would give a practical calibration of how much data completeness actually needs.","The hard cases are sampling problems, not learning problems: Grid and Sokoban fail only because random walks get trapped in dead-ends, which predicts that guided traces from a strong planner would succeed where random traces fail."],"forward_implications":["In any domain whose traces are complete, SIFT yields a learned instance in which every training trace is applicable and every action that is inapplicable after a trace in the hidden instance is inapplicable in the learned instance (Theorem 15).","A finite set of extended traces distinguishes valid features from invalid ones: a feature is consistent with all of them exactly when it appears in the maximal domain Dmax, so the learned domain becomes equivalent to the hidden domain (Theorem 20).","Full and partial state graphs, which carry state equalities, yield 100% verification in all tested domains including n-puzzle with about 500,000 edges; plain traces alone fail in Grid (0%) and Sokoban (4%), and the paper argues Grid is not learnable from plain traces alone.","Because the learned domain is equivalent to the hidden one on complete traces, the learned schemas generalize to larger instances than the one used for training, which is how the paper verifies its results.","The 2-CNF consistency test is polynomial, so feature pruning is cheap; the main scalability win comes from generating typed, ordered features from trace-level type information, cutting candidate counts by orders of magnitude."],"supporting_citations":[{"why":"Introduces LOCM, which also accepts action traces and outputs lifted domain descriptions; SIFT shares its scalable spirit but replaces heuristics with a consistency test.","marker":"Cresswell, McCluskey, and West 2013"},{"why":"Earlier LOCM formulation defining the consecutive-action intuition on traces that SIFT formalizes through action patterns and features.","marker":"Cresswell and Gregory 2011"},{"why":"Weighted Max-SAT approach that learns first-order representations from labeled state graphs; supplies the sound-but-not-scalable benchmark SIFT matches.","marker":"Bonet and Geffner 2020"},{"why":"Extends the SAT/ASP approach to black-box states; example of graph-based learning that does not scale, which SIFT is contrasted against.","marker":"Rodriguez et al. 2021"},{"why":"LATPLAN learns STRIPS-like models from image-based traces in latent space; the propositional deep baseline SIFT positions against.","marker":"Asai and Fukunaga 2018"},{"why":"Extended latent-space planning system showing that deep learning approaches do not yet yield meaningful lifted representations, motivating SIFT's symbolic output.","marker":"Asai et al. 2022"}],"fun_headline_variants":["SIFT: sound, complete, scalable learning from action traces alone","Consistency test unlocks STRIPS models without any prior predicates","From traces to rules: SIFT learns planning domains with no restrictions","Scalable soundness: SIFT learns hidden predicates from action traces","Action traces alone? SIFT discovers the hidden planning rules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantees rely on the training traces being complete for the hidden instance: drawn from its initial state, affecting every predicate of the domain, and carrying enough information to recover the initial truth values of the affected atoms; a learner cannot verify this completeness from the traces themselves.","fun_headline_variants_meta":{"raw":{"variants":["SIFT: sound, complete, scalable learning from action traces alone","Consistency test unlocks STRIPS models without any prior predicates","From traces to rules: SIFT learns planning domains with no restrictions","Scalable soundness: SIFT learns hidden predicates from action traces","Action traces alone? SIFT discovers the hidden planning rules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1732,"prompt_tokens":977,"completion_tokens":755,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":668}},"tokens_in":593,"tokens_out":755,"duration_ms":8738,"temperature":1.0,"reasoning_tokens":668,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:38:39.952965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a hand-built STRIPS domain with a predicate of arity 3 whose only effect is through one action pattern set, enumerate all extended traces up to a modest bound from a complete-data protocol, and run SIFT on the full state graph; if any feature outside Dmax survives the consistency test on all complete traces, the equivalence guarantee of Theorem 20 is false. A practical version is the paper's own Grid domain: if plain traces sampled from the initial state ever reach 100% verification on larger instances, the claim that Grid is not learnable from plain traces alone is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier LOCM formulation defining the consecutive-action intuition on traces that SIFT formalizes through action patterns and features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Weighted Max-SAT approach that learns first-order representations from labeled state graphs; supplies the sound-but-not-scalable benchmark SIFT matches."},{"cited_title":"D.; Bonet, B.; Romero, J.; and Geffner, H","cited_arxiv_id":null,"evidence_quote":"Extends the SAT/ASP approach to black-box states; example of graph-based learning that does not scale, which SIFT is contrasted against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LATPLAN learns STRIPS-like models from image-based traces in latent space; the propositional deep baseline SIFT positions against."}],"review_version":1}