{"id":"0b1faebd-38db-4b25-b9ea-5f9160aad30e","arxiv_id":"2506.12784","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LPMLN programs translate into answer set programs with weak constraints so that MAP models coincide, and full P-log programs translate into LPMLN so that possible worlds and probabilities are preserved.","lead":"Two translation theorems connect three formalisms for probabilistic logic programming: LPMLN, weak constraints, and P-log. The translations let standard answer set solvers compute maximum probability models and embed the full P-log language into LPMLN.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'full P-log' claim in Theorem 3 rests on a modified possibility semantics that diverges from Baral-Gelfond-Rushton for intervened atoms; the translation does not verify the original P-log semantics.","rationale":"The paper's central contribution is two-fold: reducing LPMLN MAP inference to weak constraint optimization, and characterizing P-log in LPMLN. The first part (Theorem 1) is a clean, well-supported translation and I found no internal flaw in its argument. The second part (Theorem 3) is the load-bearing claim for 'full P-log'. The reader's verdict identifies the weakest assumption precisely: the paper silently revises the definition of 'possible' in footnote 8, excluding intervened atoms from the probability computation. This is not just a cosmetic change; it removes a probability factor that the original P-log semantics would assign to the intervened attribute. The LPMLN translation's rule (8) encodes the modified definition, and the proof of Theorem 3 relies on that definition throughout. Consequently, the theorem establishes equivalence with a modified P-log, not with the semantics of Baral, Gelfond, and Rushton 2009, despite the abstract claiming a 'full P-log' characterization. This is a genuine overclaim, but it is not fatal to the underlying technical work: the translation is plausibly correct for the modified semantics, and the paper could be repaired by explicitly proposing the semantics change as a contribution and qualifying the theorem statement. The reader's CONDITIONAL verdict is therefore appropriate; my analysis does not move the verdict. I considered other potential issues, such as the use of α as a weak-constraint weight in Corollary 2, but that is a minor representational slip rather than a central correctness failure. The footnote-8 divergence is the single most load-bearing concern because it directly attacks the target of the main theorem.","tokens_in":33015,"tokens_out":20641,"duration_ms":239123,"concrete_test":"Run the following analytical/computational check. Take a P-log program with two Boolean attributes A and B, random rules random(A) and random(B), an action Do(A=a), and a random rule for A whose body contains B=c and whose possible-value filter p(x) is such that when B=c the only possible value is a and when B=d the possible values are {a,b}. Give A a pr-atom with condition depending on B (or none), so the default probability of A=a differs between B=c and B=d under the original semantics. Compute the normalized probabilities of the two possible worlds (A=a,B=c) and (A=a,B=d) under (i) the original Baral-Gelfond-Rushton semantics, (ii) the paper's modified semantics, and (iii) plog2lpmln(Π). If (iii) equals (ii) but not (i), Theorem 3 does not cover the original P-log for intervention scenarios.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In 'Review: P-log', footnote 8, the paper changes the definition of 'possible' to require ¬Intervene(c(u)). In the original P-log semantics (Baral, Gelfond, Rushton 2009), an intervened attribute still receives a probability (the default probability). The LPMLN encoding inherits this change in rule (8), 'Poss_r(c(u)=v) ← Body, p(v), not Intervene(c(u))'. Since the probability factors in Theorem 3 are attached only to atoms that are 'possible', any stable model corresponding to a world with Do(c(u)=v) omits the probability contribution that the original semantics assigns to c(u)=v. Thus Theorem 3 establishes a 1-1 correspondence and probability equality for the paper's modified P-log, not for the P-log of Baral et al. The abstract and introduction nonetheless claim the translation 'applies to full P-log' and 'completely characterizes' P-log. This is an overclaim for exactly the intervention scenarios that distinguish the semantics, so the central claim is not established for the intended target language unless the target semantics is explicitly revised.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two translations linking LPMLN with weak constraints and with P-log. The first translation, lpmln2wc, turns any ground LPMLN program into an ASP program with weak constraints so that the most probable stable models of the LPMLN program coincide with the optimal stable models of the weak-constraint program (Theorem 1); the paper also gives variants suitable for ASP solver input (Corollaries 1–4) and derives a corresponding reduction for Markov Logic (Theorem 2). The second translation, plog2lpmln, maps a P-log program to an LPMLN program and claims a one-to-one, probability-preserving correspondence between nonzero-probability possible worlds and probabilistic stable models (Theorem 3). The paper includes worked examples, including a four-door Monty Hall variant, and an appendix with detailed proofs using splitting, completion, and results from Lee and Wang (2016).","tokens_in":33272,"tokens_out":6303,"duration_ms":92266,"significance":"If the claims hold, the paper has practical and theoretical value: it reduces MAP inference for LPMLN—and, via known embeddings, for Markov Logic, ProbLog, and Pearl's causal models—to weak-constraint optimization in standard ASP solvers, and it provides a modular embedding of P-log into LPMLN. The LPMLN-to-weak-constraints direction is simple and appears to be a genuine, useful bridge, and the appendix contains substantial proof machinery rather than only examples. The main caveat is that the P-log direction is proved for a modified notion of 'possible' that diverges from the original Baral-Gelfond-Rushton semantics in intervention scenarios, so the scope of the claimed 'full P-log' characterization needs to be made precise.","major_comments":[{"comment":"The paper changes the definition of when an atom c(u)=v is possible in a possible world W by requiring not Intervene(c(u)), as stated in footnote 8 and implemented in rule (8), Poss_r(c(u)=v) <- Body, p(v), not Intervene(c(u)). Footnote 8 explicitly acknowledges that in the original P-log semantics of Baral, Gelfond, and Rushton (2009), an intervened attribute still receives a probability determined by the default probability. Since Theorem 3 and its mapping phi are built on this modified definition, the theorem establishes a correspondence for the paper's revised P-log semantics, not for the original semantics. The abstract and Section 1 nevertheless claim that the translation 'applies to full P-log' and 'completely characterizes' P-log. This overclaim is load-bearing because intervention scenarios are exactly the cases that distinguish the two semantics. The authors should either prove Theorem 3 under the original definition of 'possible' or explicitly declare the modified semantics as the intended target and revise the abstract and introduction accordingly.","section":"Review: P-log, footnote 8; Turning P-log into LPMLN, rule (8)"},{"comment":"The weak-constraint rules in Corollary 2 use the symbol alpha as a weight for hard formulas, as in ':∼ unsat(i) [wi@l]' with l=1 when wi is alpha. In the weak-constraint syntax of Buccafurri, Leone, and Rullo (2000) and ASP-Core-2, weights are real numbers, and alpha is not a real number. The intended reading is presumably that any sufficiently large finite weight at the higher level dominates all level-0 penalties, but this should be stated formally. This is a correctness detail for the solver-oriented corollaries rather than for the main equivalence in Theorem 1, but it should be fixed or clarified.","section":"Corollary 2 and the surrounding translation lpmln2wcpnt,rule"}],"minor_comments":[{"comment":"In the line 'α : ⊥ ←RemDefPr(Prize, x), x≤ 0', the predicate name appears to be a typo for RemPr, since RemDefPr is not defined elsewhere.","section":"Example 3 Continued"},{"comment":"The phrases 'applies to full P-log' and 'completely characterized in LPMLN' should be qualified in light of the modified possibility definition; this is related to the first major comment, but the abstract and introduction should at least mention the modification even if the authors choose to keep it.","section":"Abstract and Introduction"},{"comment":"The symbol σ is used both for the signature of τ(Π) in Theorem 3 and for the components σ1 and σ2 of a P-log signature; this overloading makes the appendix harder to follow. A distinct symbol for the signature of τ(Π) would improve readability.","section":"Review: P-log, notation"},{"comment":"The proof of Corollary 2 relies on Lemma 1 and the Splitting Theorem, but Lemma 1 is introduced only in the appendix; a forward reference or a brief statement in the main text would help the reader understand the proof strategy.","section":"Alternative Translations, Corollary 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically substantial and the LPMLN-to-weak-constraints result appears sound and useful. The main concern for the editor is the scope of the P-log claim: the authors have knowingly modified the semantics of 'possible' in a footnote and then asserted 'full P-log' characterization. This is fixable by revising the claims or by proving the result for the original semantics, but it is load-bearing for the contribution. I would not reject the paper on this basis, since the internal proof is coherent under the stated definition and the modification is disclosed; however, the public claims overreach."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new result is the inverse translation from LPMLN to weak constraints. It is simple in the right way: each weighted formula becomes a choice formula and a weak constraint whose penalty is the negated weight, hard rules sit at a higher level so they dominate, and the appendix proof is an algebraic verification plus Proposition 1. The worked examples match hand computation, and the rule-form variant with unsat atoms is plausible as an actual CLINGO input pipeline. The practical claim attached to it—MAP inference for LPMLN, and through known embeddings for MLN, ProbLog, and Pearl nets, reduces to weak-constraint optimization—is supported by the theorem, though no implementation or benchmark appears. That is a minor gap for a theory paper.\n\nThe P-log-to-LPMLN encoding is the more ambitious half. It handles dynamic default probabilities, which the earlier Lee-Wang fragment did not, and the Monty Hall example checks out. The proof is long and uses splitting and completion in a way that seems legitimate, and the auxiliary predicates really do carry the numerator and denominator of the default probability. I did not find circularity: the external lemmas are prior results, not the target conclusions.\n\nThe soft spot is real and is exactly where the stress test points. In the review of P-log, footnote 8 changes the definition of 'possible' by requiring ¬Intervene(c(u)). The original Baral-Gelfond-Rushton semantics, as the footnote itself admits, gives an intervened attribute its default probability. Rule (8) and the probability factors in Theorem 3 therefore apply to a revised P-log semantics, not to BGR's P-log. The abstract says the translation 'applies to full P-log' and 'completely characterizes' it; for programs with Do actions, that is not established. If the authors reframed this as a deliberate repair of the intervention semantics, the theorem would be fine. As it stands, the target semantics is not the one cited.\n\nI would send this to a serious referee. The weak-constraint theorem is publishable on its own; the P-log half needs either a statement of the semantic revision or an argument that the original semantics behaves as the paper defines. I would also ask for one small experiment or at least a note that the solver-ready translation has been run.","headline":"Worth engaging: the LPMLN-to-weak-constraints translation is a clean, useful inverse result, and the P-log encoding is serious work, but the 'full P-log' claim overstates what is proved because the paper quietly revises the intervention semantics.","tokens_in":33798,"tokens_out":4473,"would_cite":true,"duration_ms":52683,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T27","68N17","68T37"],"pacs":[],"model":"deepseek-v4-flash","headline":"Two exact translations connect LPMLN to weak constraints and to P-log: MAP estimates reduce to optimal stable models, and P-log's possible worlds map one-to-one into LPMLN with unchanged probabilities.","keywords":["LPMLN","weak constraints","P-log","answer set programming","MAP inference","probabilistic logic programming","stable models","Markov Logic"],"falsifier":"Take a P-log program with one random Boolean attribute $c$, no probability atoms, and the action $Do(c=t)$. The original P-log semantics gives the single possible world a nonzero default probability $1/2$ for $c=t$; the translation $\\mathrm{plog2lpmln}$ treats $c=t$ as not possible because $Intervene(c)$ holds, so the world receives no default-probability factor and its normalized probability differs. Checking this program against the original P-log semantics would settle whether the correspondence holds for interventions.","tokens_in":32807,"feed_emoji":"🎲","tokens_out":13000,"duration_ms":134659,"temperature":0.7,"pith_summary":"This paper establishes two exact bridges around LPMLN, a weighted extension of answer set programming. It proves that the most probable stable models of any LPMLN program are exactly the optimal stable models of a program with weak constraints, obtained by replacing every weighted rule with a choice rule and a penalty constraint. It also proves that a consistent P-log program has a one-to-one correspondence into LPMLN that preserves possible worlds and their probabilities, under the paper's adjusted definition of 'possible.' If the results hold, MAP estimation for LPMLN and, through known embeddings, for Markov Logic, ProbLog, and causal models, can be done with ordinary ASP solvers, and P-log can be computed the same way.","feed_headline":"LPMLN's top models become weak-constraint optima","feed_subtitle":"A second encoding puts P-log inside LPMLN, so standard ASP solvers can handle both formalisms.","key_machinery":"The load-bearing object in the first translation is the choice formula $\\{F\\}^{\\mathrm{ch}}$, which lets a stable model decide whether to satisfy $F$ while a weak constraint $:\\sim F\\,[-w@L]$ subtracts $w$ from the penalty; two priority levels, 1 for hard rules and 0 for soft rules, mirror the lexicographic ordering of LPMLN's infinite and finite weights. For P-log, the machinery is a layer of auxiliary predicates $\\mathrm{Poss}$, $\\mathrm{PossWithAssPr}$, $\\mathrm{PossWithDefPr}$, $\\mathrm{NumDefPr}$, $\\mathrm{RemPr}$, and $\\mathrm{TotalDefPr}$, plus soft rules of the form $\\ln(p):\\bot\\leftarrow\\neg A$ that contribute multiplicative factors to the weight of a stable model. Aggregates $\\#\\mathrm{count}$ and $\\#\\mathrm{sum}$ compute the denominator and numerator of the default-probability formula.","core_discovery":"The central discovery is a pair of exact equivalences. Theorem 1 says that for every LPMLN program $\\Pi$, the stable models with maximum normalized weight are precisely the optimal stable models of the weak-constraint program $\\mathrm{lpmln2wc}(\\Pi)$; the translation turns each weighted formula $w:F$ into a choice formula $\\{F\\}^{\\mathrm{ch}}$ plus $:\\sim F\\,[-w@L]$, with level 1 recording hard rules and level 0 recording soft weights. Theorem 3 says that for any consistent P-log program $\\Pi$, $\\mathrm{plog2lpmln}(\\Pi)$ has probabilistic stable models in one-to-one correspondence with the non-zero-probability possible worlds of $\\Pi$, and corresponding worlds have the same probability. Together they show that the optimization part of LPMLN is just two-level weak-constraint optimization, and that P-log's assigned and default probabilities can be encoded as weighted rules.","pith_inferences":["If the modified 'possible atom' definition is accepted, the P-log translation implies that intervention behaves as a logical constraint rather than a probabilistic choice, because an intervened attribute contributes no default-probability factor; the paper does not draw out that consequence.","The choice-rule plus two-level-penalty pattern is a general recipe: any finite probabilistic logic whose weight is the exponential of a sum of satisfied formula weights can likely be compiled to weak-constraint optimization in the same style.","The P-log encoding shows that default probabilities with renormalized denominators can be expressed through soft rules guarded by aggregates, without probabilistic choice atoms; this pattern may transfer to other distribution-semantics languages, though the paper does not claim that.","A direct numerical comparison on a small intervention program would settle whether the Footnote 8 modification changes P-log's intended answers in practice."],"forward_implications":["MAP inference for LPMLN programs reduces to finding optimal stable models of a weak-constraint program, so any ASP solver supporting weak constraints can perform it.","The same reduction applies to Markov Logic, ProbLog, and causal models through their known embeddings into LPMLN, giving a common solver-based route to MAP estimates.","Combining the two translations, P-log inference reduces first to LPMLN and then to weak-constraint optimization, so P-log can be evaluated by standard ASP solvers.","Hard rules are protected by placing their penalties at a higher weak-constraint level, so the translation preserves the priority of definite knowledge over soft preferences."],"supporting_citations":[{"why":"Defines LPMLN semantics and proves the embeddings of Markov Logic and ProbLog into LPMLN that the extensions of Theorem 1 depend on.","marker":"Lee and Wang 2016"},{"why":"Introduces weak constraints and optimal stable models, the target formalism of the first translation.","marker":"Buccafurri, Leone, and Rullo 2000"},{"why":"Defines P-log syntax, possible worlds, and the probability scheme that the second translation must reproduce; its semantics is the one modified in Footnote 8.","marker":"Baral, Gelfond, and Rushton 2009"},{"why":"Supplies the first-order stable model semantics that allows rules to be treated as formulas, justifying choice formulas and the splitting arguments.","marker":"Ferraris, Lee, and Lifschitz 2011"},{"why":"Defines Markov Logic, whose reduction to LPMLN extends the weak-constraint MAP result.","marker":"Richardson and Domingos 2006"},{"why":"Defines ProbLog, another formalism whose LPMLN embedding inherits the MAP reduction.","marker":"De Raedt, Kimmig, and Toivonen 2007"},{"why":"Provides the embedding of causal models into LPMLN used to extend the weak-constraint MAP reduction.","marker":"Lee, Meng, and Wang 2015"},{"why":"Provides the complementary LPMLN-to-P-log translation that frames this paper's P-log-to-LPMLN direction.","marker":"Balai and Gelfond 2016"}],"fun_headline_variants":["LPMLN's MAP models are weak-constraint optima","P-log's worlds match LPMLN's probabilistic stable models","Two translations: LPMLN to weak constraints, P-log to LPMLN","LPMLN's MAP equals weak-constraint optima","P-log inside LPMLN: same probabilities, computable via ASP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an intervened random attribute should no longer count as a possible outcome, so it contributes no default-probability factor; the original P-log semantics instead gives such outcomes a default probability, and the equivalence is proved only under the paper's modified version.","fun_headline_variants_meta":{"raw":{"variants":["LPMLN's MAP models are weak-constraint optima","P-log's worlds match LPMLN's probabilistic stable models","Two translations: LPMLN to weak constraints, P-log to LPMLN","LPMLN's MAP equals weak-constraint optima","P-log inside LPMLN: same probabilities, computable via ASP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3798,"prompt_tokens":932,"completion_tokens":2866,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2773}},"tokens_in":548,"tokens_out":2866,"duration_ms":23288,"temperature":1.0,"reasoning_tokens":2773,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:43:29.875696+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a P-log program with one random Boolean attribute $c$, no probability atoms, and the action $Do(c=t)$. The original P-log semantics gives the single possible world a nonzero default probability $1/2$ for $c=t$; the translation $\\mathrm{plog2lpmln}$ treats $c=t$ as not possible because $Intervene(c)$ holds, so the world receives no default-probability factor and its normalized probability differs. Checking this program against the original P-log semantics would settle whether the correspondence holds for interventions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines LPMLN semantics and proves the embeddings of Markov Logic and ProbLog into LPMLN that the extensions of Theorem 1 depend on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the complementary LPMLN-to-P-log translation that frames this paper's P-log-to-LPMLN direction."}],"review_version":1}