LPMLN, Weak Constraints, and P-log

Joohyung Lee; Zhun Yang

REVIEW 2 major objections 4 minor 18 references

LPMLN, Weak Constraints, and P-log

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2506.12784 v1 pith:E2ZFABLG submitted 2025-06-15 cs.AI

Joohyung Lee , Zhun Yang This is my paper · ORCID

classification cs.AI MSC 68T2768N1768T37

keywords LPMLNweakconstraintsP-loganswersetprogrammingMAPinferenceprobabilisticlogicstablemodelsMarkov

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

The 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.

I 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.

Referee Report

2 major / 4 minor

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).

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 (2)

[Review: P-log, footnote 8; Turning P-log into LPMLN, rule (8)] 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.
[Corollary 2 and the surrounding translation lpmln2wcpnt,rule] 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.

minor comments (4)

[Example 3 Continued] 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.
[Abstract and Introduction] 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.
[Review: P-log, notation] 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.
[Alternative Translations, Corollary 2] 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.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the LPMLN-to-weak-constraints result is a direct algebraic equivalence, and the P-log-to-LPMLN theorem is a self-contained encoding proof; the footnote-8 change to P-log's 'possible' semantics is a disclosed scope limitation, not a circular step.

full rationale

The two main theorems do not assume their conclusions. Theorem 1 is proved by first showing (Proposition 1) that SM[Π] equals the stable models of lpmln2wc(Π), then by a chain of equalities that argmax of exponentiated LPMLN weights equals argmin of weak-constraint penalties; the hard/soft priority is matched by levels. This is a reduction between independently defined formalisms, not a fit of the conclusion into the premises. Theorem 2 inherits the MLN-to-LPMLN embedding from (Lee and Wang 2016) as a prior result with its own proof, so it is not a self-citation chain. Theorem 3 is a modular encoding: the soft weights ln(p), ln(1/m), and ln(x) are chosen to reproduce the P-log probability factors, and the proof verifies each factor and the stable-model correspondence. This is translational correctness work, not circularity. The one caveat is footnote 8: the paper redefines 'possible' by adding ¬Intervene(c(u)), and rule (8) encodes exactly that; in the original Baral-Gelfond-Rushton semantics an intervened atom still receives a default probability. Consequently the 'full P-log' claim in the introduction is not established for intervention scenarios under BGR 2009. This is a scope/correctness overclaim, disclosed by the authors, and it does not make the derivation circular: the theorem is self-contained relative to the paper's stated variant. No load-bearing self-citation was found; the cited LPMLN foundations are prior, parameter-free results outside the target claim.

Assumptions & free parameters 0 free parameters · 5 assumptions · 2 invented entities

No free parameters are fitted; all numeric weights come from the source P-log or LPMLN programs. The central results rely on the semantics of LPMLN and P-log as imported from the cited literature, plus the paper's stated modification of when an attribute is possible under intervention. The translation-only auxiliary predicates add descriptive vocabulary but not new empirical entities.

assumptions (5)

standard math Stable model semantics for first-order formulas, including the Ferraris-Lifschitz reduct and splitting/completion results.
Invoked throughout the appendix proofs, for example Lemma 1 from Bartholomew and Lee 2013 and the splitting theorem from Ferraris, Lee, Lifschitz, and Palla 2009.
domain assumption LPMLN semantics with real weights and the alpha-limit for hard rules, following Lee and Wang 2016.
Theorems 1 and 3 are stated inside this semantics; Lemma 4 imports Proposition 2 of Lee and Wang 2016.
domain assumption P-log semantics of Baral, Gelfond, and Rushton 2009, including random selection rules, pr-atoms, observations, actions, and Conditions 1 and 2, with the paper's modified definition of possible for intervened attributes.
The P-log translation is verified against this semantics; footnote 8 changes the original definition of possible in W regarding Intervene.
domain assumption Finite Herbrand universe for grounding LPMLN programs.
Stated in the LPMLN review; needed for grounding and for representing aggregates by finite formulas.
domain assumption Real-valued weights and #sum and #count aggregates are expressible in the formal encodings, and practical ASP solvers require integer approximation of real weights.
The solver-facing translation in Corollary 2 depends on approximating real weights to integers, a practical assumption that is acknowledged but not given formal error bounds.

invented entities (2)

Auxiliary atoms unsat(i) in lpmln2wcpnt,rule
purpose: Mark violated rules so weak constraints can penalize them
Introduced by the translation; they have no truth value on the original signature and no testable consequence outside the encoding.
Auxiliary predicates Poss, PossWithAssPr, AssPr, PossWithDefPr, NumDefPr, RemPr, and TotalDefPr in plog2lpmln
purpose: Track possible values, assigned probabilities, default denominators, and default numerators in the LPMLN encoding of P-log
They are internal bookkeeping predicates; the one-to-one correspondence fixes their values from the P-log possible world, and they have no independent empirical content.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LPMLN, Weak Constraints, and P-log." pith.science (2026). https://pith.science/paper/E2ZFABLG

@misc{pith2026250612784,
  author       = {Pith},
  title        = {Pith review of: LPMLN, Weak Constraints, and P-log},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E2ZFABLG}},
  note         = {Machine review of arXiv:2506.12784}
}

read the original abstract

LPMLN is a recently introduced formalism that extends answer set programs by adopting the log-linear weight scheme of Markov Logic. This paper investigates the relationships between LPMLN and two other extensions of answer set programs: weak constraints to express a quantitative preference among answer sets, and P-log to incorporate probabilistic uncertainty. We present a translation of LPMLN into programs with weak constraints and a translation of P-log into LPMLN, which complement the existing translations in the opposite directions. The first translation allows us to compute the most probable stable models (i.e., MAP estimates) of LPMLN programs using standard ASP solvers. This result can be extended to other formalisms, such as Markov Logic, ProbLog, and Pearl's Causal Models, that are shown to be translatable into LPMLN. The second translation tells us how probabilistic nonmonotonicity (the ability of the reasoner to change his probabilistic model as a result of new information) of P-log can be represented in LPMLN, which yields a way to compute P-log using standard ASP solvers and MLN solvers.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages

[1]

Hinge-Loss Markov Random Fields and Probabilistic Soft Logic

Bach, S. H.; Broecheler, M.; Huang, B.; and Getoor, L. 2015. Hinge-loss markov random fields and probabilistic soft logic. arXiv:1505.04406 [cs.LG]

work page Pith review arXiv 2015
[2]

Balai, E., and Gelfond, M. 2016. On the relationship between P -log and L P ^ M L N . In Proceedings of International Joint Conference on Artificial Intelligence ( IJCAI )

work page 2016
[3]

Baral, C.; Gelfond, M.; and Rushton, J. N. 2009. Probabilistic reasoning with answer sets. TPLP 9(1):57--144

2009
[4]

Buccafurri, F.; Leone, N.; and Rullo, P. 2000. Enhancing disjunctive datalog by constraints. Knowledge and Data Engineering, IEEE Transactions on 12(5):845--860

2000
[5]

Calimeri, F.; Faber, W.; Gebser, M.; Ianni, G.; Kaminski, R.; Krennwallner, T.; Leone, N.; Ricca, F.; and Schaub, T. 2013. A S P - C ore-2 input language format

work page 2013
[6]

De Raedt, L.; Kimmig, A.; and Toivonen, H. 2007. P rob L og: A probabilistic prolog and its application in link discovery. In IJCAI , volume 7, 2462--2467

2007
[7]

Ferraris, P.; Lee, J.; and Lifschitz, V. 2011. Stable models and circumscription. Artificial Intelligence 175:236--263

work page 2011
[8]

Ferraris, P. 2011. Logic programs with propositional connectives and aggregates. ACM Transactions on Computational Logic (TOCL) 12(4):25

work page 2011

Show all 18 references

[9]

Gelfond, M., and Lifschitz, V. 1988. The stable model semantics for logic programming. In Kowalski, R., and Bowen, K., eds., Proceedings of International Logic Programming Conference and Symposium , 1070--1080. MIT Press

1988
[10]

J.; Lifschitz, V.; and Yang, F

Harrison, A. J.; Lifschitz, V.; and Yang, F. 2014. The semantics of gringo and infinitary propositional formulas. In Principles of Knowledge Representation and Reasoning: Proceedings of the Fourteenth International Conference, KR 2014

2014
[11]

Lee, J., and Meng, Y. 2012. Stable models of formulas with generalized quantifiers (preliminary report). In Technical Communications of the 28th International Conference on Logic Programming , 61--71

2012
[12]

Lee, J., and Wang, Y. 2016. Weighted rules under the stable model semantics. In Proceedings of International Conference on Principles of Knowledge Representation and Reasoning (KR) , 145--154

2016
[13]

Lee, J.; Meng, Y.; and Wang, Y. 2015. Markov logic style weighted rules under the stable model semantics. In Technical Communications of the 31st International Conference on Logic Programming

2015
[14]

Nickles, M., and Mileo, A. 2014. Probabilistic inductive logic programming based on answer set programming. In 15th International Workshop on Non-Monotonic Reasoning (NMR 2014)

2014
[15]

Pearl, J. 2000. Causality: models, reasoning and inference , volume 29. Cambridge Univ Press

2000
[16]

Richardson, M., and Domingos, P. 2006. Markov logic networks. Machine Learning 62(1-2):107--136

2006
[17]

On the stable model semantics for intensional functions

Bartholomew, Michael and Lee, Joohyung 2013. On the stable model semantics for intensional functions. Cambridge University Press

2013
[18]

Symmetric splitting in the general theory of stable models

Paolo Ferraris and Joohyung Lee and Vladimir Lifschitz and Ravi Palla 2009. Symmetric splitting in the general theory of stable models. Proceedings of International Joint Conference on Artificial Intelligence (IJCAI)

2009

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Hinge-Loss Markov Random Fields and Probabilistic Soft Logic

Bach, S. H.; Broecheler, M.; Huang, B.; and Getoor, L. 2015. Hinge-loss markov random fields and probabilistic soft logic. arXiv:1505.04406 [cs.LG]

work page Pith review arXiv 2015

[2] [2]

Balai, E., and Gelfond, M. 2016. On the relationship between P -log and L P ^ M L N . In Proceedings of International Joint Conference on Artificial Intelligence ( IJCAI )

work page 2016

[3] [3]

Baral, C.; Gelfond, M.; and Rushton, J. N. 2009. Probabilistic reasoning with answer sets. TPLP 9(1):57--144

2009

[4] [4]

Buccafurri, F.; Leone, N.; and Rullo, P. 2000. Enhancing disjunctive datalog by constraints. Knowledge and Data Engineering, IEEE Transactions on 12(5):845--860

2000

[5] [5]

Calimeri, F.; Faber, W.; Gebser, M.; Ianni, G.; Kaminski, R.; Krennwallner, T.; Leone, N.; Ricca, F.; and Schaub, T. 2013. A S P - C ore-2 input language format

work page 2013

[6] [6]

De Raedt, L.; Kimmig, A.; and Toivonen, H. 2007. P rob L og: A probabilistic prolog and its application in link discovery. In IJCAI , volume 7, 2462--2467

2007

[7] [7]

Ferraris, P.; Lee, J.; and Lifschitz, V. 2011. Stable models and circumscription. Artificial Intelligence 175:236--263

work page 2011

[8] [8]

Ferraris, P. 2011. Logic programs with propositional connectives and aggregates. ACM Transactions on Computational Logic (TOCL) 12(4):25

work page 2011