Praline: Dual Systems for Probabilistic Inference

• Praline refers to two distinct probabilistic reasoning systems: one using Markov Logic Networks for elementary-science QA and another employing a Datalog extension for precise inference with correlated inputs.
• The MLN-based system integrates lexical and structural alignments with controlled proof dependencies, acyclicity, and false-unless-proven semantics to enhance answer accuracy and runtime efficiency.
• The Datalog-based variant leverages correlation classes and constrained optimization to compute exact or δ-exact probability bounds, providing rigorous guarantees even with partially known input correlations.

Searching arXiv for the cited Praline papers and closely related context. Praline is the name of two distinct probabilistic reasoning systems in the arXiv literature. In "Markov Logic Networks for Natural Language Question Answering," it denotes PRobabilistic ALignment and INferencE, the third and most effective Markov Logic Network formulation proposed for elementary-science multiple-choice question answering; it uses MLNs to align lexical elements and to define and control how inference should be performed (Khot et al., 2015). In "Probabilistic Inference for Datalog with Correlated Inputs," it denotes a probabilistic extension to Datalog designed for precise probabilistic inference in the presence of partially known input correlations; a program is written as $\mathcal{P}=(\mathcal{C},R,\rho_R,\rho_I)$ and yields exact or $\delta$-exact probability bounds for output facts (Wang et al., 21 Aug 2025).

1. Terminological scope and disambiguation

A common source of confusion is nominal rather than technical: the name Praline refers here to two different formal systems with different substrates, tasks, and semantics.

Variant	Formal substrate	Stated goal
Praline (PRobabilistic ALignment and INferencE)	Markov Logic Networks	Elementary-science question answering
Praline	Probabilistic Datalog	Precise inference with correlated inputs

The 2015 system is a question-answering formulation over knowledge extracted automatically from science textbooks and expressed in a subset of first-order logic. The 2025 system is a Datalog extension in which rule probabilities are independent, input facts are partitioned into correlation classes, and output probabilities are computed as exact values or intervals depending on how fully the input correlations are specified. The shared name therefore does not imply a shared implementation or shared semantics (Khot et al., 2015, Wang et al., 21 Aug 2025).

2. The MLN-based Praline for natural-language question answering

In the 2015 formulation, Praline is the third of three MLN-based approaches investigated for elementary-science QA. The preceding approaches either used extracted science rules directly as MLN clauses, which produced an unmanageable number of groundings because the domain had long and complex rules, or interpreted science rules as prototypical entities mapped directly to grounded MLN assertions and clustered via existing entity resolution methods, which simplified the network but remained brittle. Praline instead uses MLNs both to perform alignment and to regulate inference, and this change in semantics is central to its reported effectiveness (Khot et al., 2015).

Its combined graph contains question elements and KB-rule elements. The relevant predicates are:

• Structural predicates: node(n), edge(n_1,n_2,\ell), setup(n), query(n), inLhs(n,r), and inRhs(n,r).
• Random variables over content and proof state: holds(n), aligns(n_1,n_2), lhsHolds(r), rhsHolds(r), proves(n_1,n_2), and ruleProves(r_1,r_2).

Every soft clause contributes a log-potential $w\cdot I[\text{clause is true}]$. The weighted clauses are organized into six groups. First, Praline introduces initial alignment based on lexical entailment by attaching soft evidence $\rho_{xy}:\operatorname{aligns}(x,y)$ to node pairs with entailment score $\rho_{xy}$. Second, it introduces structural alignment via four propagation clauses, with higher weights when edge labels match and lower weights when they differ. Third, it enables propagation of holds via alignment through

$$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$

Fourth, it models rule firing by decomposing each KB rule $r$ into four binary clauses of weight $w_4$:

$$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$

$$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$

$\delta$0

$\delta$1

Fifth, it adds acyclicity constraints at both node and rule levels. The proof layer is built from proves and ruleProves, together with transitivity clauses and the hard constraints $\delta$2 and $\delta$3. Sixth, it enforces false-unless-proven (FUP) through the hard constraint

$\delta$4

so that a node can hold only if it is given in the setup or derived by firing some rule. Given evidence $\delta$5, inference computes marginal probabilities $\delta$6 for query nodes, and the selected answer option is the one whose query nodes attain the highest joint probability (Khot et al., 2015).

3. Probabilistic alignment and explicit control of inference

Praline’s alignment mechanism couples lexical similarity with graph consistency. Lexical entailment $\delta$7 is computed via WordNet-based or entailment-classifier scores and then used as soft evidence on aligns(x,y). Structural consistency is enforced by the four propagation clauses that push alignment down from parents to children and symmetrically up from children to parents, with higher weight $\delta$8 when edge labels match and lower weight $\delta$9 otherwise. This jointly ties local lexical similarity to consistency of entire subgraphs (Khot et al., 2015).

The system’s control over inference is equally explicit. Rule chaining is mediated by the binary predicates lhsHolds(r) and rhsHolds(r), rather than by direct unrestricted implication from antecedent strings to consequent strings. Clause (3), $w\cdot I[\text{clause is true}]$0, means that only if all soft evidence for the antecedent is positive will the consequent become likely. The proof-graph layer prevents feedback loops in which one rule supports a fact used by another rule that then re-supports the first. FUP then hard-wires the Problog-style semantics that nothing is true unless proven, replacing the default MLN tendency to assign probability $w\cdot I[\text{clause is true}]$1 to unmentioned atoms.

The high-level algorithmic procedure for one multiple-choice question begins by building a graph $w\cdot I[\text{clause is true}]$2 from string-constant nodes in the setup, query, and all candidate KB rules. Hard evidence asserts setup(n)=true and holds(n)=true for setup nodes, marks query nodes, and records all extracted edge, inLhs, and inRhs facts. Soft evidence is added for lexical alignment scores. The Praline clauses are then added, the MLN is grounded, hard-constraint backbone reduction is applied to fix variables implied by the setup, and MC-SAT or other MLN inference estimates marginals such as $w\cdot I[\text{clause is true}]$3 and $w\cdot I[\text{clause is true}]$4. For answer choice $w\cdot I[\text{clause is true}]$5, the score is

$w\cdot I[\text{clause is true}]$6

and the prediction is $w\cdot I[\text{clause is true}]$7 (Khot et al., 2015).

4. Experimental behavior of the QA formulation

The reported experiments use Dev-108, consisting of 108 4th-grade science multiple-choice questions without diagrams, and Unseen-68, a held-out set of 68 questions. The KB contains approximately 47,000 automatically extracted IF-THEN sentences from texts and web search. The rule selector chooses the top 30 KB sentences by word overlap. Entailment uses WordNet-derived $w\cdot I[\text{clause is true}]$8-scores converted to soft evidence. The solver setup is Tuffy+Glucose grounding-reduction with a 10 minute total limit, 6 minutes for grounding, and MC-SAT with 500 samples (Khot et al., 2015).

On Dev-108, FO-MLN answered 82/108 questions, achieved 33.6% accuracy, used 35 rules on average, produced 524 ground clauses, and took 280 seconds per question on average. ER-MLN answered 107/107, achieved 34.5% accuracy, used 41 rules, produced 2308 ground clauses, and took 188 seconds. Praline answered 108/108, achieved 48.8% accuracy, used 51 rules, produced 219 ground clauses, and took 17 seconds. On Unseen-68, Praline scored 46.3% versus approximately 33% for the baselines and ran approximately $w\cdot I[\text{clause is true}]$9 faster than ER-MLN or FO-MLN grounding plus inference.

The ablation study isolates the role of chaining, acyclicity, and FUP. With one rule, Praline achieved 48.8% on Dev-108 and 46.3% on Unseen-68; with chain length 2, it achieved 50.3% and 52.7%. Removing acyclicity reduced performance to 44.7% and 36.0% for one rule, and to 43.6% and 30.9% for chain length 2. Removing FUP yielded 35.0% and 30.9% for one rule, and 42.1% and 29.4% for chain length 2. Removing both FUP and acyclicity yielded 37.3% and 34.2% for one rule, and 36.6% and 24.3% for chain length 2. The paper therefore attributes the empirical gains to node-level modeling of holds, the joint alignment-plus-inference MLN, and the explicit suppression of spurious reinforcement loops and unsupported facts. In the authors’ summary, these elements yield a +15 percentage point accuracy improvement over earlier FO-MLN and ER-MLN formulations together with a $\rho_{xy}:\operatorname{aligns}(x,y)$0 runtime speed-up (Khot et al., 2015).

5. The Datalog-based Praline: syntax, semantics, and exact inference

The 2025 Praline is a probabilistic-Datalog extension for handling partially known correlations among input facts. A program is a tuple

$\rho_{xy}:\operatorname{aligns}(x,y)$1

where $\rho_{xy}:\operatorname{aligns}(x,y)$2 is a partition of input facts into correlation classes, $\rho_{xy}:\operatorname{aligns}(x,y)$3 is a set of Datalog rules whose heads are output predicates, $\rho_{xy}:\operatorname{aligns}(x,y)$4 assigns a probability to each rule, and $\rho_{xy}:\operatorname{aligns}(x,y)$5 specifies input-fact probabilities and conditional probabilities. Rules are mutually independent. A rule of the form

$\rho_{xy}:\operatorname{aligns}(x,y)$6

means

$\rho_{xy}:\operatorname{aligns}(x,y)$7

and an input declaration

$\rho_{xy}:\operatorname{aligns}(x,y)$8

means

$\rho_{xy}:\operatorname{aligns}(x,y)$9

Absent any rule linking two inputs, they are assumed independent (Wang et al., 21 Aug 2025).

The semantics is given by possible worlds. After grounding all rules and facts into a finite set of ground rules $\rho_{xy}$0, a possible world $\rho_{xy}$1 chooses a subset of the probabilistic rules consistent with $\rho_{xy}$2 and $\rho_{xy}$3. A joint distribution $\rho_{xy}$4 over worlds must satisfy the rule marginals $\rho_{xy}$5 and the conditional-probability constraints induced by the input declarations. For an output fact $\rho_{xy}$6,

$\rho_{xy}$7

If all conditional probabilities in $\rho_{xy}$8 fully determine each correlation class’s joint distribution, $\rho_{xy}$9 is a singleton; otherwise it is an interval.

Partially known correlations are encoded by introducing, for each correlation class $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$0 of size $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$1, the $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$2 joint-probability variables

$$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$3

subject to $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$4 and $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$5. A conditional fact $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$6 within the same class yields the linear equality

$$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$7

Inputs in different classes are treated as independent.

Exact inference proceeds through a Datalog derivation hypergraph $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$8, whose nodes are ground atoms and whose hyperedges record rule firings with probabilities $$w_3:\ \operatorname{holds}(x)\land \operatorname{aligns}(x,y)\Rightarrow \operatorname{holds}(y).$$9. Each node $r$0 receives a symbolic polynomial $r$1 over the variables $r$2. If $r$3 is an input fact $r$4, then $r$5. If $r$6 has outgoing hyperedges $r$7, then

$r$8

and for a hyperedge $r$9,

$w_4$0

For each queried output $w_4$1, Praline then solves two constrained optimization problems,

$w_4$2

subject to the linear constraints on the $w_4$3. The resulting interval $w_4$4 is a sound lower/upper bound on $w_4$5 (Wang et al., 21 Aug 2025).

6. $w_4$6-exact inference, theoretical guarantees, and empirical results

Because exact inference is reducible to bilinear constrained optimization and is NP-hard in general, the 2025 system introduces a more scalable $w_4$7-exact procedure, written as $w_4$8. The algorithm first grounds the program and computes the derivation graph $w_4$9, then forms the linear constraints $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$0 over the variables $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$1. It next infers correlation types for pairs of subexpressions—definitely positively correlated $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$2, negatively correlated $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$3, independent $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$4, or unknown $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$5—by solving small optimization or SAT subproblems over $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$6 and checking whether $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$7 is greater than, less than, or incomparable to $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$8. It then computes fast sound interval bounds bottom-up using Fréchet inequalities refined by the known correlation types, and finally performs iterative refinement: for each output $$\operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \operatorname{lhsHolds}(r),$$9, it pushes up the lower bound and pushes down the upper bound in steps of size $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$0, checks satisfiability of $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$1, and then binary-searches until the width is at most $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$2 (Wang et al., 21 Aug 2025).

The key algebraic operations are also fixed. If $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$3 and $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$4, conjunction collapses to

$$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$5

while disjunction uses inclusion-exclusion,

$$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$6

The paper states that every $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$7, $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$8, and $$\neg \operatorname{holds}(x)\land \operatorname{inLhs}(x,r)\Rightarrow \neg \operatorname{lhsHolds}(r),$$9 rewrite preserves the exact probabilistic semantics, that the approximate-bound rules are sound over the four correlation types, and that iterative refinement converges from the outside in. The resulting guarantee is

$\delta$00

When correlation classes are small, full constraints can be enforced in refinement and exactness with $\delta$01 can be recovered.

The empirical study uses 19 side-channel benchmarks, including Datalog analyses for power-side-channel leakage over real cryptographic kernels such as AES, SHA3, Keccak, and MAC, with derivation graphs up to 200 K nodes and correlation classes up to approximately 700 facts. It also uses 11 Bayesian networks from the bnlearn repository, with 10 to 223 nodes and correlation classes of size 2 to 13. Exact-opt serves as ground truth when it finishes, which occurred on 12 of 30 benchmarks. With $\delta$02, the reported average lower/upper-bound errors on the 12 solvable benchmarks comprising 3,179 facts are 0.00168 and 0.00177, with maximum error below 0.01. Over 97,056 total queried facts, at least 73% are guaranteed $\delta$03-exact. Runtime results show 60% of benchmarks completing in under 10 seconds, 76% in under 100 seconds, and the largest 200 K-node benchmark in approximately 1,400 seconds. The ablation study reports that ApproxOnly and Approx+Corr lose 17–28 times tightness, while full $\delta$04 gives the best trade-off of precision versus runtime (Wang et al., 21 Aug 2025).

7. Comparative significance and recurring themes

The two systems named Praline are formally distinct, but each addresses probabilistic reasoning under incomplete specification. The MLN-based Praline is designed for elementary-science QA with incomplete and noisy automatically extracted rules; its central devices are lexical and structural alignment, node-level holds, explicit proof dependencies, acyclicity, and false-unless-proven semantics. The Datalog-based Praline is designed for inference with partially known input correlations; its central devices are correlation classes, possible-world constraints, symbolic derivation expressions, constrained optimization, and $\delta$05-exact refinement (Khot et al., 2015, Wang et al., 21 Aug 2025).

A common misconception is therefore to treat Praline as a single logic-programming framework. In the cited literature it is instead a homonymous label for two separate research contributions. A plausible implication is that the shared name reflects a recurring emphasis on controlled probabilistic inference rather than on any common formal language: the 2015 system regulates proof search inside an MLN for question answering, whereas the 2025 system computes exact or near-exact probability bounds in Datalog when correlations are only partially known. Both cases place the burden of robustness on the semantics of inference rather than on purely syntactic matching.