Local Inconsistency Resolution in PDGs
- Local Inconsistency Resolution (LIR) is a heuristic method for reducing mismatches in probabilistic dependency graphs by focusing on localized attention and controlled parameter updates.
- It employs attention and control masks to iteratively address subsets of a model, thereby recovering established procedures like EM, belief propagation, and GAN training as special cases.
- Experimental evaluations reveal that different refocus strategies significantly impact inconsistency reduction and stability, highlighting trade-offs between local adjustment and global accuracy.
Local Inconsistency Resolution (LIR) is a generic algorithm for learning and approximate inference in probabilistic models that proceeds by iteratively focusing on a subset of a model and resolving inconsistencies using the parameters under control. In the formulation built on Probabilistic Dependency Graphs (PDGs), LIR gives an explicitly epistemic interpretation to attention and control, treats inconsistency as a formal optimization target, and recovers a range of established procedures as special cases of local inconsistency reduction rather than as unrelated algorithms (Richardson et al., 18 Apr 2026).
1. Conceptual basis
LIR is presented as a heuristic iterative procedure over a parametric PDG. At each iteration, one chooses an attention mask that determines which parts of the model currently matter and a control mask that determines which parameters are allowed to change. The ensuing update reduces the inconsistency of the attended submodel, not necessarily that of the full model. This locality is central in two senses: locality of attention, because one does not evaluate all inconsistencies simultaneously, and locality of control, because one may treat some beliefs or parameters as fixed while revising others (Richardson et al., 18 Apr 2026).
The framework is motivated by settings in which a model contains multiple probabilistic beliefs or constraints that may not be mutually satisfiable. The paper explicitly treats this as both a computational and an epistemic issue. Detecting or minimizing full inconsistency may be intractable, whereas local inconsistency may be tractable; and in realistic reasoning one often revises only the beliefs one is willing or able to change. In this reading, focusing on a subset of the model means that only some arcs or constraints are currently “in mind,” and resolving inconsistency using parameters under control means that only selected mutable beliefs are revised (Richardson et al., 18 Apr 2026).
LIR distinguishes global inconsistency, defined on the whole PDG, from local inconsistency, defined after applying attention to a restricted or reweighted submodel. The local objective is generally an underestimate of the global one. As a result, resolving one attended inconsistency may leave other inconsistencies untouched or may worsen them elsewhere. The framework is therefore not a guaranteed global solver; it is a local inconsistency-reduction scheme whose behavior depends on the schedule by which focus is reassigned (Richardson et al., 18 Apr 2026).
2. Formal basis in Probabilistic Dependency Graphs
A parametric PDG is written as
$\dg M(\Theta) = (\mathcal X, \mathcal A, \Theta{=} \{\Theta_a \}_{a \in \mathcal A}, \mathbb P {=}\{ \mathbf p_a \}_{a \in \mathcal A} ),$
where is a set of variables, is a directed hypergraph of arcs, each arc has source variables $\Src a$ and target variables $\Tgt a$, each arc has a parameter space , and each arc has a parameterized conditional distribution
$\mathbf p_a : \Theta_a \times \mathcal V\Src a \to \Delta \mathcal V\Tgt a.$
PDGs are the representational substrate of LIR because they can explicitly encode arbitrary collections of probabilistic beliefs even when those beliefs are inconsistent (Richardson et al., 18 Apr 2026).
The paper decomposes inconsistency into two terms. The first is observational incompatibility,
$\OInc_{\dg M}(\mu) := \sum_{a \in \mathcal A} \beta_a\, D_{\mathrm{KL}}\!\Big( \mu(\Tgt a,\Src a) \,\big\|\, \mathbf p_a(\Tgt a \mid \Src a)\,\mu(\Src a) \Big),$
which measures disagreement between a candidate joint distribution and the local conditional beliefs carried by the arcs. The second is structural deficiency,
0
which measures mismatch between 1 and the structural decomposition induced by the hypergraph. Full 2-inconsistency is then
3
Most examples in the paper use 4, while the factor-graph and belief-propagation constructions use 5 (Richardson et al., 18 Apr 2026).
Attention is not an external annotation layered on top of the semantics; it is realized by the same weighting machinery. The paper writes attention as 6 and denotes the inconsistency of the attended model by 7. Thus a focus may amount either to hard restriction to a subset of arcs or to a continuous reweighting of structural and observational terms (Richardson et al., 18 Apr 2026).
3. Update rule, local optimization, and exact-solve regimes
Algorithm 1 of the paper defines LIR by the update
8
where 9 is the attention mask, 0 is the control mask, and Refocus supplies a sequence of foci 1. The state being updated is the parameter collection 2. Each step therefore consists of choosing what inconsistency is visible and choosing what part of the model may respond to it (Richardson et al., 18 Apr 2026).
Two limiting interpretations are explicit. When 3 is small, the update is approximated by an ordinary gradient step on the attended inconsistency. When 4 is infinite on the controlled components, the update often reduces to exact local minimization,
5
The latter regime is how the paper reconstructs exact E-steps, exact message updates, and other alternating optimization procedures (Richardson et al., 18 Apr 2026).
For discrete PDGs, the implementation minimizes an inner objective over the joint distribution 6, computing
7
and then uses the envelope theorem so that gradients with respect to 8 are taken as 9 without backpropagating through the inner solve. The paper emphasizes warm-starting the inner optimization and using either Adam or ODE solvers for the outer step (Richardson et al., 18 Apr 2026).
Theoretical guarantees are intentionally limited. The paper proves a convexity result when parameterizations are constant or unconditional and log-concave and when 0, but it also states that inconsistency is not always convex in parameters of conditional probability distributions. The framework is therefore positioned as a general heuristic with selective exact subroutines rather than as a universally convergent optimization method (Richardson et al., 18 Apr 2026).
4. Recovery of established algorithms
The unifying claim of LIR is that many standard algorithms can be expressed as particular choices of attention and control over a PDG (Richardson et al., 18 Apr 2026).
| Procedure | Attention/control pattern | Resulting objective or update |
|---|---|---|
| EM | Alternate full control of 1 and 2 | 3 |
| Belief propagation | Local message focus; full control of one message | Standard BP message equations |
| GANs | Alternate generator and discriminator foci | GAN min-max / JS decomposition |
| Adversarial training | Alternate control of perturbed input and model parameters | Attack/training alternation |
| GFlowNets | Surprisal-based local attention on trajectories | 4 |
For latent-variable models, the paper constructs a PDG in which the inconsistency is exactly negative ELBO: 5 Alternating full control of 6 and 7 recovers the E-step and M-step, and the paper states that 8. In the same formalism, allowing 9 to depend on $\Src a$0 yields the variational-autoencoder setting.
For factor graphs, LIR recovers belief propagation by defining foci that expose a local message context and grant full control to a single message variable. The resulting optimality conditions are the standard BP updates
$\Src a$1
The paper’s proposition states that the possible runs of LIR are precisely those of belief propagation under different message schedules.
For GANs, the PDG contains a real source, a fake source, a discriminator, and an additional belief $\Src a$2 asserting that the real/fake coin should be unpredictable from the image. The derived inconsistency is
$\Src a$3
By changing which beliefs are attended to and which parameters are under control, LIR alternates discriminator and generator updates. Adversarial training is treated analogously: controlling perturbed inputs performs attack generation, while controlling model parameters performs repair against those attacks.
The same localist reading is extended to transformer inference and to GFlowNets. In the transformer case, choosing attention weights $\Src a$4 and controlling $\Src a$5 recovers the fixed point
$\Src a$6
In the GFlowNet case, centered surprisal-based attention yields a length-normalized trajectory-balance-like loss rather than standard trajectory balance.
5. GFlowNet modification and empirical behavior
The clearest new algorithmic proposal derived from LIR is the modified GFlowNet loss
$\Src a$7
together with the analogous modified LPV loss
$\Src a$8
Relative to standard trajectory balance, the distinguishing factor is $\Src a$9. The paper argues that this factor is more natural because the inconsistency is derived by choosing a uniformly random transition index in a trajectory and averaging local inconsistency per step rather than per full trajectory (Richardson et al., 18 Apr 2026).
The experimental program has two parts. First, the paper studies randomly generated discrete PDGs with intentionally conflicting conditional distributions under several refocus strategies. All strategies reduced inconsistency, but the reported average resolution percentages differed substantially: Uniform $\Tgt a$0, Partial $\Tgt a$1, and Hub $\Tgt a$2. The corresponding average total-variation distortions from the globally optimized joint distribution were Uniform $\Tgt a$3, Partial $\Tgt a$4, and Hub $\Tgt a$5. The paper interprets this carefully: hub focus causes more absolute distortion than partial focus, but partial focus induces more distortion per unit of inconsistency resolved (Richardson et al., 18 Apr 2026).
Second, the paper evaluates modified GFlowNet losses on four discrete HyperGrid environments: Original, Cosine, Bitwise XOR, and Multiplicative Coprime. Modified losses are reported to show faster early convergence and, in harder environments, improved learned-distribution fidelity. On Bitwise XOR, the appendix reports final $\Tgt a$6 error $\Tgt a$7 for TB and $\Tgt a$8 for ModTB. The appendix also reports much larger median best gradient clipping thresholds for modified losses, with differences of approximately $\Tgt a$9, 0, 1, and 2 for the four environments listed above, which the paper interprets as evidence of more stable gradients (Richardson et al., 18 Apr 2026).
6. Related local repair notions in other domains
Outside the PDG framework, the phrase “local inconsistency resolution” or closely related local repair ideas recur in several technical literatures, but usually with narrower object classes and without the same general probabilistic semantics.
In R2RML and RML-based RDF generation, the closest operationalization of LIR is detecting violations at the mapping-rule level, localizing them to specific Triples Maps, Term Maps, or vocabulary terms, and repairing those local rule fragments so that future RDF generation no longer propagates the same inconsistencies. The paper frames inconsistency as incorrect application of vocabulary terms to data, with rule validation preferred because it “points exactly to the root causing the inconsistency and prevent violations from being propagated.” The Resglass workflow then detects rule inconsistencies, clusters rules by Triples Map, ranks clusters, and supports expert refinement of the implicated local rules (Dimou, 2020).
In direct modeling of CAD models, local inconsistency resolution takes the form of resolving geometry-associativity inconsistency after a direct edit by decomposing the affected model into minimal over-constrained parts and maximal well-constrained parts. Valid repair actions are then limited to removals inside the minimal dependent part or additions bridging maximal well-constrained parts, and the paper states that the method allows a “complete exclusion of invalid resolution options.” Prioritization is based on a change-rate measure intended to reduce subsequent model variation (Zou et al., 2019).
In large-scale systems, a more architectural precursor appears in the language of Subjective Consistency, Eventual Consistency, solipsistic transactions, focused transactions, and SOUPS (“Single Object Update per Process Step”). That paper does not define LIR as a named framework, but it argues that systems should permit local progress under incomplete or conflicting state and then manage inconsistency through warnings, queued events, compensation, and apology-oriented computing. The result is a design philosophy in which local inconsistency is admitted, bounded, logged, and later reconciled rather than globally prevented at all times (0909.1782).
In deep learning, a mathematically distinct use of local inconsistency is the label-free quantity
3
interpreted as worst-case local output instability under parameter perturbations. The associated methods IAM-D and IAM-S minimize or adversarially train against this instability. Although this literature does not adopt the PDG formalism, it is another instance in which inconsistency is localized, quantified, and then actively reduced using only a controlled subset of model degrees of freedom (Kim et al., 29 May 2026).
7. Limitations, open problems, and terminological scope
The PDG paper is explicit that LIR is a heuristic framework rather than a completed general theory. It leaves open the questions of when LIR converges, how attention and control schedules should be learned or adapted, whether hybrid algorithms can be systematically designed by composing foci, and how the framework behaves when inconsistency is nonconvex in the parameters of conditional probability distributions. The implementation results are confined to small discrete PDGs, and the paper states that full PDG inconsistency is NP-hard in general, with bounded-tree-width settings as a principal tractable exception (Richardson et al., 18 Apr 2026).
The broader literature also shows that “local inconsistency” does not have a single cross-domain semantics. In probabilistic models it denotes reduction of attended PDG inconsistency; in RDF generation it refers to localization of violations to mapping fragments; in CAD it refers to repair of structurally isolated constraint inconsistencies; and in deep learning it denotes KL-instability under local parameter perturbation. A plausible implication is that the phrase functions more as a methodological pattern—restrict scope, expose a conflict, revise only what is controllable—than as a universally fixed formal object.
Context also matters because the acronym LIR is overloaded. In elliptic PDE theory, LIR denotes the Local Increasing Regularity Method, a bootstrap method turning local a priori estimates for 4 into global 5 and Sobolev solvability results on Riemannian manifolds (Amar, 2018). In Venus atmospheric studies, LIR denotes Akatsuki’s Longwave Infrared Camera, used to reveal the global structure of thermal tides in the upper cloud layer (Burkart, 2019). Within research writing, “LIR” therefore identifies a specific topic only when the surrounding disciplinary context is explicit.