Attribute-Computation Paths: Methods & Applications

• Attribute-computation path is a computational object that defines the sequence or structure used to derive, propagate, and constrain attributes in domains like LLMs, causal fairness, and network systems.
• The concept appears in varied forms—from minimal transformer layer subsets for factual retrieval to integration paths in feature attribution and formal chains in description logics—highlighting domain-specific formulations.
• These paths drive practical improvements in explainability, fairness, and operational efficiency by structuring how attributes are computed, optimized, and aggregated.

Attribute-computation path denotes, across several research areas, a path-like computational object through which an attribute is produced, propagated, constrained, or retrieved. In the cited literature, the term and closely related notions denote different formal objects rather than a single standardized formalism: a subset of transformer layers sufficient for factual recall in LLMs (Hochman et al., 19 Jun 2026), an integration path in axiomatic attribution methods (Ward et al., 2020), a selected set of causal paths from a sensitive attribute to a decision (Chiappa et al., 2018), a serial/parallel aggregation scheme that lifts edge attributes to paths and path sets (Fiaschi et al., 7 Nov 2025), or an explicit attribute-mediated inference branch in visual recognition systems (Mbongo et al., 6 Dec 2025). Taken together, these works treat the path not as a peripheral implementation detail but as the object that fixes the semantics of attribution, retrieval, fairness, or composition.

1. Scope and definitions

Across the cited works, the same phrase is instantiated with different state spaces, path elements, and sufficiency criteria.

Domain	Path object	Representative formulation
LLM factual retrieval	Subset of layers over an entity representation	Minimal attribute-sufficient computation path (Hochman et al., 19 Jun 2026)
Path attribution	Integration path from reference to explanation point	Straight-line IG vs multi-path interventional Shapley (Ward et al., 2020)
Causal fairness	Selected causal paths from $A$ to $Y$	Path-specific counterfactual fairness (Chiappa et al., 2018)
Network systems	Attribute aggregation from edges to paths/path sets	$\Psi(P)=OP_s \circ OP_p \circ T$ (Fiaschi et al., 7 Nov 2025)
Visual composition	Attribute-mediated transformation or composition branch	Classifier composition, latent transformations, or prompt-guided paths (Radenovic et al., 2021)

In the LLM setting, an attribute-computation path is defined as “a sequence of computations over the entity representation, that together enable the LLM to answer the prompt correctly,” and a minimal path is “a subset of layers that is both sufficient and necessary for computing the desired output” (Hochman et al., 19 Jun 2026). In path-based explainability, the central object is the path $\gamma$ from baseline to input, with feature attributions obtained by integrating gradients along $\gamma$ (Ward et al., 2020). In causal fairness, the relevant path is a selected set of causal routes by which a sensitive attribute affects a decision, so that only unfair routes are neutralized (Chiappa et al., 2018). In network systems, the path becomes a compositional operator over edge attributes, later lifted to path sets by serial and parallel aggregation plus structural transformation (Fiaschi et al., 7 Nov 2025). In visual composition, several papers do not use the phrase explicitly, but they implement attribute-mediated computation by classifier composition, learned transformations, graph propagation, or text-guided cross-attention (Radenovic et al., 2021).

An adjacent formal usage appears in type theory, where a computational path is a “composition of rewrites” witnessing equality $a =_{s} b$ and later rewritten by $LND_{EQ}$-TRS (Ramos et al., 2016). This usage is not attribute-specific, but it provides a precise syntactic model of path as explicit computational evidence.

2. Path-based attribution and explainability

In feature attribution, the path determines what explanation means. The additive decomposition is anchored by

$f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$

and, for path methods,

$$f(\vec{x}_b)-f(\vec{x}_a) = \int_{\vec{x}_a}^{\vec{x}_b}\vec{\nabla}f(\vec{x}) \cdot d\vec{x}.$$

Integrated Gradients instantiates this with the straight line between reference and explanation points, whereas interventional Shapley is reinterpreted as averaging over the $n!$ orthogonal feature-ordering paths between those points (Ward et al., 2020).

This path view sharpens the distinction between atomic-game and infinitesimal-game attribution. Interventional Shapley changes one feature at a time, completely, in every possible ordering; Aumann–Shapley and IG use a single path, typically the straight line; GIG extends straight-line integration to models with discontinuities by combining gradient integration on continuous segments with local Shapley assignments at discontinuities (Ward et al., 2020). The paper proves that Aumann–Shapley and interventional Shapley agree for

$Y$0

and explains the equivalence by Stokes-theorem symmetry: interventional Shapley’s path family is symmetric around the straight line for pairwise bilinear interactions (Ward et al., 2020).

The empirical consequence is that path choice affects off-manifold behavior. In a toy classification setting with increasing unsupported-region data, interventional Shapley shifts attribution strongly toward one feature, while GIG remains more stable; with 100 uniform points the feature-attribution ratio exceeds 2, and with 200 uniform points the first feature can appear up to 6 times more important than the second (Ward et al., 2020). The stated explanation is geometric: orthogonal detours pass through regions where the model has little support, whereas the straight line “almost always pass[es] closer to the data manifold” (Ward et al., 2020).

A complementary line of work treats path selection itself as an optimization problem. “Path Choice Matters for Clear Attribution in Path Methods” formalizes attribution as $Y$1, restricts search to manipulation paths $Y$2, and proposes the Concentration Principle, which maximizes attribution variance

$Y$3

SAMP then greedily chooses the next manipulated features using the gradient-projection score $Y$4, while the infinitesimal constraint caps $Y$5 and the momentum strategy smooths path construction (Zhang et al., 2024). This suggests that “attribute-computation path” in attribution research is simultaneously a semantic object and an optimization variable.

3. Mechanistic attribute computation in LLMs

The most explicit recent definition occurs in factual retrieval in LLMs. A decoder-only transformer with layers $Y$6 is analyzed at the entity token position $Y$7, with layer activations

$Y$8

The paper defines the earliest attribute-sufficient layer by a lock operation

$Y$9

and

$\Psi(P)=OP_s \circ OP_p \circ T$0

It then searches greedily for a minimal path using an isolate operation that feeds clean entity computation only through selected layers while patching excluded layers with counterfactual activations (Hochman et al., 19 Jun 2026).

The reported findings are that these paths are typically multi-layer, often non-contiguous, and frequently redundant. On LLaMA 3.1 8B and Qwen3 8B, the mean earliest sufficient layer is $\Psi(P)=OP_s \circ OP_p \circ T$1 and $\Psi(P)=OP_s \circ OP_p \circ T$2, mean primary path length is $\Psi(P)=OP_s \circ OP_p \circ T$3 and $\Psi(P)=OP_s \circ OP_p \circ T$4, paths skip at least one layer in $\Psi(P)=OP_s \circ OP_p \circ T$5 and $\Psi(P)=OP_s \circ OP_p \circ T$6 of cases, and non-identical alternative minimal paths exist for $\Psi(P)=OP_s \circ OP_p \circ T$7 and $\Psi(P)=OP_s \circ OP_p \circ T$8 of prompts (Hochman et al., 19 Jun 2026). Intervention tests further show that the endpoint representation is usually necessary but often not sufficient by itself, implying that the path acts as a coherent functional unit rather than a single localized state (Hochman et al., 19 Jun 2026). A plausible implication is that factual retrieval is better modeled as a distributed computation over the entity representation than as retrieval from one layer or one MLP block.

Related LLM work generalizes the path concept from layers to explicit reasoning units and feature graphs. AttriCoT constructs a structural causal model over chain-of-thought units, using

$\Psi(P)=OP_s \circ OP_p \circ T$9

and attributes each output unit by the conditional log-probability scalarizer

$\gamma$0

Its key result is a local causal attribution over a realized reasoning path using $\gamma$1 forward passes, where $\gamma$2 is the number of units (Wei et al., 20 Jun 2026). ADAG operates at a different granularity: it starts from an attribution graph $\gamma$3, defines attribution profiles by input attributions and output contributions, clusters internal features by a multiview similarity matrix, and then describes the resulting supernodes with an explainer–simulator loop (Arora et al., 8 Apr 2026). Together, these papers suggest a shift from simple path localization toward explicit causal structure over units, features, and graph edges.

4. Causal, logical, and formal path semantics

In fairness research, path semantics becomes normative. “Path-Specific Counterfactual Fairness” starts from the premise that a sensitive attribute $\gamma$4 may affect a decision $\gamma$5 through both fair and unfair pathways, so fairness is not equivalent to deleting every effect of $\gamma$6 (Chiappa et al., 2018). In simple mediation, nested counterfactuals such as

$\gamma$7

separate direct and indirect effects; in more complex graphs, the paper uses recursively nested counterfactuals to isolate the selected unfair path set (Chiappa et al., 2018). The main methodological move is not to compute the path-specific effect explicitly, but to correct observed descendants adversely affected by $\gamma$8, infer latent individualized factors with variational inference, and regenerate the descendants under a path-specific intervention before prediction. In the linear example, this appears as deleting the unfair $\gamma$9-dependent components from $\gamma$0, $\gamma$1, and $\gamma$2 while preserving other terms (Chiappa et al., 2018). This suggests a distinct sense of attribute-computation path: a selected subset of causal routes whose influence is to be blocked or retained.

In description logics, attribute paths are chains of functional relationships. An attribute chain $\gamma$3 is interpreted as a composition of partial or total functions, and same-as constraints

$\gamma$4

require both paths to be defined and to end at the same value (Borgida et al., 2011). The distinction between partial and total attributes is decisive: under total attributes,

$\gamma$5

while under partial attributes this implication fails (Borgida et al., 2011). The paper proves that for partial attributes the least common subsumer exists and is polynomial-time computable, whereas for total attributes it may not exist and may be exponential in size (Borgida et al., 2011). Here the path is not a learned computational route but a formal composition of attribute functions with strong consequences for subsumption and definability.

A further formalization appears in type theory, where a computational path $\gamma$6 from $\gamma$7 to $\gamma$8 is “a composition of rewrites,” written $\gamma$9, and internalized into the identity type by

$a =_{s} b$0

Operations such as reflexivity $a =_{s} b$1, symmetry $a =_{s} b$2, and transitivity $a =_{s} b$3 endow these paths with a groupoid-like structure up to rewrite, while $a =_{s} b$4-TRS eliminates redundancies such as $a =_{s} b$5 (Ramos et al., 2016). This is an adjacent, non-attribute-specific path formalism, but it clarifies a general theme: a path may be treated as explicit computational evidence rather than as an implicit semantic abstraction.

5. Visual and multimodal attribute pathways

Several vision papers implement attribute-mediated computation paths even when the phrase itself is not formalized. In “Large-Scale Attribute-Object Compositions,” CompNet learns separate attribute and object classifier vectors $a =_{s} b$6, composes them with an MLP,

$a =_{s} b$7

and scores an image by

$a =_{s} b$8

The attribute pathway therefore runs from attribute classifier parameters, not merely attribute probabilities, into a composition classifier for unseen attribute-object pairs (Radenovic et al., 2021).

“Learning Single/Multi-Attribute of Object with Symmetry and Group” treats attributes as latent-space transition operators $a =_{s} b$9, with transformations such as

$LND_{EQ}$0

Its Relative Moving Distance compares the displacement caused by adding versus removing an attribute,

$LND_{EQ}$1

so that attribute recognition is based on path behavior in feature space rather than static pattern matching (Li et al., 2021). The paper explicitly states that it does not use the phrase “Attribute-Computation Path,” but interprets its own framework in exactly those terms (Li et al., 2021).

A graph-based composition model for unseen single- and multi-attribute-object recognition implements a different path: label word embeddings $LND_{EQ}$2 are propagated by a GCN,

$LND_{EQ}$3

candidate compositions are assembled by a mask matrix $LND_{EQ}$4 as $LND_{EQ}$5, and recognition is performed by nearest-composition matching to clustered image embeddings $LND_{EQ}$6 (Chen et al., 2020). In long-range gender recognition, the “attribute-mediated path” is an explicit second branch of a dual-path transformer: CLIP text embeddings $LND_{EQ}$7 for soft-biometric prompts query visual tokens $LND_{EQ}$8 via

$LND_{EQ}$9

the per-attribute vectors $f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$0 are summed to

$f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$1

and this attribute-mediated embedding is fused with a direct visual path for ternary gender prediction (Mbongo et al., 6 Dec 2025). These works differ architecturally, but all treat attributes as active mediators of computation rather than passive labels.

6. Network and path-set attribute computation

In network systems, attribute-computation path denotes the way link-level quantities are propagated to end-to-end path or path-set quantities. “QR-MO” formulates routing in a 5G-MEC graph $f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$2 with edge cost vectors

$f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$3

specifically packet-loss probability, latency, and jitter (Sarah et al., 23 Mar 2025). For each objective,

$f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$4

and QR-MO learns vector-valued Q-estimates for next-hop decisions while selecting actions through a dominance-based heuristic rather than scalarization (Sarah et al., 23 Mar 2025). The path here is the route itself, and attribute computation is edge-attribute aggregation to a path-attribute vector.

A more abstract formalization appears in “A Formal Model for Path Set Attribute Calculation in Network Systems,” which defines path-set evaluation by

$f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$5

Here $f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$6 is serial composition along a path, $f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$7 is parallel composition across paths, and $f(\vec{x}_b)-f(\vec{x}_a) = \sum_{i=1}^{n}\phi_i$8 is a structural transformation such as identity, union, or cut transformation (Fiaschi et al., 7 Nov 2025). The paper instantiates this framework for delay, administrative cost, capacity, unavailability probability, and fault probability, showing that overlapping path sets often require cut-based rather than naive per-path aggregation (Fiaschi et al., 7 Nov 2025). This is a direct formal model of how an attribute is computed along a path and then lifted to a path set.

“Path computation in multi-layer networks” adds stateful protocol semantics. A feasible path is not determined by topology alone, but by current protocol and the stack of pending encapsulations induced by conversion, encapsulation, and decapsulation actions (Lamali et al., 2016). The problem is solved in the unconstrained additive case by compiling the network into a weighted push-down automaton and then a weighted CFG, while bandwidth-constrained and multi-QoS variants become NP-complete in general (Lamali et al., 2016). Taken together, these papers broaden the notion of attribute-computation path from explanation and recognition to operational route synthesis under typed path state and multi-attribute cost structure.

7. Cross-cutting themes and unresolved issues

Several themes recur across these literatures. First, path choice or path discovery determines semantics. In attribution, different paths between the same endpoints yield different feature attributions (Ward et al., 2020); in factual retrieval, different minimal layer subsets can compute the same attribute (Hochman et al., 19 Jun 2026); in fairness, different causal paths from the same sensitive attribute are explicitly classified as fair or unfair (Chiappa et al., 2018). This suggests that the path is often the hidden variable behind what an attribute explanation or decision rule actually means.

Second, many papers replace monolithic localization with distributed or structured computation. The LLM retrieval work reports multiple, functionally equivalent minimal paths and non-contiguous layer subsets (Hochman et al., 19 Jun 2026). ADAG groups neurons into supernodes by shared input and output effects rather than treating single features as self-explanatory (Arora et al., 8 Apr 2026). The path-set model in networking likewise shows that path attributes may have to be computed over unions or cuts rather than over individual paths (Fiaschi et al., 7 Nov 2025). A plausible implication is that “where the attribute is” is often less informative than “through which sequence or structure the attribute is computed.”

Third, path form is constrained by each domain’s admissible state transitions. In description logics, definedness and totality determine how attribute chains compose (Borgida et al., 2011). In multi-layer routing, the protocol stack makes feasibility context-free rather than regular (Lamali et al., 2016). In SymNet, add/remove operators are regularized by symmetry, invertibility, and conditional commutativity (Li et al., 2021). These are not interchangeable path notions; they are domain-specific semantics for what counts as a valid computation.

Finally, each literature exposes a different unresolved issue. Attribution path methods face off-manifold ambiguity and the absence of a universally meaningful default path (Ward et al., 2020). LLM mechanistic work still leaves open whether alternative discovered paths are naturally used or only patch-induced (Hochman et al., 19 Jun 2026). Path-specific fairness depends on correct causal graphs and path judgments (Chiappa et al., 2018). Attributed feature-model synthesis remains parameterized by domain knowledge for attribute/feature classification, null values, and legal placement choice (Bécan et al., 2015). The common conclusion is not that attribute-computation path is a unified theory, but that it is a recurring research pattern: attributes are increasingly analyzed through the specific computational, causal, logical, or structural routes by which they arise.