Attribution Contract in Generative Models

• Attribution Contract is a formal specification that defines explicit components (S, C, O, P, E) essential for attributing features in generative language models.
• It distinguishes between various attribution settings—local next-token, prompt-conditioned, and span-level—thereby clarifying misunderstandings such as the self-attribution fallacy.
• The concept extends to diffusion models with state-level, denoising-stage, and prompt-to-output contracts, enabling method–contract evaluations across generative processes.

The Attribution Contract is a specification for feature-attribution claims in generative LLMs that makes explicit what is being explained, which features are eligible to receive attribution, what generative process is assumed, what is held fixed, and what model score is being attributed. In "The Attribution Contract: Feature Attribution for Generative LLMs" (Nguyen, 21 May 2026), this specification is introduced to address a conceptual limitation of carrying classifier-era feature attribution directly into generative language modeling. The central claim is that many disagreements about feature attribution in generative LLMs are not disagreements about attribution algorithms, but about unstated explanatory contracts.

1. Conceptual basis and problem statement

Feature attribution in classifier-era settings assumes a static mapping $x \to f(x)$ in which features are fixed components of $x$, such as pixels or tokens, and the model produces a single prediction. Importance is then assigned relative to a scalar target such as a class logit or loss (Nguyen, 21 May 2026). In generative LLMs, by contrast, the model makes many predictions in sequence or via iterative refinement, so the identity of the relevant feature set is not fixed in advance.

In autoregressive LLMs, the prompt $x$ and all previously generated tokens $y_{<t}$ feed into the next-token prediction $y_t$. Previously generated tokens are therefore simultaneously outputs and future inputs. In diffusion LLMs, generation proceeds through an iterative denoising or unmasking chain $z_T \to z_{T-1} \to \dots \to z_0 \to y$, so there is no fixed left-to-right next token, and explanation can target intermediate states or stages rather than only the final output (Nguyen, 21 May 2026).

The paper treats this ambiguity as a conceptual problem rather than an implementation detail. In autoregressive settings, attribution over $y_t$ can legitimately award high importance to $y_{<t}$ because those tokens are predictive context. But if the explanatory goal is “which prompt tokens explain the meaning of the generated span,” attributing to $y_{<t}$ is misleading. This mismatch is named the self-attribution fallacy: reading attribution to generated-prefix tokens as if it answered a prompt-level question without stating the explanatory contract (Nguyen, 21 May 2026). In diffusion settings, the same attribution method can likewise be interpreted as explaining the wrong object if prompt tokens, intermediate states, and denoising stages are not explicitly distinguished.

A related but broader framing appears in "Unifying Corroborative and Contributive Attributions in LLMs" (Worledge et al., 2023), which decomposes attribution systems into input, model, output, attributable units, attribution domain, evaluator, and attribution set. This suggests that the Attribution Contract belongs to a wider effort to make attribution claims modular, use-case specific, and evaluable.

2. Formal specification

The paper defines an Attribution Contract, or explanatory setting, as a tuple

$A = (S, C, O, P, E),$

read as SCOPE (Nguyen, 21 May 2026).

Component	Meaning
$x$0	model score being attributed
$x$1	what is held fixed
$x$2	output being explained
$x$3	assumed generative process
$x$4	features eligible to receive attribution

The score $x$5 can be a logit, probability, surprisal, loss, or log-likelihood. The held-fixed component $x$6 has a computational meaning: the variable remains in the forward pass at its actual value but is removed from the attribution path or perturbation set, as when a variable is not interpolated in Integrated Gradients. The output $x$7 can be a next token $x$8, a generated sequence $x$9, an intermediate diffusion state $x$0, or the final output $x$1. The generative process $x$2 names the relevant decoding or denoising process. The eligibility set $x$3 specifies who can receive credit, including prompt tokens $x$4, previously generated tokens $x$5, intermediate states $x$6, or denoising stages $x$7 (Nguyen, 21 May 2026).

The paper also gives a notational correspondence to the requested form

$x$8

with $x$9, $y_{<t}$0, $y_{<t}$1, $y_{<t}$2, and $y_{<t}$3 (Nguyen, 21 May 2026). The central interpretive claim is contract relativity: the same attribution method can yield different answers under different contracts because $y_{<t}$4, $y_{<t}$5, $y_{<t}$6, and $y_{<t}$7 change the meaning of an attribution map.

3. Autoregressive LLMs

For autoregressive generation, the paper uses the factorization

$y_{<t}$8

with token-level score

$y_{<t}$9

surprisal

$y_t$0

and sequence-level score

$y_t$1

(Nguyen, 21 May 2026).

Three example contracts are distinguished. Local next-token attribution sets $y_t$2, $y_t$3, $y_t$4, $y_t$5 nothing held fixed, and $y_t$6. Its use case is “Which available context features (prompt + prefix) made $y_t$7 likely?” Under this contract, high mass on $y_t$8 is informative because it shows predictive context, but it does not explain how the prompt caused the overall response (Nguyen, 21 May 2026).

Prompt-conditioned token attribution keeps $y_t$9 and $z_T \to z_{T-1} \to \dots \to z_0 \to y$0, but restricts eligibility to prompt tokens only, $z_T \to z_{T-1} \to \dots \to z_0 \to y$1, while holding $z_T \to z_{T-1} \to \dots \to z_0 \to y$2 fixed in $z_T \to z_{T-1} \to \dots \to z_0 \to y$3. The score remains $z_T \to z_{T-1} \to \dots \to z_0 \to y$4. Here the prefix participates in the forward pass but is not eligible for attribution, so mass that would have gone to $z_T \to z_{T-1} \to \dots \to z_0 \to y$5 is redistributed among $z_T \to z_{T-1} \to \dots \to z_0 \to y$6 (Nguyen, 21 May 2026).

Span-level prompt attribution sets $z_T \to z_{T-1} \to \dots \to z_0 \to y$7, $z_T \to z_{T-1} \to \dots \to z_0 \to y$8, $z_T \to z_{T-1} \to \dots \to z_0 \to y$9, $y_t$0 held fixed, and

$y_t$1

Its use case is “Which prompt tokens explain the whole generated span?” Evaluation re-scores the fixed span under prompt perturbations; regenerating a new span tests a different contract (Nguyen, 21 May 2026).

The paper’s translation example illustrates how these contracts differ. With the prompt Translate to French: “The dog is black.” and the generation Le chien est noir., local next-token attribution to “noir” concentrates on the generated prefix “Le chien est” because it predicts “noir.” If the same output token is explained under prompt-conditioned attribution, with the prefix fixed and excluded from eligibility, attribution mass shifts to the prompt token “black.” If the output is the full span Le chien est noir and the eligible features are prompt tokens, attribution concentrates on prompt terms such as “dog” and “black” that explain the sequence (Nguyen, 21 May 2026).

The hallucination example is used to state the self-attribution fallacy more sharply. Running Integrated Gradients on a hallucinated token with $y_t$2 yields high mass on $y_t$3. Reading this as “the prompt did not cause the hallucination” is described as a contract error. For that question, the paper states that the proper contract is span-level prompt attribution with the hallucinated span held fixed and source-document tokens as the eligible features (Nguyen, 21 May 2026).

4. Diffusion LLMs

Diffusion LLMs are described as an iterative refinement chain

$y_t$4

where $y_t$5 are intermediate states such as masked sequences, discrete latents, or continuous latents depending on the diffusion-language-model family (Nguyen, 21 May 2026). Because there is no autoregressive per-step factorization analogous to next-token prediction, explanation targets must be defined contractually.

The paper identifies three contracts. State-level attribution sets $y_t$6, $y_t$7, $y_t$8, $y_t$9 nothing held fixed, and

$y_{<t}$0

Its use case is “Which prompt features or earlier states influenced the current state?” (Nguyen, 21 May 2026).

Denoising-stage attribution instead takes the final output as the object of explanation, $y_{<t}$1, and makes denoising stages eligible features, $y_{<t}$2. The score is defined by perturbation:

$y_{<t}$3

The paper states that there is no autoregressive-style per-step factorization for diffusion, so stage importance is defined as the change in likelihood under perturbing stage $y_{<t}$4 by ablation, noise-schedule tweak, or substitution. Its use case is “Which stages committed key aspects of the final output (e.g., meaning vs wording)?” (Nguyen, 21 May 2026).

Prompt-to-output attribution sets $y_{<t}$5, $y_{<t}$6, $y_{<t}$7, $y_{<t}$8 nothing held fixed, and $y_{<t}$9. The paper describes this as analogous to span-level prompt attribution in autoregressive models, except that influence propagates through the denoising chain rather than through autoregressive factorization (Nguyen, 21 May 2026).

The masked diffusion example makes the distinction concrete. With the prompt The capital of France is [MASK]., an early $y_{<t}$0 resolves [MASK] to “city,” and a later $y_{<t}$1 refines it to “Paris.” State-level attribution at the early stage explains “city”; stage attribution identifies the late stage committing to “Paris”; prompt-to-output attribution explains “Paris” from prompt tokens such as “France” and “capital” (Nguyen, 21 May 2026). The paper therefore treats state-level attribution as informative for understanding partial commitments and evolution, and denoising-stage attribution as informative for sequencing responsibility, while warning against reading stage-level effects as token-level feature importance.

5. Evaluation as method–contract pairs

A central principle of the paper is that attribution methods must be evaluated together with their contracts. It explicitly states that a single “good attribution map” standard does not exist across contracts (Nguyen, 21 May 2026). Evaluation is therefore contract specific.

For local next-token attribution in autoregressive models, the proposed interventions are to delete, insert, or replace highly attributed prompt or prefix tokens and to measure the change in $y_{<t}$2. Faithfulness requires that high-attribution features yield large score changes. For prompt-conditioned token attribution, prompt tokens are perturbed while $y_{<t}$3 is held fixed, again measuring the change in $y_{<t}$4; invariance requires that changes to the fixed prefix not alter the score or attribution. For span-level prompt attribution, prompt tokens are perturbed while keeping $y_{<t}$5 fixed and re-scoring rather than regenerating, with sensitivity measured by the change in $y_{<t}$6 (Nguyen, 21 May 2026).

For diffusion models, state-level attribution perturbs prompt tokens or swaps earlier states $y_{<t}$7 with $y_{<t}$8, continuing the chain to $y_{<t}$9 and measuring the change in $A = (S, C, O, P, E),$0. Denoising-stage attribution perturbs or substitutes a denoising stage, runs the process to completion, and measures $A = (S, C, O, P, E),$1; stability is defined relative to seed handling when $A = (S, C, O, P, E),$2 specifies seeds. Prompt-to-output attribution perturbs prompt tokens and regenerates through the full chain, measuring the change in $A = (S, C, O, P, E),$3, which the paper notes is computationally heavier because the full process must be rerun (Nguyen, 21 May 2026).

Additional checks are contract dependent. Sensitivity and invariance must be consistent with the held-fixed component $A = (S, C, O, P, E),$4. Stability across sampling seeds matters when seeds are held fixed or varied in the contract. Completeness or additivity is meaningful only when the score decomposes, as in autoregressive sequence log-likelihood, and is not applicable for diffusion stage attribution where the score is defined via perturbation differences (Nguyen, 21 May 2026). The paper’s case-study implication is that the same method, including Integrated Gradients, produces different maps under local next-token, prompt-conditioned, and span-level contracts, so false disagreements between methods can arise when evaluations do not reflect the chosen contract.

6. Reporting, pitfalls, and adjacent developments

The paper’s practical guidance is to declare the Attribution Contract explicitly when publishing or building tools. The required elements are the output being explained, the eligible features, the generative process, the held-fixed conditions, and the score being attributed. It further lists implementation details that should be reported: model and tokenizer, context length and tokenization specifics, decoding configuration such as temperature, top-$A = (S, C, O, P, E),$5, top-$A = (S, C, O, P, E),$6, or beam width, whether randomness or seeds are fixed, whether previously generated tokens are treated as fixed features or re-sampled inputs, and score units and aggregation such as per-token, per-span, per-stage, log-space, probability, or loss (Nguyen, 21 May 2026).

Several pitfalls are identified. One is conflating outputs with inputs by attributing to generated-prefix tokens and reading that as prompt-level causation. Another is leaving decoding policies and randomness implicit, since temperature, top-$A = (S, C, O, P, E),$7, top-$A = (S, C, O, P, E),$8, beam search, and seeds influence $A = (S, C, O, P, E),$9 and thereby attributions. A third is ambiguity about the score itself: “probability,” “log-probability,” and “loss” do not have the same interpretation. A fourth is silently changing what is held fixed, as when a span-level contract regenerates sequences even though $x$00 should be fixed. A fifth is comparing token-attribution maps in autoregressive settings to stage-attribution maps in diffusion settings as if they answered the same question (Nguyen, 21 May 2026).

The broader literature supplied alongside the paper shows that the term “Attribution Contract” is also being adapted outside feature attribution for generative LLMs. In LLM agents, "4D-ARE: Bridging the Attribution Gap in LLM Agent Requirements Engineering" (Yu et al., 8 Jan 2026) describes a compilable specification that binds an agent to produce attribution-complete answers across Results, Process, Support, and Long-term dimensions. In human-centric data attribution, "A Human-Centric Framework for Data Attribution in LLMs" (Wührl et al., 11 Feb 2026) presents a negotiated model with parameters for granularity, modality, compensation, consent, privacy, auditability, and enforcement. In market design, "What's a Credit Worth? A Market Framework for Attribution-Aware Compensation in Generative Music" (Zhang et al., 1 Jul 2026) and "AME: A Multi-Type Contributor Attribution Framework in Generative AI Markets" (Shi et al., 15 Jun 2026) use attribution contracts to tie valuation signals to payments, rights mapping, and execution. Other domain-specific uses include vendor-side canary tracing for agent attribution (Chocron et al., 15 May 2026), conformance contracts for resident KV reuse under active KV pressure (Stepanek, 22 May 2026), operator-gated blockchain provenance registries (Moore, 3 Apr 2026), and robust payment design under manipulable distributed-learning attribution (Gao et al., 15 May 2026). This suggests that the term has become a general device for making attribution claims auditable by specifying outputs, evidence, permissions, and failure modes.

Within the narrower attribution literature, the paper positions itself against classifier-era assumptions also discussed in the survey "Algorithmic Contract Theory: A Survey" (Duetting et al., 2024), where payments are tied to observable performance signals that are imperfect proxies for hidden effort, and alongside work that distinguishes corroborative attribution from contributive attribution in LLMs (Worledge et al., 2023). In that context, the Attribution Contract can be understood as a domain-specific formalism for generative-model explanation: not a new attribution algorithm, but a way to state what a given attribution algorithm is claiming to explain (Nguyen, 21 May 2026).