The Attribution Contract: Feature Attribution for Generative Language Models

Technical Summary: The Attribution Contract

Problem Statement
Feature attribution methods, designed to identify which input features drive a model's output, face a fundamental conceptual ambiguity when applied to generative language models (LMs). In classical classification settings, the input-output schema is static ($x \to f(x)$), and features are clearly defined as input components. However, generative LMs operate via dynamic processes where earlier outputs become inputs for subsequent predictions (in autoregressive models) or where generation proceeds through iterative denoising (in diffusion models).

This structural difference creates an "attribution ambiguity": it is often unclear what constitutes a "feature" or what specific "output" is being explained. For instance, in an autoregressive model, previously generated tokens are both outputs of the model and inputs to the next prediction. Consequently, the same attribution algorithm can yield different results depending on whether the goal is to explain a local next-token prediction or the influence of the original prompt on the entire generated sequence. The paper argues that many perceived disagreements in the literature regarding feature attribution in generative models are not conflicts between algorithms, but rather conflicts arising from unstated or mismatched explanatory contracts.

Methodology: The Attribution Contract
To resolve this ambiguity, the authors introduce the Attribution Contract, a formal specification tuple $A = (S, C, O, P, E)$ that defines the explanatory scope of any feature-attribution claim. The components are:

• $S$ (Score): The specific model quantity being attributed (e.g., logit, probability, loss, or log-likelihood).
• $C$ (Conditioning): What variables are held fixed during the attribution computation (e.g., holding the generated prefix fixed while attributing only to the prompt).
• $O$ (Output): The specific target being explained (e.g., a single next token, an intermediate diffusion state, or a full generated span).
• $P$ (Process): The assumed generative process (e.g., autoregressive chain $x \to y_1 \to \dots \to y_t$ or diffusion chain $z_T \to \dots \to z_0 \to y$).
• $E$ (Eligible Features): The set of features allowed to receive attribution (e.g., prompt tokens only, prompt plus generated prefix, or denoising stages).

The paper demonstrates that the "Attribution Contract" framework allows for the precise instantiation of different settings. It categorizes existing and potential settings into:

1. Autoregressive Settings:
   • Local Next-Token: Attributes the prediction of token $y_t$ to both prompt and generated prefix tokens.
   • Prompt-Conditioned Token: Attributes $y_t$ only to prompt tokens, holding the generated prefix fixed.
   • Span-Level Prompt: Attributes the probability of a full generated span $y_{1:T}$ to prompt tokens, holding the span fixed.
2. Diffusion Settings:
   • State-Level: Attributes an intermediate state $z_t$ to prompt and prior states.
   • Denoising-Stage: Attributes properties of the final output to specific stages of the denoising process.
   • Prompt-to-Output: Attributes the final output $y$ to prompt tokens, treating the denoising chain as the process.

Key Contributions
The paper makes four primary contributions:

1. Identification of Contract Ambiguity: It identifies "contract ambiguity" as a distinct failure mode in generative feature attribution, where the lack of a specified explanatory setting leads to misinterpretation.
2. The Attribution Contract Framework: It proposes the $(S, C, O, P, E)$ tuple to explicitly define the target, eligible features, generative process, conditioning regime, and attributed score for any claim.
3. The Self-Attribution Fallacy: It defines a specific interpretive error where attribution to generated-prefix tokens (valid under a local next-token contract) is incorrectly interpreted as answering a prompt-level question (e.g., "did the source document cause this hallucination?"). The paper illustrates that answering the latter requires a prompt-conditioned contract where the prefix is held fixed.
4. Method-Contract Evaluation: It argues that feature-attribution methods in generative LMs should be evaluated as method–contract pairs rather than as context-free algorithms. A method cannot be judged "good" or "bad" in isolation; its validity depends on whether it aligns with the specific contract being tested.

Results and Analysis
The paper does not present new empirical benchmarks or novel algorithms but provides a theoretical analysis and re-evaluation of existing literature through the lens of the Attribution Contract.

• Reconciliation of Literature: The authors show that methods often perceived as competing (e.g., ContextCite vs. Inseq) are actually answering different questions defined by different contracts. ContextCite operates under a prompt-to-output contract, while Inseq typically operates under a local next-token contract.
• Evaluation Implications: The paper details how evaluation metrics must change based on the contract. For example, evaluating a "span-level prompt attribution" requires perturbing the prompt and measuring the change in the log-probability of the original fixed generation, not regenerating a new sequence. Regenerating a sequence would test a different contract (e.g., prompt-to-output).
• Diffusion Specifics: The analysis highlights that diffusion models require contracts that account for intermediate states and denoising stages, which have no direct analog in autoregressive models.

Significance and Claims
The paper claims that the Attribution Contract is necessary to bring rigor to feature attribution in generative modeling. By making the explanatory setting explicit, the framework:

• Clarifies why the same attribution method can produce different maps depending on instantiation.
• Prevents the "self-attribution fallacy," ensuring that practitioners do not misattribute the cause of a generation to the prompt when the attribution was actually driven by the model's own prior tokens.
• Establishes that there is no single "good" attribution map for generative models; evaluation must be contract-specific.

The authors position this work not as a replacement for existing attribution algorithms, but as a necessary specification layer to make the claims of these methods comparable and interpretable. The significance lies in shifting the discourse from debating algorithms to defining the precise explanatory questions those algorithms are intended to answer.