· Xiaojing Yang · Explainability and Responsible AI · 4 min read
中文Utility Functions: Which Model Behaviours Are We Explaining?
Training-data attribution depends on the utility function. Quality, factuality, style, safety, fairness, and terminology may point to different influential data.
Core idea
There is no attribution score in the abstract. Every score is attribution with respect to a utility: accuracy, COMET, factuality, toxicity, terminology preservation, style, robustness, or a combined objective.
This note is part of my series Training Data Attribution for NLP and LLM Research. The series is written as both a research notebook and an interview preparation path: each article should help me explain the idea clearly, connect it to my thesis, and identify what would become future PhD work.
Guiding question: When we assign data attribution, what model behaviour is the score about?
Intuition
A dataset can help one behaviour and hurt another. Domain data may improve petroleum terminology while making general-domain fluency worse. Safety filtering may reduce harmful output while lowering helpfulness on some queries. If the utility is vague, the attribution claim becomes vague too.
In NLP and LLM research, this matters because model behaviour is deeply shaped by data mixture. A model may be fluent because of broad web text, domain-accurate because of specialised documents, safer because of curated instruction data, or biased because of repeated patterns in a subset of the corpus. Training-data attribution gives us language for asking these questions systematically instead of only saying “the data matters”.
Formal lens
Define u_j(D) for behaviour j. A scalar utility gives one attribution vector phi^j. A multi-faceted evaluation gives several vectors: phi^quality, phi^terminology, phi^style, phi^fairness. Comparing these vectors is often more informative than collapsing everything into one number.
The important discipline is to define the attribution setup before interpreting the score:
| Design choice | Question to answer |
|---|---|
| Attribution unit | What receives credit: source, group, document, example, or token? |
| Utility function | Which behaviour is being explained: quality, terminology, style, factuality, or safety? |
| Intervention | Are we adding, deleting, reweighting, correcting, or retraining? |
| Estimator | Is the score exact, sampled, gradient-based, surrogate-based, or heuristic? |
| Uncertainty | How stable is the score across seeds, samples, metrics, and evaluation sets? |
NLP / LLM example
In English—Norwegian MT, I would not rely only on BLEU. I would report chrF or COMET for general quality, terminology metrics for domain terms, and style or written-standard indicators when relevant. The same data group may have positive attribution for terminology and negative attribution for style consistency.
This is why I do not want to treat attribution as a generic interpretability topic. For my profile, the natural connection is multilingual and domain-specific NLP: low-resource settings, technical terminology, written-standard variation, and evaluation beyond one headline metric.
Connection to my thesis
In my thesis narrative, training-data attribution is useful because it turns a vague data question into an experimental design:
- define interpretable data units;
- define the model behaviour to explain;
- compare controlled data coalitions or interventions;
- estimate contribution;
- report uncertainty and limitations;
- decide what evidence is strong enough to support a causal-style claim.
That structure helps me avoid overclaiming. A score is not automatically a causal explanation. It is a measurement produced by a specific setup.
What I have done, understand, and would extend
| Level | Status |
|---|---|
| Already completed / thesis-ready | Group-level attribution, coalition thinking, metric-based utilities, cautious interpretation, random baselines, bootstrap-style reliability checks. |
| I understand but may not fully implement yet | Instance-level gradient attribution, influence functions, TracIn, Monte Carlo Shapley, surrogate/datamodel approximations. |
| Strong PhD extension | Hierarchical attribution, intervention-based validation, factuality/style-specific utilities, scalable attribution for LLM data mixtures. |
Interview answer
A key design choice in my project is the utility function. I would define separate utilities for translation quality, terminology, and style, because a single Shapley value cannot explain all behaviours. This also makes the work more research-like: it turns attribution from a ranking exercise into a structured evaluation problem.
References and reading path
- Lloyd Shapley, A Value for n-Person Games.
- Ghorbani and Zou, Data Shapley: Equitable Valuation of Data for Machine Learning.
- Koh and Liang, Understanding Black-box Predictions via Influence Functions.
- Pruthi et al., Estimating Training Data Influence by Tracing Gradient Descent.
- Ilyas et al., Datamodels: Predicting Predictions from Training Data.
- Rei et al., COMET: A Neural Framework for MT Evaluation.