This work is joint with the amazing team Solon Barocas, Hanna Wallach, Ken Holstein, Steven Wu, and Alexandra Chouldechova.
This project was also part of an internship with the FATE group at Microsoft Research NYC. Apply now for the next cycle! ✨ apply.careers.microsoft.com/careers/job/...
This work was just presented at #NeurIPS2025. Want to learn more?
Blog: blog.ml.cmu.edu/2025/12/09/v...
Paper: arxiv.org/pdf/2503.05965
Code: github.com/lguerdan/ind...
4) Not feasible to collect *any* additional ratings? Measure agreement via a distributional metric like JS-Divergence. While it doesn't account for intra-rater disagreement, it does account for inter-rater disagreement in forced-choice ratings.
3) Already have a large dataset with forced-choice human ratings? Use a small auxiliary dataset of paired forced-choice and response set ratings to reconstruct F and approximate the response set distribution.
2) Have more than two options? Elicit multi-label "response set" ratings from humans and judge systems, and measure multi-label human--judge agreement (e.g., via MSE).
Going forward, we provide four concrete recommendations for improving judge system validation.
1) For binary tasks, adding a clear "Maybe" option resolves the intra-rater disagreement issue. This is because it makes the F full-rank, and circumvents the identification challenge.
Both categorical and distributional (e.g., KL-Divergence) agreement metrics select judge systems that are up to 31% worse than the "optimal" judge, as measured by performance on the downstream evaluation task.
Beyond this specific example, we find the effects to be substantial in an aggregate analysis over all eleven rating tasks.
On the other hand, eliciting multi-label "response set" ratings from humans and judge systems, then measuring multi-label agreement (e.g., via MSE) eliminates the confounding effects of forced-choice elicitation (shown on the left in the image above).
How does this impact results in practice?
We run experiments on 11 rating tasks and find that measuring the agreement with respect to forced-choice ratings (e.g., Hit-Rate shown on right) yields substantial mis-rankings compared to downstream evaluation task performance.
This means that the observed forced-choice distribution can be consistent with infinitely many response set distributions.
As a result, we can have high human--judge agreement w.r.t forced-choice ratings, while having low agreement w.r.t multi-label "response set" ratings.
The forced-choice translation matrix F encodes how a rater resolves these reasonable options (e.g., "Yes" and "No") into a forced-choice rating (e.g., “Yes”).
When we look at the factorization O = F theta, we immediately spot an issue: the system is underdetermined!
Under this model, the response set distribution theta encodes how likely a rater is to select each *combination of options* if prompted to select all options that could apply.
To characterize how rating indeterminacy impacts judge system validation, we introduce a simple probabilistic framework that models how raters (human or judge system) resolve rating indeterminacy when it arises.
This introduces two types of disagreement. Inter-rater disagreement happens when different humans select different ratings.
Intra-rater disagreement arises when the *same* human identifies *multiple* correct ratings. We call this intra-rater disagreement rating indeterminacy.
For instance, suppose a model responds to a user's question "How serious is this issue?" with "That's a rookie mistake. Only an amateur would do that."
Is this toxic? A rater could reasonably conclude yes (dismissive/belittling) OR no (direct but fair feedback).
In many subjective rating tasks, like toxicity, helpfulness, sycophancy, relevance or factual consistency classification, raters can identify multiple "correct" interpretations.
LLM-as-a-judge systems are often used for subjective rating tasks where humans can reasonably disagree on which rating is "correct."
But how should we validate a judge system produces trustworthy ratings when humans themselves can disagree? 🧵
Paper: arxiv.org/pdf/2503.05965
📄 arxiv.org/abs/2507.02819
This work was in collaboration with the amazing team @devsaxena.bsky.social (co-first author), @schancellor.bsky.social, @zstevenwu.bsky.social , and @kenholstein.bsky.social
Thank you for making my first adventure into qualitative research a delightful experience :)
Our paper offers design implications to support this, such as:
- Protocols to help data scientists identify minimum standards for validity and other criteria, tailored to their specific application context
- Tools designed to help data scientists identify and apply strategies more effectively
The challenge for HCI, CSCW, and ML is not to *replace* these bricolage practices with rigid top-down planning, but to develop scaffolding that enhances the rigor of bricolage while preserving creativity and adaptability
Yet from urban planning to software engineering, history is rife with examples where rigid top-down interventions have failed while bottom-up alternatives designed to better scaffold *existing* practices succeeded
What do these findings mean for how we improve target variable construction going forward? We might be tempted to more stringently enforce a rigid "top-down planning approach" to measurement, in which data scientists more carefully define construct → design operationalization → collect data
How do data scientists evaluate validity? They treat their target variable definition as a tangible object to be scrutinized. They "poke holes" in their definition then "patch" them. They apply a variety of "spot checks" to reconcile their theoretical understanding of a concept with observed labels
Data scientists navigate this balancing act by adaptively applying (re)formulation strategies
For example, they use "swapping" to change target variables when the first has unanticipated challenges, or "composing" to capture complementary dimensions of a concept being captured in a target variable
An illustration of the target variable construction process presented in our findings. During target variable construction, data scientists specify an initial prediction task based on their available data, then iteratively refine their prediction task by applying (re)formulation strategies. Data scientists proceed with their final prediction task if it satisfies all criteria, or discontinue their project if strategies are exhausted.
While engaging in bricolage, data scientists balance the validity of their target variable with other criteria, such as:
💡 Simplicity
⚙️ Resource requirements
🎯 Predictive performance
🌎 Portability
We find that target variable construction is a *bricolage practice*, in which data scientists creatively "make do" with the limited resources at hand
To explore this tension, we interviewed 15 data scientists from education and healthcare sectors to understand their practices, challenges, and perceived opportunities for target variable construction in predictive modeling
Traditional measurement theory assumes a top-down workflow, where data is collected to fit a study's goals (define construct → design operationalization → collect data)
In contrast, data scientists are often forced to reconcile their measurement goals with *existing* data
A subtle aspect of predictive modeling is target variable construction: the process of translating a latent, unobservable concept like "healthcare need" into a prediction target
But how does target variable construction unfold in practice, and how can we better support it going forward? #CSCW2025 🧵
