Yingxiao Ye

Philippe Chlenski, Zachariah Carmichael, Ayush Warikoo et al.

Mechanistic interpretability (MI) requires full access to model internals, yet the APIs for most widely deployed language models at best expose log-probabilities over output tokens. This creates a surrogate problem: when do measurements made on open models allow us to make claims about a closed model? We evaluate surrogate fidelity at the prediction, attribution, and representation levels. For binary classification tasks, log-odds provide an API-compatible scalar readout of the model's representation space, and leave-one-out attributions provide insight into model behavior. Across eleven models spanning four families (Llama, Qwen, GPT, and Gemini), we find that prediction fidelity substantially overstates attribution fidelity: models that agree on what the answer is often disagree on why. We document an access-validity inversion: white-box signals like attention patterns and perturbation magnitudes are highly stable across models but only weakly predictive of causal attributions, which black-box input ablations capture by design. Mechanistic insight does not automatically transfer to closed targets, and prediction-level agreement is insufficient to warrant such transfer. Code and results are available at https://github.com/facebookresearch/surrogate.

Phebe Vayanos, Yingxiao Ye, Duncan McElfresh et al.

We study the problem of eliciting the preferences of a decision-maker through a moderate number of pairwise comparison queries to make them a high quality recommendation for a specific problem. We are motivated by applications in high stakes domains, such as when choosing a policy for allocating scarce resources to satisfy basic needs (e.g., kidneys for transplantation or housing for those experiencing homelessness) where a consequential recommendation needs to be made from the (partially) elicited preferences. We model uncertainty in the preferences as being set based and} investigate two settings: a) an offline elicitation setting, where all queries are made at once, and b) an online elicitation setting, where queries are selected sequentially over time in an adaptive fashion. We propose robust optimization formulations of these problems which integrate the preference elicitation and recommendation phases with aim to either maximize worst-case utility or minimize worst-case regret, and study their complexity. For the offline case, where active preference elicitation takes the form of a two and half stage robust optimization problem with decision-dependent information discovery, we provide an equivalent reformulation in the form of a mixed-binary linear program which we solve via column-and-constraint generation. For the online setting, where active preference learning takes the form of a multi-stage robust optimization problem with decision-dependent information discovery, we propose a conservative solution approach. Numerical studies on synthetic data demonstrate that our methods outperform state-of-the art approaches from the literature in terms of worst-case rank, regret, and utility. We showcase how our methodology can be used to assist a homeless services agency in choosing a policy for allocating scarce housing resources of different types to people experiencing homelessness.