acceptodds
Desk rejected at ICLR 2027

The Experiment Defines What Is Measurable

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Must causal representation learning reconstruct an entire latent world when the scientific task asks for one target-wise measurement? Two stochastic hard interventions on a known target define such a measurement before any representation model is chosen: their density ratio cancels every non-target mechanism, and the Bayes environment logit reads out the remaining scalar quantity. Observation determines what this latent estimand becomes. Under an arbitrary invariant observation channel, the observable ratio is a posterior average of the latent intervention ratio, yielding a coordinate, a quotient, or a posterior observable. Alignment across entities or acquisition conditions imposes a separate requirement. The intervention laws also control the semantics and conditioning of the measurement. Exponential tilts program a chosen target statistic; among monotone tilts, affine statistics uniquely optimize worst-case conditioning under a weak symmetric-KL budget. Controlled experiments recover the programmed statistic across linear, monotone, and non-injective choices and predict a renderer-induced quotient before training. Matched comparisons with ILCM, CITRIS, BISCUIT, iVAE, and GSCALE-I produce the corresponding crossed profile: target-wise classification favors the programmed quotient, whereas interfaces designed to recover state or intervention structure retain more signed-state and intervention information. Natural-image and chemical-sensor studies further separate within-context discrimination from alignment under unseen entities and acquisition drift. The first identification question is therefore which measurement the experiment induces, before attempting full latent reconstruction.

Primary area:
causal reasoning
Keywords:
Causal Representation Learning, Interventional Causal Inference, Identifiability, Intervention Design, Density-Ratio Estimation

Related papers

Discussion (0)

Sign in to comment.