Driving · 2026-10-28
Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention
Ayush Mittal, Dhruv Gupta
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors study how changing the deployment policy can break a fixed world model’s uncertainty calibration when only the query distribution over pre-outcomes changes but the predictor stays fixed. Under shared conditional mechanism P(Y|X)=Q(Y|X), they identify an intervention likelihood-ratio w(X)=dQX/dPX and use weighted conformal calibration plus learned ratio estimation to restore target-law coverage; a coverage penalty depends on ratio error (via TV and classifier log-loss regret).
Why it matters
World-model uncertainty guarantees are calibration statements about the distribution under which residuals are assessed. If deployment changes the policy, it changes the measure over queries X, which can invalidate unweighted conformal sets. The paper provides a transport mechanism—weighted conformal calibration using an intervention-induced density ratio—plus bounds connecting ratio-estimation error to target-law coverage, and demonstrates it in a controlled physical robot setting while keeping the world-model predictor frozen.
Method
- Assume intervention preserves the conditional mechanism P(Y|X)=Q(Y|X), so the intervention is identified by the pre-outcome change of measure w(X)=dQX/dPX (for policy-controlled dynamics it factors into occupancy and action-propensity).
- Use weighted conformal calibration that keeps the world-model predictor m fixed and transports marginal coverage from source to target law with penalty terms involving TV(Q, Qhat) and E_P|w−what|.
- Estimate w(X) without outcomes using source–target domain classification: fit a clipped target posterior h(x), then bound ratio error using population binary log-loss regret Δlog(h). The authors machine-check the finite-support algebra/causal identities in Lean 4.
Abstract (from arXiv)
Changing a deployment policy can invalidate a world model's uncertainty while leaving the physical transition mechanism and predictor unchanged. For one-step prediction, if source and target laws share the conditional mechanism $P(Y|X)=Q(Y|X)$, then the intervention is identified by the pre-outcome change of measure $w(X)=dQ_X/dP_X$. For policy-controlled dynamics this ratio factors into state-occupancy and action-propensity terms. Weighted conformal calibration transports marginal coverage to the target law; with an estimated ratio $\hat w$, $P_Q\{Y \in C_{\hat w}(X)\} \ge 1-\alpha-TV(Q,\hat Q) \ge 1-\alpha-E_P|w-\hat w|$. Population log-loss regret for a source-versus-target domain classifier bounds the same ratio error. If $h$ is its clipped target posterior, $\rho$ is the target-domain prior, and $\Delta(h)$ is population log-loss regret, then $E_P|w-\hat w_h| \le (1/(\rho\gamma))\sqrt{\Delta(h)/2}$. An exact cancellation improves the dependence from a generic odds-map bound of order $\gamma^{-2}$ to $(\rho\gamma)^{-1}$. Our experiments use a released DINO-feature world model and physical Franka trajectories. Under the strongest controlled state-action shift, ordinary 90% conformal sets cover 55.0%, whereas oracle and learned-ratio calibration cover 91.0% and 90.9%. The learned surrogate remains within 0.9% total variation of the target law. Calibration is relative to the deployment law: the model can remain fixed while calibration follows the intervention. We machine-check in Lean 4 the finite-support algebraic and causal identities used in the classifier-regret argument, including its coverage-composition step.
Related papers
- Calibration-risk routing for controlled world-model adaptation
- One-Step Prediction Fidelity Is a Poor Proxy for Planning Performance in JEPA World Models
- Beyond Prediction: Steering VLM Agents with Retrospective World Modeling
- How Representational Priors Shape Generalization in World Action Models
- pico-WAM: Evaluating Counterfactual Prediction and Planning Error in a Compact World Action Model
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-08-20 · Indexed by Pub-AI: 2026-10-28 UTC
Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention
Preprint posted: 2026-08-20 · Original record ↗
- arxiv
- 2610.35877
Explicit source versions (1)
- v1 · 2026-08-20
- Preprint available
Source date: 2026-08-20 · First detected: 2026-10-28 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.35877v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Changing a deployment policy can invalidate a world model's uncertainty while leaving the physical transition mechanism and predictor unchanged. For one-step prediction, if source and target laws share the conditional mechanism $P(Y|X)=Q(Y|X)$, then the intervention is identified by the pre-outcome change of measure $w(X)=dQ_X/dP_X$. For policy-controlled dynamics this ratio factors into state-occupancy and action-propensity terms. Weighted conformal calibration transports marginal coverage to the target law; with an estimated ratio $\hat w$, $P_Q\{Y \in C_{\hat w}(X)\} \ge 1-α-TV(Q,\hat Q) \ge 1-α-E_P|w-\hat w|$. Population log-loss regret for a source-versus-target domain classifier bounds the same ratio error. If $h$ is its clipped target posterior, $ρ$ is the target-domain prior, and $Δ(h)$ is population log-loss regret, then $E_P|w-\hat w_h| \le (1/(ργ))\sqrt{Δ(h)/2}$. An exact cancellation improves the dependence from a generic odds-map bound of order $γ^{-2}$ to $(ργ)^{-1}$. Our experiments use a released DINO-feature world model and physical Franka trajectories. Under the strongest controlled state-action shift, ordinary 90% conformal sets cover 55.0%, whereas oracle and learned-ratio calibration cover 91.0% and 90.9%. The learned surrogate remains within 0.9% total variation of the target law. Calibration is relative to the deployment law: the model can remain fixed while calibration follows the intervention. We machine-check in Lean 4 the finite-support algebraic and causal identities used in the classifier-regret argument, including its coverage-composition step.
