Search arXiv⌕ Search

arXiv · 2610.35877

Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention

Abstract

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.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ayush Mittal, Dhruv Gupta. 2026-08-20. Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention. https://arxiv.org/abs/2610.35877

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

A Theoretical Bridge Between Long-Tailed Recognition and Continual Learning

We theoretically and empirically establish a previously unexplored connection between Long-Tailed Recognition (LTR) and Continual Learning (CL). Specifically, we show that training on a long-tailed dataset with imbalance factor $\mathrm{IF}$ drives model parameters into an $O(1/\sqrt{\mathrm{IF}})$ neighborhood of the dominant-class solution, under both uniform and exponentially decaying cardinality distributions. Building on this result, we establish an objective-level reformulation of LTR as a sequential learning problem and prove that the CL objective upper-bounds the balanced LTR loss up to a constant determined by the residual Head loss. Motivated by this connection, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled framework that leverages standard off-the-shelf CL methods to sequentially learn Head and Tail classes while mitigating catastrophic forgetting. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions and demonstrate strong performance across diverse LTR benchmarks. In the foundation-model regime, CLTR with DualPrompt on a CLIP backbone outperforms standard adaptation baselines and is competitive with specialized methods using external semantic supervision. Our work bridges LTR and CL both theoretically and empirically, providing a principled approach for addressing long-tailed learning using standard CL strategies.

cs.LG↗

TRAM: Test-Time Risk Adaptation with Mixture of Agents

Deployed reinforcement learning agents often face safety requirements that are specified only after training: new hazard maps, revised risk thresholds, or behavioral alignment constraints. We study zero-update deployment-time adaptation in which a fixed library of risk-neutral source policies must be reused under a newly specified reward-risk tradeoff. We propose TRAM (Test-time Risk Adaptation with Mixture of Agents), a source-scored composition rule that evaluates each source under the target reward and an occupancy-based deployment risk, then selects actions using risk-adjusted source scores. Unlike training-time risk-sensitive methods tied to a fixed surrogate such as return variance, TRAM supports spatial barrier exposure, divergence to a reference behavior, and local volatility risks specified at test time. We make the surrogate nature of the method explicit: TRAM is not claimed to solve the full occupancy-control problem of the stitched policy, but it admits a measurable source-hull mismatch term connecting source-scored risk to realized risk. Experiments in gridworlds, MuJoCo Reacher, Safety-Gymnasium, and an inference-time LLM alignment setting show that TRAM improves deployment risk while preserving reward and requires no parameter updates at test time.

cs.LG↗

Towards the Worst-case Robustness of Large Language Models

Recent studies have revealed the vulnerability of large language models to adversarial attacks, where adversaries craft specific inputs to induce incorrect or harmful outputs. Although various empirical defenses have been proposed, their worst-case robustness remains poorly understood, raising concerns about how they would behave against stronger future attacks. Here, we study the worst-case robustness of LLMs from both empirical and theoretical perspectives. First, we provide tighter empirical upper bounds for the tested deterministic defenses using stronger white-box attacks, under which most tested configurations achieve nearly 0\% robustness. Second, we formulate randomized-smoothing certification for bounded and binary base functions as fractional and 0-1 knapsack problems, respectively, and derive grouped solvers for two token kernels. For example, on AdvBench, we certify a GPT-4o safety detector smoothed by a uniform kernel to an average certified $\ell_0$ radius of 2.02. In our stated suffix-attack setting, the same kernel certifies a Llama-3-8B safety detector to an average certified radius of 6.41.

cs.LG↗