This position paper argues that AI agents should be regulated based on the sequence of actions they autonomously take. AI agents with long-term planning and strategic capabilities can pose significant risks of human extinction and irreversible global catastrophes. While existing regulations often focus on computational scale as a proxy for potential harm, we contend that such measures are insufficient for assessing the risks posed by AI agents whose capabilities arise primarily from inference-time computation. To support our position, we discuss relevant regulations and recommendations from AI scientists regarding existential risks, as well as the advantages of action sequences over existing impact measures that require observing environmental states.

The humanoid robot (Unitree G1) demonstrates diverse agile whole-body skills, showcasing the control policies' agility: (a) Cristiano Ronaldo's signature celebration involving a jump with a 180-degree mid-air rotation; (b) LeBron James's "Silencer" celebration involving single-leg balancing; and (c) Kobe Bryant's famous fadeaway jump shot involving single-leg jumping and landing; (d) 1.5m-forward jumping; (e) Leg stretching; (f) 1.3m-side jumping. Abstract -- Humanoid robots hold the potential for unparalleled versatility for performing human-like, whole-body skills. However, achieving agile and coordinated whole-body motions remains a significant challenge due to the dynamics mismatch between simulation and the real world. Existing approaches, such as system identification (SysID) and domain randomization (DR) methods, often rely on labor-intensive parameter tuning or result in overly conservative policies that sacrifice agility. In this paper, we present ASAP (Aligning Simulation and Real Physics), a two-stage framework designed to tackle the dynamics mismatch and enable agile humanoid whole-body skills. Then ASAP fine-tunes pre-trained policies with the delta action model integrated into the simulator to align effectively with real-world dynamics. We evaluate ASAP across three transfer scenarios--IsaacGym to IsaacSim, IsaacGym to Genesis, and IsaacGym to the real-world Unitree G1 humanoid robot. Our approach significantly improves agility and whole-body coordination across various dynamic motions, reducing tracking error compared to SysID, DR, and delta dynamics learning baselines. ASAP enables highly agile motions that were previously difficult to achieve, demonstrating the potential of delta action learning in bridging simulation and real-world dynamics. These results suggest a promising sim-to-real direction for developing more expressive and agile humanoids. I NTRODUCTION For decades, we have envisioned humanoid robots achieving or even surpassing human-level agility. However, most prior work [46, 74, 47, 73, 107, 19, 95, 50] has primarily focused on locomotion, treating the legs as a means of mobility. Recent studies [10, 25, 24, 26, 32] have introduced whole-body expressiveness in humanoid robots, but these efforts have primarily focused on upper-body motions and have yet to achieve the agility seen in human movement.

Digital twins have been actively explored in many engineering applications, such as manufacturing and autonomous systems. However, model discrepancy is ubiquitous in most digital twin models and has significant impacts on the performance of using those models. In recent years, data-driven modeling techniques have been demonstrated promising in characterizing the model discrepancy in existing models, while the training data for the learning of model discrepancy is often obtained in an empirical way and an active approach of gathering informative data can potentially benefit the learning of model discrepancy. On the other hand, Bayesian experimental design (BED) provides a systematic approach to gathering the most informative data, but its performance is often negatively impacted by the model discrepancy. In this work, we build on sequential BED and propose an efficient approach to iteratively learn the model discrepancy based on the data from the BED. The performance of the proposed method is validated by a classical numerical example governed by a convection-diffusion equation, for which full BED is still feasible. The proposed method is then further studied in the same numerical example with a high-dimensional model discrepancy, which serves as a demonstration for the scenarios where full BED is not practical anymore. An ensemble-based approximation of information gain is further utilized to assess the data informativeness and to enhance learning model discrepancy. The results show that the proposed method is efficient and robust to the active learning of high-dimensional model discrepancy, using data suggested by the sequential BED. We also demonstrate that the proposed method is compatible with both classical numerical solvers and modern auto-differentiable solvers.

Rebuttal: thank you for your clarifications. I still think that learning kernel (parameters) from multiple realizations of a GP is not very novel in general, but sufficiently novel in your specific context to get discussed at NIPS. The authors use Gaussian processes to learn human function extrapolation behaviour from human sample data. After a comprehensive literature review, they introduce the main idea of the paper: learn the kernel parameters by maximizing the conditional probability of the extrapolation data given the training data. To allow for flexible kernel shapes, they use spectral mixture kernels.

Weaknesses: A few comments that are needed to be addressed: 1) The first comment is about the presentation of the derivations. There are steps in the appendix, and also in the main text that are skipped. Some of them took me a while to rederive, some I couldn't spend more time to rederive. Some steps are also taken as granted in the main text. It is useful to elaborate on them more.

Summary and Contributions: This paper establishes a concentration inequality in operator norm for the sample co-occurrence matrix C of a regular finite-state Markov chain. This is the matrix whose (i,j) entry counts the fraction of times states i and j co-occur within a time window of fixed size T, with a potential weighting by the difference of their occurrence times. The probability that C-E[C] exceeds eps is shown to be exponentially small in eps 2 * L, where L is the sample size (i.e. the total length of the chain). This concentration inequality is established as a corollary of a general result for the concentration of sample averages for a bounded symmetric-matrix-valued function applied to samples from an ergodic Markov chain. This general result is of independent interest.

This work tackles the problem of estimating the co-occurrence matrix of a Markov chain. The referees were unanimous in the assessment that this is a solid contribution, worthy of being accepted. The only reservations were some missing references in related work, which the authors agreed to discuss in the revision.

The paper describes tricks to scale Bayesian network structure learning to thousands of variables. This is achieved by developing new heuristics for candidate parent set identification and the subsequent order based structure optimization. In general, the paper is clearly written and easy to read. There are issues in editing and style, but the problems do not affect readability (much). The suggested heuristics feel bit ad-hoc, thus the value of the work is eventually judged by empirical evaluation.

Summary and Contributions: The paper considers the problem of sampling from the posterior distribution in Bayesian inference. To be more precise, the paper approaches the question of stochastic sampling that relies only on minibatches of data at each iteration. To achieve rapid mixing between isolated modes, the authors consider parallel tempered chains and introduce replica-exchange steps into the stochastic Nose-Hoover Dynamics. The crux of this approach is the stochastic test for the replica-exchange step. To develop such a test, the authors follow the paper [An efficient minibatch acceptance test for metropolis-hastings], which introduces the concept of correction distribution.

The paper proposes a novel MCMC-type algorithm to perform Bayesian inference on large datasets. The paper is a mixture of replica exchange, Nose-Hoover dynamics and non-standard acceptance criterion to deal with mini-batches. All the reviewers participated actively to the discussion after the rebuttal was made available. Although all the ingredients of the proposed method do exist, their combination is original and potentially useful for the ML literature as pointed out by most reviewers. Theorem 2 is also neat and proposes a nice way to propose swaps between replicas using mini-batches.