arXiv subjects

Longxu Zhang

Publications and source records attributed to Longxu Zhang.

3 recordsLinked to original sources

arXiv2026

Action-Prior Denoising for Smooth Real-Time Chunking

Real-time chunking (RTC) lets chunked action policies operate under inference delay by conditioning a newly generated action chunk on actions already committed by the previous chunk. Training-time RTC simulates this delay during learning and avoids expensive guidance at deployment, but its binary prefix mask treats all non-prefix tokens as fully unconstrained. This under-models asynchronous execution: early overlap actions are fixed, while later overlap actions remain editable but should still stay close to the previous plan. We propose Soft RTC, a training-time RTC generalization based on action-prior denoising. Soft RTC constructs corrupted overlap tokens from partially denoised states instead of pure noise and injects the aligned previous chunk as the same prior during inference through a lightweight token-wise blending rule. On the 12 released large Kinetix levels, a short soft window nearly matches hard training-time RTC in overall solve rate (0.809 vs. 0.815), while a medium window reduces high-delay action delta and jerk by 9.1% and 9.6% relative to hard RTC. Both variants keep near-naive runtime, unlike inference-time RTC baselines. A small preliminary real-robot sorting study provides additional evidence that training-time RTC can improve completion and that Soft RTC gives the lowest commanded-action finite-difference metrics among the tested policies.

cs.RO↗

arXiv2026

Revealing Strategic Interactions in Network Games Under Decaying Active Probing

Revealing the interaction topology underlying strategic behavior is fundamental to prediction, intervention, and policy design in networked systems. Yet the interaction matrix is often unobservable, and passive observation of repeated actions fails to provide sufficient excitation for reliable recovery. This paper studies topology recovery in repeated linear-quadratic network games under decaying active probing, where probing inputs are injected into a subset of players and the unknown interaction matrix is inferred from the resulting action trajectories. We first characterize a structural recoverability condition that determines when noiseless probing experiments can make the interaction matrix identifiable. We then show that, under suitable stability and controllability assumptions, a concrete decaying probing signal guarantees exact finite-step recovery while preserving convergence of the repeated-play process. To handle decision perturbations, we further develop a reweighted sparse estimator that achieves almost-sure consistency together with finite-time exact support recovery. These results clarify what can be recovered in both noiseless and perturbed settings.

math.OC↗

arXiv2025

On the Fundamental Limit of the Stochastic Gradient Identification Algorithm Under Non-Persistent Excitation

Stochastic gradient (SG) methods are fundamental to system identification and machine learning, enabling online parameter estimation in large-scale and streaming-data settings. As a classical identification method, the SG algorithm has been extensively studied for decades. Under non-persistent excitation, the strongest currently available convergence result assumes that the condition number of the Fisher information matrix is \(O((\log r_n)^\alpha)\), where \(r_n = 1 + \sum_{i=1}^n \|\varphi_i\|^2\). Existing theory establishes strong consistency when \(\alpha \le 1/3\), whereas the same condition with \(\alpha > 1\) is insufficient to guarantee strong consistency. We prove that strong consistency holds throughout the range \(0 \le \alpha < 1\). The proof is based on a new algebraic framework that yields substantially sharper matrix norm bounds. This result nearly resolves the four-decade-old Chen--Guo conjecture by establishing strong consistency throughout the previously open range \(1/3 < \alpha < 1\).

math.OC↗