arXiv subjects

Runxiong Wu

Publications and source records attributed to Runxiong Wu.

6 recordsLinked to original sources

arXiv2026

You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model

A frozen language model on reasoning tasks has two coupled weaknesses: it under-uses evidence its own residual stream already encodes, and it fails to detect when the input is insufficient to answer, so it confabulates. This paper consolidates two research lines that address these on the same residual stream: a conditional steering probe writes the stream at mid-stack layers and recovers reasoning accuracy from a frozen backbone, and a zero-shot sufficiency direction reads the stream and abstains when information is insufficient. Deployed in one forward pass they interfere: the steering write shifts the state the direction reads, costing up to 8 AUROC points of cross-domain transfer on small models; a separate clean pass doubles inference cost. We keep the direction fixed and train a small network to reconstruct the pre-steering residual from the steered one -- mean-squared error on (steered, clean) pairs, no sufficiency labels -- and read the direction on the reconstruction. The resulting system, YOPO (You Only Pass Once), answers, steers, and abstains in one forward pass of a frozen Qwen2.5 backbone (1.5B/3B/7B). End to end, three-way accuracy more than doubles the frozen baseline (0.375->0.798 on 1.5B alphaNLI) and one pass beats the two-pass reference at every scale (0.798/0.830/0.893 vs 0.753/0.790/0.863) and on ten backbones across six model families. We chart the capacity-transfer frontier quantifying the principle that abstention should not be trained in; a source-side audit catches our own alphaNLI construction leaking a surface artifact, so architectural claims are anchored on native-label replications (SQuAD2, RepLiQA, MuSiQue); and on the standard four-domain suite we contribute, to our knowledge, the first answer-or-abstain benchmark, where our gate tops every in-domain dataset and the label-free direction is the only gate family to survive domain transfer.

cs.CL↗

arXiv2025

Communication-Efficient Decentralized Optimization via Double-Communication Symmetric ADMM

This paper focuses on decentralized composite optimization over networks without a central coordinator. We propose a novel decentralized symmetric ADMM algorithm that incorporates multiple communication rounds within each iteration, derived from a new constraint formulation that enables information exchange beyond immediate neighbors. While increasing per-iteration communication, our approach significantly reduces the total number of iterations and overall com- munication cost. We further design optimal communication rules that minimize the number of rounds and variables transmitted per iteration. The proposed algorithm is shown to achieve linear convergence under standard and relatively weak assumptions (e.g., metric subregularity). Extensive experiments on regression and classification tasks validate the theoretical results and demonstrate superior performance compared to existing decentralized optimization methods.

math.OC↗

arXiv2025

On the Relationship Between CoCoA and ADMM for Distributed Empirical Risk Minimization

Distributed empirical risk minimization (ERM) is often studied through two influential yet seemingly separate families of methods: CoCoA-type algorithms, derived from distributed dual coordinate ascent, and ADMM-type algorithms, derived from consensus and proximal splitting. In this paper, we investigate the connection of the two types of algorithms from a unified primal-dual perspective. We show that consensus ADMM, linearized consensus ADMM, two distributed proximal ADMM variants, and ridge-regularized CoCoA can all be written in a common update form involving a global primal variable and block dual variables. This reformulation makes several previously hidden connections explicit: For ridge-regularized ERM, CoCoA coincides with a particular proximal ADMM scheme at the level of the dual update. Moreover, consensus ADMM on the primal problem is equivalent to proximal ADMM on the dual problem under an explicit parameter mapping together with a sign reversal of the saddle objective; similar correspondences also hold for the linearized variants. These results indicates that the ADMM-type algorithms, when fine tuned, performs at least as good as CoCoA, under ridge regularized ERM problems. The unified view also yields a natural primal-dual gap stopping criterion for consensus ADMM and a unified $O(1/T)$ ergodic convergence analysis for the ADMM-type methods. Experiments on synthetic regression problems and real SVM datasets support the predicted relationships, clarify the role of tuning parameters, and show that suitably tuned ADMM variants can outperform CoCoA in the ridge-regularized setting.

math.OC↗

arXiv2023

Optimal Sparse Sliced Inverse Regression via Random Projection

We propose a novel sparse sliced inverse regression method based on random projections in a large $p$ small $n$ setting. Embedded in a generalized eigenvalue framework, the proposed approach finally reduces to parallel execution of low-dimensional (generalized) eigenvalue decompositions, which facilitates high computational efficiency. Theoretically, we prove that this method achieves the minimax optimal rate of convergence under suitable assumptions. Furthermore, our algorithm involves a delicate reweighting scheme, which can significantly enhance the identifiability of the active set of covariates. Extensive numerical studies demonstrate high superiority of the proposed algorithm in comparison to competing methods.

stat.ME↗

arXiv2021

High-Dimensional Sparse Single-Index Regression Via Hilbert-Schmidt Independence Criterion

Hilbert-Schmidt Independence Criterion (HSIC) has recently been used in the field of single-index models to estimate the directions. Compared with some other well-established methods, it requires relatively weaker conditions. However, its performance has not yet been studied in the high-dimensional scenario, where the number of covariates is much larger than the sample size. In this article, we propose a new efficient sparse estimate in HSIC based single-index model. This new method estimates the subspace spanned by the linear combinations of the covariates directly and performs variable selection simultaneously. Due to the non-convexity of the objective function, we use a majorize-minimize approach together with the linearized alternating direction method of multipliers algorithm to solve the optimization problem. The algorithm does not involve the inverse of the covariance matrix and therefore can handle the large p small n scenario naturally. Through extensive simulation studies and a real data analysis, we show our proposal is efficient and effective in the high-dimensional setting. The Matlab codes for this method are available online.

stat.ME↗

arXiv2021

MM Algorithms for Distance Covariance based Sufficient Dimension Reduction and Sufficient Variable Selection

Sufficient dimension reduction (SDR) using distance covariance (DCOV) was recently proposed as an approach to dimension-reduction problems. Compared with other SDR methods, it is model-free without estimating link function and does not require any particular distributions on predictors (see Sheng and Yin, 2013, 2016). However, the DCOV-based SDR method involves optimizing a nonsmooth and nonconvex objective function over the Stiefel manifold. To tackle the numerical challenge, we novelly formulate the original objective function equivalently into a DC (Difference of Convex functions) program and construct an iterative algorithm based on the majorization-minimization (MM) principle. At each step of the MM algorithm, we inexactly solve the quadratic subproblem on the Stiefel manifold by taking one iteration of Riemannian Newton's method. The algorithm can also be readily extended to sufficient variable selection (SVS) using distance covariance. We establish the convergence property of the proposed algorithm under some regularity conditions. Simulation studies show our algorithm drastically improves the computation efficiency and is robust across various settings compared with the existing method. Supplemental materials for this article are available.

stat.ML↗