arXiv subjects

Zichen Xu

Publications and source records attributed to Zichen Xu.

10 recordsLinked to original sources

arXiv2026

Is More Privileged Information Better? From Solution Traces to Problem-Solving Structure in Self-Distilled Reasoning

On-policy self-distillation (OPSD) improves reasoning by using a privileged view of a model conditioned on reference solutions to supervise a student view that observes only the question. However, the teacher-provided token-level targets may depend on reference-specific information unavailable at inference time. We propose Problem-Space-Guided OPSD (PS-OPSD), which replaces the complete solution with trajectory-grounded guidance describing the initial state, goal conditions, constraints, and a selected state-transition path. The student rollout and OPSD objective remain unchanged. Across three mathematical reasoning benchmarks and model scales ranging from 1.7B to 8B, PS-OPSD achieves the highest aggregate question-only accuracy among the compared methods. Controlled experiments further indicate that guidance relevance and path coherence contribute to these gains, highlighting the representation of privileged information as an important design choice in OPSD.

cs.AI↗

arXiv2026

Latent Thought Credit: Multi-Answer Credit Assignment for Latent Reasoning

Latent reasoning allows language models to carry out intermediate reasoning in continuous latent representations rather than fully externalizing it as discrete chains of thought. However, assigning credit to such latent thoughts from answer-only rewards is difficult: a single final answer mixes thought quality with answer-sampling noise. We propose \textbf{Latent Thought Credit (LTC)}, a hierarchical credit-assignment framework for latent reasoning. For each prompt, LTC samples multiple latent thoughts, fixes the context after each thought, and estimates thought-level expected reward by averaging rewards over multiple answers generated from that fixed context. LTC uses thought-level advantages to optimize the latent-thought phase, answer-level advantages to optimize the answer phase, and an advantage-weighted thought-matching objective that helps the policy reproduce high-credit latent thoughts. We instantiate LTC in a GRPO-style on-policy training framework and evaluate it across mathematical reasoning and STEM multiple-choice tasks. LTC achieves the best average accuracy among the compared methods, while ablations and fixed-context diagnostics show that multi-answer estimation reduces reward-estimation error and mitigates ambiguous or incorrect thought-level credit.

cs.AI↗

arXiv2026

TARPO: Token-Wise Latent-Explicit Reasoning via Action-Routing Policy Optimization

Latent reasoning has emerged as a promising alternative to discrete Chain-of-Thought (CoT) in large language models (LLMs), enabling more expressive reasoning by operating over continuous representations. However, the inherently deterministic nature of continuous representations limits policy exploration in reinforcement learning (RL). To address this, we propose TARPO (Token-Wise Latent-Explicit Reasoning via Action-Routing Policy Optimization), a pure RL framework that adaptively switches between discrete token generation and continuous latent reasoning at each step. TARPO introduces a lightweight action head router that observes the current hidden state and samples a routing decision from a binary mode-selection space, preserving the stochasticity of discrete token sampling from the vocabulary. The LLM backbone and router are jointly optimized end-to-end with a shared group-relative advantage signal. Extensive experiments across Qwen2.5 (from 1.5B to 7B) and Llama-3.1-8B backbones demonstrate that TARPO consistently outperforms existing explicit and latent reasoning RL baselines across diverse benchmarks. Further analysis shows that TARPO learns adaptive token-wise switching behaviors while maintaining stable training dynamics. Our code is available at https://github.com/NKU-LITI/TARPO-master.

cs.CL↗

arXiv2026

Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning

Post-training has become central to turning pretrained large language models (LLMs) into aligned, capable, and deployable systems. Recent progress spans supervised fine-tuning (SFT), preference optimization, reinforcement learning (RL), process supervision, verifier-guided methods, distillation, and multi-stage pipelines. Yet these methods are often discussed in fragmented ways, organized by labels or objectives rather than by the behavioral bottlenecks they address. This survey argues that LLM post-training is best understood as structured intervention on model behavior. We organize the field first by trajectory provenance, which defines two primary regimes: off-policy learning on externally supplied trajectories and on-policy learning on learner-generated rollouts. We then interpret methods through two recurring roles -- effective support expansion, which makes useful behaviors more reachable, and policy reshaping, which improves behavior within already reachable regions -- together with a complementary systems-level role, behavioral consolidation, which preserves, transfers, and amortizes useful behavior across stages and model transitions. Under this view, SFT may serve either support expansion or policy reshaping; preference optimization is usually off-policy reshaping, though online variants move closer to learner-generated states. On-policy RL often improves behavior on learner-generated states, but stronger guidance can also make hard-to-reach reasoning paths reachable. Distillation is often better understood as consolidation rather than only compression, and hybrid pipelines emerge as coordinated multi-stage compositions. Overall, the framework helps diagnose post-training bottlenecks and reason about stage composition, suggesting that progress increasingly depends on coordinated systems design rather than any single dominant objective.

cs.CL↗

arXiv2026

CD-Raft: Reducing the Latency of Distributed Consensus in Cross-Domain Sites

Today's massive AI computation loads push heavy data synchronization across sites, i.e., nodes in data centers. Any reduction in such consensus latency can significantly improve the overall performance of desired systems. This consensus challenge explosively peaks at cross-domain sites. In this paper, we proposed CD-Raft to address the cross-domain latency challenge, an optimized Raft protocol for strong consistency in cross-domain sites. CD-Raft can significantly reduce consensus latency by optimizing cross-domain round-trip time (RTT) for reads and writes, as well as carefully positioning the leader node. We verified the correctness of CD-Raft in a formal specification using the TLA+ specification, guaranteeing the strong consistency across sites. We have prototyped CD-Raft and evaluated it using the YCSB benchmark. Empirical results show that compared to the classic Raft, CD-Raft reduces the average latency by 32.90% and (99th percentile) tail latency by 49.24% for renown traces across multiple sites.

cs.DC↗

arXiv2026

Nezha: A Key-Value Separated Distributed Store with Optimized Raft Integration

Distributed key-value stores are widely adopted to support elastic big data applications, leveraging purpose-built consensus algorithms like Raft to ensure data consistency. However, through systematic analysis, we reveal a critical performance issue in such consistent stores, i.e., overlapping persistence operations between consensus protocols and underlying storage engines result in significant I/O overhead. To address this issue, we present Nezha, a prototype distributed storage system that innovatively integrates key-value separation with Raft to provide scalable throughput in a strong consistency guarantee. Nezha redesigns the persistence strategy at the operation level and incorporates leveled garbage collection, significantly improving read and write performance while preserving Raft's safety properties. Experimental results demonstrate that, on average, Nezha achieves throughput improvements of 460.2%, 12.5%, and 72.6% for put, get, and scan operations, respectively.

cs.DC↗

arXiv2022

Cost-effective BlackWater Raft on Highly Unreliable Nodes at Scale Out

The Raft algorithm maintains strong consistency across data replicas in Cloud. This algorithm divides nodes into leaders and followers, to satisfy read/write requests spanning geo-diverse sites. With the increase of workload, Raft shall provide scale-out performance in proportion. However, traditional scale-out techniques encounter bottlenecks in Raft, and when the provisioned sites exhaust local resources, the performance loss will grow exponentially. To provide scalability in Raft, this paper proposes a cost-effective mechanism for elastic auto-scaling in Raft, called BlackWater-Raft or BW-Raft. BW-Raft extends the original Raft with the following abstractions: (1) secretary nodes that take over expensive log synchronization operations from the leader, relaxing the performance constraints on locks. (2) massive low cost observer nodes that handle reads only, improving throughput for typical data intensive services. These abstractions are stateless, allowing elastic scale-out on unreliable yet cheap spot instances. In theory, we demonstrate that BW-Raft can maintain Raft's strong consistency guarantees when scaling out, processing a 50X increase in the number of nodes compared to the original Raft. We have prototyped the BW-Raft on key-value services and evaluated it with many state-of-the-arts on Amazon EC2 and Alibaba Cloud. Our results show that within the same budget, BW-Raft's resource footprint increments are 5-7X smaller than Multi-Raft, and 2X better than original Raft. Using spot instances, BW-Raft can reduces costs by 84.5\% compared to Multi-Raft. In the real world experiments, BW-Raft improves goodput of the 95th-percentile SLO by 9.4X, thus serving as an alternative for services scaling out with strong consistency.

cs.DC↗

arXiv2021

Wavelet Transform-assisted Adaptive Generative Modeling for Colorization

Unsupervised deep learning has recently demonstrated the promise of producing high-quality samples. While it has tremendous potential to promote the image colorization task, the performance is limited owing to the high-dimension of data manifold and model capability. This study presents a novel scheme that exploits the score-based generative model in wavelet domain to address the issues. By taking advantage of the multi-scale and multi-channel representation via wavelet transform, the proposed model learns the richer priors from stacked coarse and detailed wavelet coefficient components jointly and effectively. This strategy also reduces the dimension of the original manifold and alleviates the curse of dimensionality, which is beneficial for estimation and sampling. Moreover, dual consistency terms in the wavelet domain, namely data-consistency and structure-consistency are devised to leverage colorization task better. Specifically, in the training phase, a set of multi-channel tensors consisting of wavelet coefficients is used as the input to train the network with denoising score matching. In the inference phase, samples are iteratively generated via annealed Langevin dynamics with data and structure consistencies. Experiments demonstrated remarkable improvements of the proposed method on both generation and colorization quality, particularly in colorization robustness and diversity.

cs.CV↗

arXiv2021

A Local Similarity-Preserving Framework for Nonlinear Dimensionality Reduction with Neural Networks

Real-world data usually have high dimensionality and it is important to mitigate the curse of dimensionality. High-dimensional data are usually in a coherent structure and make the data in relatively small true degrees of freedom. There are global and local dimensionality reduction methods to alleviate the problem. Most of existing methods for local dimensionality reduction obtain an embedding with the eigenvalue or singular value decomposition, where the computational complexities are very high for a large amount of data. Here we propose a novel local nonlinear approach named Vec2vec for general purpose dimensionality reduction, which generalizes recent advancements in embedding representation learning of words to dimensionality reduction of matrices. It obtains the nonlinear embedding using a neural network with only one hidden layer to reduce the computational complexity. To train the neural network, we build the neighborhood similarity graph of a matrix and define the context of data points by exploiting the random walk properties. Experiments demenstrate that Vec2vec is more efficient than several state-of-the-art local dimensionality reduction methods in a large number of high-dimensional data. Extensive experiments of data classification and clustering on eight real datasets show that Vec2vec is better than several classical dimensionality reduction methods in the statistical hypothesis test, and it is competitive with recently developed state-of-the-art UMAP.

cs.LG↗

arXiv2021

DEAL: Decremental Energy-Aware Learning in a Federated System

Federated learning struggles with their heavy energy footprint on battery-powered devices. The learning process keeps all devices awake while draining expensive battery power to train a shared model collaboratively, yet it may still leak sensitive personal information. Traditional energy management techniques in system kernel mode can force the training device entering low power states, but it may violate the SLO of the collaborative learning. To address the conflict between learning SLO and energy efficiency, we propose DEAL, an energy efficient learning system that saves energy and preserves privacy with a decremental learning design. DEAL reduces the energy footprint from two layers: 1) an optimization layer that selects a subset of workers with sufficient capacity and maximum rewards. 2) a specified decremental learning algorithm that actively provides a decremental and incremental update functions, which allows kernel to correctly tune the local DVFS. We prototyped DEAL in containerized services with modern smartphone profiles and evaluated it with several learning benchmarks with realistic traces. We observed that DEAL achieves 75.6%-82.4% less energy footprint in different datasets, compared to the traditional methods. All learning processes are faster than state-of-the-practice FL frameworks up to 2-4X in model convergence.

cs.LG↗