arXiv subjects

Yingquan Wu

Publications and source records attributed to Yingquan Wu.

5 recordsLinked to original sources

arXiv2026

K2-V2: A 360-Open, Reasoning-Enhanced LLM

We introduce K2-V2, a 360-open LLM built from scratch as a superior base for reasoning adaptation, in addition to functions such as conversation and knowledge retrieval from general LLMs. It stands as the strongest fully open model, rivals open-weight leaders in its size class, outperforms Qwen2.5-72B and approaches the performance of Qwen3-235B. We actively infuse domain knowledge, reasoning, long-context, and tool use throughout the training process. This explicitly prepares the model for complex reasoning tasks. We demonstrate this potential using simple supervised fine-tuning, establishing a strong baseline that indicates significant headroom for advanced alignment. By releasing the full training history and data composition, we maximize the effectiveness of continuous training, a key open source production scenario. We release the model weights and signature LLM360 artifacts, such as complete training data, to empower the community with a capable, reasoning-centric foundation.

cs.LG↗

arXiv2018

Scalable Block-Wise Product BCH Codes

In this paper we comprehensively investigate block-wise product (BWP) BCH codes, wherein raw data is arranged in the form of block-wise matrix and each row and column BCH codes intersect on one data block. We first devise efficient BCH decoding algorithms, including reduced-1-bit decoding, extra-1-bit list decoding, and extra-2-bit list decoding. We next present a systematic construction of BWP-BCH codes upon given message and parity lengths that takes into account for performance, implementation and scalability, rather than focusing on a regularly defined BWP-BCH code. It can easily accommodate different message length or parity length at minimal changes. It employs extended BCH codes instead of BCH codes to reduce miscorrection rate and an inner RS code to lower error floor. We also describe a high-speed scalable encoder. We finally present a novel iterative decoding algorithm which is divided into three phases. The first phase iteratively applies reduced BCH correction capabilities to correct lightly corrupted rows/columns while suppressing miscorrection, until the process stalls. The second phase iteratively decodes up to the designed correction capabilities, until the process stalls. The last phase iteratively applies the proposed list decoding in a novel manner which effectively determines the correct candidate. The key idea is to use cross decoding upon each list candidate to pick the candidate which enables the maximum number of successful cross decoding. Our simulations show that the proposed algorithm provides a significant performance boost compared to the state-of-the-art algorithms.

cs.IT↗

arXiv2013

On Rational-Interpolation Based List-Decoding and List-Decoding Binary Goppa Codes

We derive the Wu list-decoding algorithm for Generalised Reed-Solomon (GRS) codes by using Gröbner bases over modules and the Euclidean algorithm (EA) as the initial algorithm instead of the Berlekamp-Massey algorithm (BMA). We present a novel method for constructing the interpolation polynomial fast. We give a new application of the Wu list decoder by decoding irreducible binary Goppa codes up to the binary Johnson radius. Finally, we point out a connection between the governing equations of the Wu algorithm and the Guruswami-Sudan algorithm (GSA), immediately leading to equality in the decoding range and a duality in the choice of parameters needed for decoding, both in the case of GRS codes and in the case of Goppa codes.

cs.IT↗

arXiv2008

New List Decoding Algorithms for Reed-Solomon and BCH Codes

In this paper we devise a rational curve fitting algorithm and apply it to the list decoding of Reed-Solomon and BCH codes. The proposed list decoding algorithms exhibit the following significant properties. 1 The algorithm corrects up to $n(1-\sqrt{1-D})$ errors for a (generalized) $(n, k, d=n-k+1)$ Reed-Solomon code, which matches the Johnson bound, where $D\eqdef \frac{d}{n}$ denotes the normalized minimum distance. In comparison with the Guruswami-Sudan algorithm, which exhibits the same list correction capability, the former requires multiplicity, which dictates the algorithmic complexity, $O(n(1-\sqrt{1-D}))$, whereas the latter requires multiplicity $O(n^2(1-D))$. With the up-to-date most efficient implementation, the former has complexity $O(n^{6}(1-\sqrt{1-D})^{7/2})$, whereas the latter has complexity $O(n^{10}(1-D)^4)$. 2. With the multiplicity set to one, the derivative list correction capability precisely sits in between the conventional hard-decision decoding and the optimal list decoding. Moreover, the number of candidate codewords is upper bounded by a constant for a fixed code rate and thus, the derivative algorithm exhibits quadratic complexity $O(n^2)$. 3. By utilizing the unique properties of the Berlekamp algorithm, the algorithm corrects up to $\frac{n}{2}(1-\sqrt{1-2D})$ errors for a narrow-sense $(n, k, d)$ binary BCH code, which matches the Johnson bound for binary codes. The algorithmic complexity is $O(n^{6}(1-\sqrt{1-2D})^7)$.

cs.IT↗

arXiv2008

On Expanded Cyclic Codes

The paper has a threefold purpose. The first purpose is to present an explicit description of expanded cyclic codes defined in $\GF(q^m)$. The proposed explicit construction of expanded generator matrix and expanded parity check matrix maintains the symbol-wise algebraic structure and thus keeps many important original characteristics. The second purpose of this paper is to identify a class of constant-weight cyclic codes. Specifically, we show that a well-known class of $q$-ary BCH codes excluding the all-zero codeword are constant-weight cyclic codes. Moreover, we show this class of codes achieve the Plotkin bound. The last purpose of the paper is to characterize expanded cyclic codes utilizing the proposed expanded generator matrix and parity check matrix. We characterize the properties of component codewords of a codeword and particularly identify the precise conditions under which a codeword can be represented by a subbasis. Our developments reveal an alternative while more general view on the subspace subcodes of Reed-Solomon codes. With the new insights, we present an improved lower bound on the minimum distance of an expanded cyclic code by exploiting the generalized concatenated structure. We also show that the fixed-rate binary expanded Reed-Solomon codes are asymptotically "bad", in the sense that the ratio of minimum distance over code length diminishes with code length going to infinity. It overturns the prevalent conjecture that they are "good" codes and deviates from the ensemble of generalized Reed-Solomon codes which asymptotically achieves the Gilbert-Varshamov bound.

cs.IT↗