arXiv subjects

Yihong Wu

Publications and source records attributed to Yihong Wu.

At least 19 recordsLinked to original sources

arXiv2026

FINSKILLOPS: A Self-Evolving Multi-Agent System for SEC Filing QA

Financial QA systems are typically improved before deployment through better retrieval, prompting, or agent coordination, leaving their reliability behavior fixed thereafter. In practice, new SEC-filing questions repeatedly expose heterogeneous errors in period, entity, evidence use, and calculation. Existing self-improvement methods can turn failures into new behaviors, but offer limited control over where a correction should apply or which previously correct answers it may break. We therefore frame post-deployment improvement as controlled behavioral maintenance: recurring failures should become scoped skill patches, and each patch should earn deployment with- out introducing regressions. We instantiate this view in FINSKILLOPS, a multi-agent system for SEC filing QA. FINSKILLOPS derives reusable skills from evidence-grounded, typed failure diagnoses and governs them through targeted validation, protected-case regression checks, negative controls, and versioned replacement or retirement. Across six financial QA benchmarks, a single frozen skill registry achieves the highest verdict-weighted correctness and reference consistency among the evaluated systems. Evolved skills raise correctness from 3.70 to 4.55 on our enhanced benchmark. In a separate 12-round operational study, only six of 33 proposed skills are promoted, while the monitoring non-correct rate falls from 20.0% to 12.5%. These results establish controlled skill scope, admission, and lifecycle management as the foundation for reliable self-improvement.

cs.AI↗

arXiv2026

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and converts that evidence into the next round of learning. We present NeoHorse-1, a family of agent-native models developed to explore this path through agentic post-training. Our system combines a heterogeneous model pool with intelligent routing, recording the predicted capability demand, selected service tier, and subsequent interaction for each user turn. These records are converted into training examples that preserve interleaved reasoning, tool calls, and harness context, and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression. Capability-guided allocation then converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B, substantially narrowing the aggregate gap between the post-trained 4B model and the 9B base model. NeoHorse-1 provides an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations.

cs.CL↗

arXiv2026

Entropy of Bernoulli Measures Conditioned on Affine Subspaces and a Problem of Ancheta--Massey

A textbook result in information theory is that linear encoders achieve the entropy for lossless compression of Bernoulli source with parameter $p$. For lossy compression, however, linearity is known to incur strict suboptimality compared to the rate-distortion function. Massey asked whether the optimal rate for linear encoding is achieved simply by compressing a fraction of the bits linearly and losslessly and estimating the rest by zero \cite{Massey1978}. For $p=\frac12$, Ancheta answered this question affirmatively \cite{Ancheta1978}. This note extends Ancheta's result to all $p<\frac12$. The key argument is to bound the entropy of the posterior distribution conditioned on an affine subspace in terms of its marginals. The proof was discovered by GPT-5.6 Sol in an interactive process guided by the author. The purpose of the present note is to communicate a simplified version of this proof and to make connections with the existing literature on coding theory and spin glass theory.

cs.IT↗

arXiv2026

Empirical Bayes linear regression in high dimensions: Method of moments and sub-linear sample complexity

We study empirical Bayes estimation of the prior in high-dimensional linear regression $\mathbf{y}=\mathbf{X}\mathbfβ+\mathbf{\varepsilon}$, where the regression coefficients are drawn independently from an unknown sub-Gaussian prior. In contrast to the sequence model, the design matrix couples the latent coefficients, so that recovering the prior requires deconvolving it from both the noise and copies of itself. We introduce the \emph{Empirical Bayes Method of Moments} (EBMoM), a computationally efficient procedure for general designs that recursively estimates the prior moments through a lower-triangular system of estimating equations and runs in time $O(np^2)$. Under mild design conditions, satisfied in particular by a broad class of correlated random designs, we show that EBMoM consistently estimates a growing number of moments and hence the prior itself, provided that $n\geq p^{1-o(1)}$. A matching information-theoretic lower bound, valid for a broad class of designs, shows that this sub-linear sample complexity is optimal for nonparametric prior estimation. This improves on existing results for likelihood-based methods whose consistency requires a linear sample size $n=Ω(p)$.

math.ST↗

arXiv2026

Diagnostic-Driven Layer-Wise Compensation for Post-Training Quantization of Encoder-Decoder ASR Models

Layer-wise post-training quantization reconstructs each layer from inputs already altered by the quantized prefix. QEP compensates for this drift with one model-wide coefficient, conflating the model-level operating point with residual variation across layers. We present FADE, which constructs layer-specific coefficients from normalized round-to-nearest distortion and a heuristic calibrated-solver response. It requires no training or per-model coefficient search and adds no inference-time operation. We evaluate seven Whisper, Moonshine, and Qwen3-ASR models at 3 and 4 bits on four English ASR benchmarks. Across 38 settings, FADE lowers mean word error rate relative to fixed QEP-0.5 in 31, although a development-tuned global coefficient recovers much of this gap. The largest absolute reductions occur in 3-bit settings whose final error remains too high for practical use; these results measure collapse mitigation rather than deployable accuracy. In two lower-WER 3-bit cases, FADE reduces the tuned-global result from 3.67 to 3.10 and from 13.03 to 11.63. Most 4-bit differences from the tuned control are within a descriptive tolerance, and one reverses. Paired reruns and within-seed permutations on an outcome-informed subset support assignment sensitivity in selected cases, but do not estimate a matrix-wide success rate. FADE is therefore a layer-wise alternative when per-model coefficient search is unavailable, not a universal replacement for tuned global compensation.

cs.SD↗

arXiv2026

Sharp regret-Hellinger bounds for Gaussian empirical Bayes via polynomial approximation

A central problem in the theory of empirical Bayes is to control the regret (excess risk) of a learned Bayes rule by the Hellinger distance between the estimated and true marginal densities. In the normal means model, the classical result of Jiang and Zhang (2009) achieves this only after regularizing the Bayes rule and incurs an extraneous cubic logarithmic factor through a delicate recursive argument. This paper introduces a new technique, based on polynomial approximation and Bernstein-type inequalities for weighted $L_2$ norms, that bounds the unregularized regret directly. The method is conceptually simpler and yields sharper, sometimes optimal, regret bounds. For compactly supported priors, we prove the sharp bound that the regret is $O(ε^2 \frac{\log(1/ε)}{\log\log(1/ε)})$, where $ε$ is the Hellinger distance between the marginal densities. The same method also extends to priors with exponential tails. Conversely, we show that regularization is genuinely necessary for heavy-tailed priors under only bounded moment assumptions. As statistical consequences, we obtain improved regret bounds for the nonparametric maximum likelihood estimator (NPMLE). Notably, for compactly supported priors, by determining the optimal Hellinger rate of mixture density estimation, we show that the optimal regret for sample size $n$ scales as $Θ(\frac{1}{n}(\frac{\log n}{\log\log n})^2)$, attained by the NPMLE within a $\log\log n$ factor.

math.ST↗

arXiv2026

Gradient flows for empirical Bayes in high-dimensional linear models

Empirical Bayes provides a powerful framework for learning and adapting to latent structure in data. In sequence models where an independent observation is associated to each latent parameter, theory and methods around empirical Bayes are well-developed. However, in models where latent parameters and observed data interact through more complex designs, many statistical and algorithmic questions remain unanswered. In this work, we study a canonical setting of empirical Bayes estimation for the distribution of regression coefficients in a high-dimensional Bayesian or random effects linear model. Computationally, we propose a new system of gradient flow equations for computing a nonparametric maximum likelihood estimator (NPMLE), which jointly optimizes over the prior and posterior distributions of the regression coefficients in a Gibbs variational representation of the marginal log-likelihood. A diffusion-based implementation yields an adaptive Langevin dynamics algorithm in which the prior evolves continuously to optimize a sequence model log-likelihood defined by the coordinates of the Langevin sample. Theoretically, we show polynomial-time convergence of the proposed gradient flow to a near-NPMLE from any initialization within a convex sub-level set of the marginal log-likelihood, by developing a high-temperature log-Sobolev inequality for the posterior law. We establish the statistical consistency of any near-NPMLE under deterministic conditions for the regression design as $n,p\rightarrow\infty$.

math.ST↗

arXiv2026

FinSAgent: Corpus-Aligned Multi-Agent RAG Framework for Evidence-Grounded SEC Filing Question Answering

Financial question answering over U.S. Securities and Exchange Commission (SEC) filings requires retrieving and synthesizing heterogeneous evidence dispersed across long, standardized, and highly redundant disclosures. Existing retrieval-augmented and multi-agent systems typically derive retrieval queries directly from the user's question and rank candidates by semantic similarity. Together, these choices create prior-corpus misalignment: a mismatch between model priors and the target filings' structure, terminology, and evidence standards. As a result, query generation misses corpus-specific evidence, while semantic reranking favors topically similar but evidentially invalid false-positive chunks. We propose FinSAgent, an evidence-grounded multi-agent framework that reframes SEC filing QA as corpus-aligned retrieval planning and corrects both ends with a single principle: inject corpus-side conditioning wherever model priors would otherwise dominate. FinSAgent combines (1) role-specialized agents anchored to the mandated 10-K item structure, (2) database-aware query decomposition that conditions each agent's sub-queries on a lightweight, summary-level view of the local corpus, and (3) multi-path retrieval with a learned feature-gated reranker that separates evidential validity from semantic similarity. Across five offline financial QA benchmarks, FinSAgent improves retrieval coverage and answer correctness over strong single-agent and multi-agent baselines; in a three-arm randomized online experiment with 1,000 anonymous user ratings, it also receives higher scores than baselines.

cs.IR↗

arXiv2026

From Evidence to Trajectory: Abductive Reasoning Path Synthesis for Retrieval-Augmented Generation Agents Development

Retrieval-augmented generation (RAG) agent development is hindered by the lack of executable ground-truth agent-environment interaction trajectories. Existing datasets provide questions, answers, and evidence, but lack fine-grained supervision for retriever invocation, dynamic planning, and stepwise decision-making. Reinforcement learning offers a potential solution, but often suffers from sparse rewards and cold-start failures when base large language models (LLMs) lack sufficient reasoning capability. Meanwhile, existing data synthesis methods mainly generate post-hoc rationales rather than executable environment-interaction trajectories. In this paper, we propose EviPath, an evidence-anchored reasoning path synthesis paradigm for RAG agent development. EviPath reverse-engineers executable trajectories from question-answer pairs and supporting evidence through three stages: (i) Abductive Subtask Planning, which decomposes questions and plans dependency-aware solution paths; (ii) Faithful Sub-question Answering, which uses supporting evidence as a proxy environment to generate grounded intermediate thoughts and answers; and (iii) Conversational Fine-Tuning, which converts complete trajectories into a dialogue format for supervised fine-tuning. Experiments on widely used question-answering benchmarks show that an 8B model trained on our synthetic corpus significantly and consistently outperforms state-of-the-art baselines, achieving a 14.7% absolute Exact Match gain in open-domain question answering.

cs.CL↗

arXiv2026

SpaR3D-MoE: Adaptive 3D Spatial Reasoning from Sparse Views Meets Geometry-Inductive Mixture-of-Experts

Recent Multimodal Large Language Models (MLLMs) struggle to bridge the representational gap between 2D semantic understanding and 3D spatial geometry. Existing 3D-aware models either rely on costly 3D-specific data or utilize RGB-only inputs with heuristic sampling and monolithic, shallow fusion, which respectively disrupt essential spatiotemporal connectivity and induce modality contention across diverse spatial tasks. To overcome these bottlenecks, we introduce SpaR3D-MoE, an end-to-end framework that enables adaptive spatial reasoning by equipping MLLMs with geometry-aware capabilities from only sparse RGB inputs. First, we propose an adaptive spatiotemporal manifold sampling mechanism that constructs a geometry-aware spatiotemporal graph to extract informative keyframes, effectively mitigating sequence redundancy while preserving the scene's topological connectivity. Second, we introduce the heterogeneous geometry-inductive Mixture-of-Experts driven by an instruction-pose aware router, which adaptively routes multimodal tokens to specialized experts, resolving the cross-modal contention inherent in monolithic fusion. Extensive experiments on VSI-Bench, ScanQA, and SQA3D demonstrate that our method achieves state-of-the-art performance. Notably, SpaR3D-MoE achieves the highest average score of 63.5 on VSI-Bench, outperforming the strongest baseline by 7.8 absolute points, alongside relative improvements of 35.4% and 51.4% in Route Plan and Relative Direction tasks, respectively.

cs.CV↗

arXiv2026

The broken sample problem revisited: Proof of a conjecture by Bai-Hsing and high-dimensional extensions

We revisit the classical broken sample problem: Two samples of i.i.d.\ data points ${\mathbf{X}}=\{X_{1},\ldots , X_{n}\}$ and ${\mathbf{Y}}=\{Y_{1},\ldots ,Y_{m}\}$ are observed without correspondence with $m\leq n$. Under the null hypothesis, ${\mathbf{X}}$ and ${\mathbf{Y}}$ are independent. Under the alternative hypothesis, ${\mathbf{Y}}$ is correlated with a random subsample of ${\mathbf{X}}$, in the sense that $(X_{π(i)},Y_{i})$'s are drawn independently from some bivariate distribution for some latent injection $π:[m] \to [n]$. Originally introduced by DeGroot, Feder, and Goel to model matching records in census data, this problem has recently gained renewed interest due to its applications in data de-anonymization, data integration, and target tracking. Despite extensive research over the past decades, determining the precise detection threshold has remained an open problem even for equal sample sizes ($m=n$). Assuming $m$ and $n$ grow proportionally, we show that the sharp threshold is given by a spectral and an $L_{2}$ condition of the likelihood ratio operator, resolving a conjecture of Bai and Hsing in the positive. These results are extended to high dimensions and settle the sharp detection thresholds for Gaussian and Bernoulli models.

math.ST↗

arXiv2026

Resolution of the Detection Threshold Conjecture for Random Geometric Graphs in the $d>n$ Regime

A random geometric graph (RGG) is generated by first sampling latent points $x_1,\ldots,x_n$ independently and uniformly from the unit sphere in $\mathbb{R}^d$, and then connecting each pair $(i,j)$ if $\langle x_i,x_j\rangle$ exceeds some threshold $τ$. We study the sharp detection threshold -- the largest dimension at which the RGG can be statistically distinguished from the Erdős--Rényi graph with the same edge density $p$. This threshold is conjectured to be $d \asymp (nh(p))^3$, where $h(p)=p \log \frac{1}{p} + (1-p) \log \frac{1}{1-p}$ is the binary entropy function. Previous works proved this conjecture for dense graphs with constant $p$ and, up to polylogarithmic factors, very sparse graphs with $p=Θ(1/n)$. In this paper, we prove that detection is impossible when $d\gg (nh(p))^3$ and $d\ge (1+ε) n$ for any constant $ε>0$, thereby resolving the conjecture in the regime $p\gtrsim n^{-2/3}/\log n$ and improving upon the state of the art in the regime $1/n \ll p \ll n^{-2/3}/\log n$. The key to our proof is a sharp analysis of the posterior distribution of the latent points given the observed graph, obtained through an information-theoretic comparison argument combined with strong log-concavity.

math.PR↗

arXiv2026

Rethinking Groups in Critic-Free RLVR

Reinforcement learning (RL) has become a central paradigm for post-training large language models. Existing critic-free RL methods typically generate a group of rollouts for the same question to estimate value baselines for advantage computation. However, this design suffers from data inefficiency, group synchronization barriers, and inflexibility with structured rollouts. In this work, we revisit the role of the ``group'' and show that its underlying function is not merely to estimate baselines but to prevent false penalties on negative samples. Building on this insight, we propose negative token filtering, a simple and effective strategy that enables stable single-rollout training. We apply it to two batch-level advantage methods, achieving comparable performance on reasoning tasks and stronger performance on agentic tasks relative to group-based RL techniques.

cs.LG↗

arXiv2026

Search-on-Graph: Iterative Informed Navigation for Large Language Model Reasoning on Knowledge Graphs

Large language models (LLMs) augmented with knowledge graphs (KGs) offer a promising approach for knowledge-intensive reasoning. Central to this approach is the selection of appropriate reasoning paths in the KG. Yet, existing methods face a common limitation: reasoning path selection is often performed by separate modules using criteria that are only weakly connected to the reasoning requirements. This often results in selecting incorrect relations or premature pruning of relevant paths. We propose Search-on-Graph (SoG), a method that strengthens the connection between path selection and reasoning by having the LLM itself select which relations to follow, informed by both the available KG structure and the complete reasoning history. SoG follows an \textit{observe-think-navigate} paradigm: at each step, the LLM observes the relational connections available at the current entity, reasons about which path best advances toward answering the question, and navigates accordingly. This context-aware navigation fully exploits the LLM's reasoning capabilities rather than relying on independent selection modules with surrogate criteria. Experiments on six knowledge graph question answering (KGQA) benchmarks demonstrate that SoG outperforms state-of-the-art methods while requiring no task-specific fine-tuning and generalizing across different KG schemas.

cs.CL↗

arXiv2026

An Entity Linking Agent for Question Answering

Some Question Answering (QA) systems rely on knowledge bases (KBs) to provide accurate answers. Entity Linking (EL) plays a critical role in linking natural language mentions to KB entries. However, most existing EL methods are designed for long contexts and do not perform well on short, ambiguous user questions in QA tasks. We propose an entity linking agent for QA, based on a Large Language Model that simulates human cognitive workflows. The agent actively identifies entity mentions, retrieves candidate entities, and makes decision. To verify the effectiveness of our agent, we conduct two experiments: tool-based entity linking and QA task evaluation. The results confirm the robustness and effectiveness of our agent.

cs.CL↗

arXiv2026

It Takes Two: Your GRPO Is Secretly DPO

GRPO has emerged as a prominent reinforcement learning algorithm for post-training LLMs. Unlike critic-based methods, GRPO computes advantages by estimating the \emph{value baselines} from group-level statistics, eliminating the need for a critic network. Consequently, the prevailing view emphasizes the necessity of large group sizes, which are assumed to yield more accurate statistical estimates. In this paper, we propose a different view that the efficacy of GRPO stems from its implicit contrastive objective in the optimization, which helps reduce variance via the control variate method. This makes GRPO structurally related to preference learning methods such as DPO. This perspective motivates 2-GRPO, a minimal group-size variant that constructs contrastive signals with only two rollouts. We provide a rigorous theoretical analysis of 2-GRPO and empirically validate its effectiveness: 2-GRPO retains $97.6\%$ of the performance of 16-GRPO, while requiring only $12.5\%$ of the rollouts and $21\%$ of the training time.

cs.LG↗

arXiv2026

Confidence-Aware Alignment Makes Reasoning LLMs More Reliable

Large reasoning models often reach correct answers through flawed intermediate steps, creating a gap between final accuracy and reasoning reliability. Existing alignment strategies address this with external verifiers or massive sampling, limiting scalability. In this work, we introduce CASPO (Confidence-Aware Step-wise Preference Optimization), a framework that aligns token-level confidence with step-wise logical correctness through iterative Direct Preference Optimization, without training a separate reward model. During inference, we propose Confidence-aware Thought (CaT), which leverages this calibrated confidence to dynamically prune uncertain reasoning branches with negligible O(V) latency. Experiments across ten benchmarks and multiple model families show that CASPO consistently improves reasoning reliability and inference efficiency. CASPO scales to Qwen3-8B-Base and surpasses tree-search baselines on AIME'24 and AIME'25 without using reward-model data. We also release a step-wise dataset with confidence annotations to support fine-grained analysis of reasoning reliability. Code is available at https://github.com/Thecommonirin/CASPO.

cs.AI↗

arXiv2026

Advancing Multi-Agent RAG Systems with Minimalist Reinforcement Learning

Large Language Models (LLMs) equipped with modern Retrieval-Augmented Generation (RAG) systems often employ multi-turn interaction pipelines to interface with search engines for complex reasoning tasks. However, such multi-turn interactions inevitably produce long intermediate contexts, as context length grows exponentially with exploration depth. This leads to a well-known limitation of LLMs: their difficulty in effectively leveraging information from long contexts. This problem is further amplified in RAG systems that depend on in-context learning, where few-shot demonstrations must also be included in the prompt, compounding the context-length bottleneck. To address these challenges, we propose Mujica-MyGo, a unified framework for efficient multi-turn reasoning in RAG. Inspired by the divide-and-conquer principle, we introduce Mujica (Multi-hop Joint Intelligence for Complex Question Answering), a multi-agent RAG workflow that decomposes multi-turn interactions into cooperative sub-interactions, thereby mitigating long-context issues. To eliminate the dependency on in-context learning, we further develop MyGO (Minimalist Policy Gradient Optimization), a lightweight and efficient reinforcement learning algorithm that enables effective post-training of LLMs within complex RAG pipelines. We provide theoretical guarantees for MyGO's convergence to the optimal policy. Empirical evaluations across diverse question-answering benchmarks, covering both text corpora and knowledge graphs, show that Mujica-MyGO achieves superior performance.

cs.CL↗