Institution profile

University of Sheffield

Academic institutioneurope · gb
Official website
Research library492linked papers
Opportunities0open roles
Selected work

Representative Papers

Demystifying Multi-Agent Debate: The Role of Confidence and Diversity

Jan 09, 2026arXiv.org

Traditional multi-agent debate (MAD) often fails to effectively enhance large language model performance due to homogeneous agents and uniform belief updating, sometimes even underperforming simple majority voting. This work proposes an improved framework that better mirrors human-like negotiation mechanisms: it employs diversity-aware initialization to increase the prior probability of correct hypotheses and introduces explicit confidence calibration in agent communication, coupled with a confidence-weighted belief update rule to systematically guide the debate toward the correct answer. Theoretical analysis and extensive experiments across six reasoning-based question-answering benchmarks demonstrate that the proposed approach significantly outperforms both conventional MAD and majority voting, thereby substantially improving the accuracy and reliability of multi-agent debates.

3 citationsRead paper

On Geometric Bipartite Graphs with Asymptotically Smallest Zarankiewicz Numbers

Oct 23, 2025

This paper investigates the Zarankiewicz problem for bipartite graphs of low Ferrers dimension—i.e., maximizing the number of edges while forbidding a $K_{k,k}$ subgraph. Focusing on Ferrers dimensions 3 and 4, we establish the first phase-transition phenomenon: edge bounds are linear in $n$ for dimension 3, whereas dimension 4 triggers a sharp complexity jump. Our approach integrates extremal graph theory, combinatorial geometry, and Ferrers-structure decomposition, augmented by a novel bichromatic diagonal argument. We derive tight asymptotic upper bounds: $2n(k-1)$ for chordal bipartite graphs and $54n(k-1)$ for grid intersection graphs—substantially improving prior exponential dependencies of the form $O(2^{O(k)}n)$. This work provides the first fine-grained threshold analysis of extremal behavior driven by Ferrers dimension and establishes optimal asymptotic orders for the Zarankiewicz problem on chordal bipartite and intersection graph classes.

2 citationsRead paper

Corpus-based approaches to Igbo diacritic restoration

Jan 31, 2019

This study addresses the challenge of diacritic restoration in Igbo, a low-resource language severely hindered by the scarcity of annotated data, which leads to significant ambiguity in diacritic placement. To tackle this problem, the work proposes the first systematic framework that integrates both traditional and modern approaches: n-gram–based contextual prediction, a sliding-window classification model, and a vector-matching strategy leveraging word embedding similarity. Evaluated under low-resource conditions, the framework effectively recovers missing diacritics, substantially improving natural language processing performance for Igbo. Moreover, it offers a reusable, multi-paradigm technical pathway that can be adapted to other under-resourced languages facing similar challenges in orthographic normalization.

2 citationsRead paper

Empirical Risk Minimization with $f$-Divergence Regularization

Jan 19, 2026

This work investigates the incorporation of $f$-divergence regularization into empirical risk minimization to enhance generalization in expected risk. By establishing equivalence conditions between $f$-divergence-regularized empirical risk minimization and expected risk minimization under an $f$-divergence constraint, the study introduces the notion of a “normalizing function,” which is characterized as a nonlinear ordinary differential equation (ODE). This characterization reveals structural equivalences across different $f$-divergence regularizations. Leveraging duality theory, ODE analysis, and numerical approximation techniques, the authors develop a unified computational framework applicable to a broad class of $f$-divergences. Numerical experiments demonstrate the practical impact of various $f$-functions on training and test risks, thereby extending the range of tractable divergences and strengthening the theoretical and algorithmic coherence of the approach.

1 citationsRead paper

SemEval-2025 Task 1: AdMIRe -- Advancing Multimodal Idiomaticity Representation

Mar 19, 2025

This work addresses the core challenge of idiom comprehension—particularly non-literal figurative meaning—in multimodal, multilingual settings. To this end, we introduce the first cross-lingual, cross-modal idiom understanding benchmark, comprising two tasks: image–idiom alignment ranking and sequential image prediction. We propose a systematic evaluation framework featuring novel multi-query fusion and a Mixture-of-Experts (MoE) mechanism to mitigate semantic biases in large language models regarding idiomatic expressions. Our approach integrates vision-language models (VLMs) and large language models (LLMs) via multi-query reasoning, expert-weighted ensemble, and semantic smoothing. Experiments demonstrate human-level performance on both tasks, with significant improvements in multilingual idiom–image semantic alignment accuracy and sequential consistency. This work establishes a new paradigm for idiom representation learning in multimodal, multilingual contexts.

1 citationsRead paper
Recent publications

Latest Papers