Sitemap
A list of all the posts and pages found on the site. For you robots out there, there is an XML version available for digesting as well.
Pages
Posts
Research Templates, Part 1: Papers and Project Pages
Published:
Finishing a paper is always a relief, but I’ve started to dread the work that comes after it. Every time, I end up doing things a little differently: asking an AI to draft a poster, fixing it by hand, and figuring out how to put together a project page.
publications
FCDS: Fusing Constituency and Dependency Syntax into Document-Level Relation Extraction
Published in LREC-COLING 2024, 2024
We introduce FCDS, a document-level relation extraction model that fuses constituency and dependency syntax. By combining sentence-level aggregation from constituency trees with dependency-based graph reasoning, FCDS better captures cross-sentence relations between entities. Experiments across multiple domains show significant performance gains, highlighting the effectiveness of integrating both syntactic views.
From Emergence to Control: Probing and Modulating Self-Reflection in Language Models
Published in TMLR 2026, 2025
We study the emergence and control of self-reflection in large language models. Our probing method reveals that pretrained models already contain a latent capacity for reflection, which can be amplified without additional training. By identifying and manipulating a “self-reflection vector” in activation space, we achieve bidirectional control over reflective behavior, improving reasoning accuracy or reducing computation as needed. This work deepens understanding of self-reflection and demonstrates how model internals can enable precise behavioral modulation.
Alleviating subgraph-induced oversmoothing in link prediction via coarse graining
Published in Neurocomputing, 2025, 2025
We address the oversmoothing problem in link prediction caused by repetitive high-degree nodes across subgraphs. Our method introduces a coarse-graining strategy that merges strongly correlated nodes, yielding more diverse receptive fields and reducing subgraph size. This not only mitigates oversmoothing but also improves scalability and efficiency of GNN-based link prediction.
AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features
Published in ICLR 2026, 2026
Sparse autoencoders (SAEs) are widely used for LLM interpretability, but existing variants often impose non-negativity that prevents single features from representing bidirectional concepts. We derive SAE variants from unrolled proximal gradient updates, identify this structural limitation, and propose AbsTopK SAE with magnitude-based hard thresholding. Across four LLMs and seven probing/steering tasks, AbsTopK improves reconstruction and interpretability while enabling single features to encode contrasting concepts.
One Token Can Be Enough: Bridging Prompting and Activation Steering with Prefix Steering
Published in arXiv preprint, 2026
Bridging prompting and activation steering, short initial interventions often retain much of full steering’s behavioral control while better preserving general capabilities.