Best Paper Award · AACL-IJCNLP 2026
LLM-Ideoplasticity: a model's political ideology is a shape over contexts, not a single point See the paper ↓

About

I am Adib Sakhawat, an NLP researcher from Bangladesh. I study what large language models can and cannot do in the places mainstream benchmarks skip: low-resource languages such as Bangla, figurative and legal text, and the political and safety behaviour of models once the language or the framing changes. I try to build datasets and evaluation protocols that could falsify a claim about model capability, not just confirm it.

Research

01

Multilingual & low-resource NLP

Bangla resources and benchmarks: a 10,361-idiom corpus, a corpus of legal acts from 1799 to 2025, and why models read idioms literally and retrieve unevenly across languages.

02

LLM evaluation

Evaluation built to test assumptions: what benchmarks actually measure, consistency in temporal reasoning, and game-based protocols for multi-turn behaviour.

03

Safety in low-resource languages

The gap between a model's ability to understand harmful content and its ability to contain it, in Bangla derogatory speech and adversarial dialogue.

04

LLMs & social science

Measuring the political behaviour of language models with instruments from political science and psychometrics, treating ideology as a context-conditioned distribution.

News

  • Best Paper Award at AACL-IJCNLP 2026 for LLM-Ideoplasticity.
  • Joined the Program Committee of BLP 2027, the Workshop on Bangla Language Processing at NAACL 2027.
  • Graduated from IUT, First Class with Honours.
  • Presented AREG in person at the MusIML Workshop @ ICML 2026 in Seoul, and volunteered at the workshop.
  • Two papers accepted at EMNLP 2026 workshops: AIDG (REALM) and Two Emojis of Difference (MRL).
  • Bangla Bagdhara idiom benchmark published at LREC 2026.

Publications

Peer-reviewed

  • LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs as a Context-Conditioned Distribution
    Adib Sakhawat, Syed Rifat Raiyan, Tahsin Islam, Takia Farhin, Hasan Mahmud, Md Kamrul Hasan
    Framing and language move a model's political position substantially, yet the cohort as a whole sits in a narrow band: a single point cannot summarise LLM political behaviour.
    AACL-IJCNLP 2026 Best Paper Award Project page ACL Anthology PDF Code & data
  • When Words Don't Mean What They Say: Figurative Understanding in Bengali Idioms
    Adib Sakhawat, Shamim Ara Parveen, Md Ruhul Amin, Tahera Khatun, Shamim Al Mahmud, Md Saiful Islam
    10,361 idioms under a 19-field expert schema. Of 30 LLMs, none exceeds 50% accuracy at inferring figurative meaning; humans reach 83.4%.
    LREC 2026 Project page ACL Anthology PDF Dataset
  • AIDG: A Formal Decomposition of Information Extraction and Containment Asymmetries in Multi-Turn LLM Dialogue
    Adib Sakhawat, Fardeen Sadab, Rakin Shahriar
    REALM @ EMNLP 2026 Project page ACL Anthology PDF Code
  • Two Emojis of Difference: What Multilingual Affective Generation Benchmarks Actually Measure
    Fardeen Sadab, Adib Sakhawat
    MRL @ EMNLP 2026 ACL Anthology PDF
  • AREG: Adversarial Resource Extraction Game for Evaluating Persuasion and Resistance in Large Language Models
    Adib Sakhawat, Fardeen Sadab, Tamjid Hasan Fahim
    MusIML @ ICML 2026 Project page OpenReview Code

Preprints

  • Literal Lock-In: How Test Language Shifts Literal Attraction in Typed Decision Models on Bengali Idioms
    Adib Sakhawat, Syed Rifat Raiyan
    Preprint Project page arXiv Code & data
  • Evaluating Temporal Reasoning in Large Language Models: Event Temporal Relation Extraction with Allen's Interval Algebra, Logical Consistency, and Relative Time in Narratives
    Adib Sakhawat, Fardeen Sadab, Tamjid Hasan Fahim, and 9 others
    Preprint arXiv Code
  • Aligned in Form, Not in Meaning: The Comprehension–Containment Decoupling of LLM Safety in Low-Resource Bangla Derogatory Speech
    Shadab Bin Habib, A K M Ferdous Reza Habib, Subarno Neel, Adib Sakhawat
    Preprint arXiv
  • BLAD: A Historically Contextualized, Multilingual Dataset of Bangladeshi Legal Acts (1799 to 2025)
    Adib Sakhawat
    Preprint arXiv Dataset Code
  • Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models
    Adib Sakhawat, Fardeen Sadab, Atik Shahriar
    Preprint arXiv Code
  • Political Alignment in Large Language Models: A Multidimensional Audit of Psychometric Identity and Behavioral Bias
    Adib Sakhawat, Tahsin Islam, Takia Farhin, Syed Rifat Raiyan, Hasan Mahmud, Md Kamrul Hasan
    Preprint arXiv Code

Leadership & volunteering

IUT Computer SocietyUniversity computing society

  • Vice President2025 – 2026
  • Assistant Director of Technical Affairs2024 – 2025 · Technical Lead, CodeRush 2.0
  • Machine Learning Team Member2023

IUT Al-Fazari Interstellar SocietyUniversity astronomy club

  • Vice President2025 – 2026
  • Joint Secretary2024 – 2025
  • Administrative Member2023

IEEE Computer Society Chapter, IUTStudent chapter

  • Vice Chair2025 – 2026

Bangladesh Red Crescent SocietyCOVID-19 response

  • Volunteer2019 – 2021 · trained in first aid and outbreak management; volunteered at sites across the pandemic

Datasets

  • Bangla Bagdhara — 10,361 Bengali idioms, 19-field annotation schema, benchmark and human-evaluation protocol.
  • Bangladesh Legal Acts Dataset — 1,484 acts (1799–2025), 35,633 sections, bilingual, annotated with the legal system in force.

CC BY 4.0 · 5,900+ downloads on Kaggle

Selected engineering

  • Reko — Rust CLI: every function in a repo, 15 languages, one IR, byte-exact reconstruction, semantic search.
  • ISHKAPON AI — local-first tutor; every number is computed in a sandboxed Python runtime, not recalled by the model.
  • Pi Daytona Chat — coding agent with an isolated sandbox per conversation and every tool call logged.