Abdulhady Abas Abdullah — AI Researcher

Building trustworthy language models — grounded, interpretable, multilingual.

AI Researcher (M.Sc.) working on large language models, mechanistic interpretability, and the first speech & NLP systems for Kurdish. 35+ publications, a regular paper at ICML 2026, and submissions to Nature Communications, AAAI, and ICLR.

Plate I

Fig. 01 · Topology- & uncertainty-aware preference optimization

ICML 2026 · Regular Paper Accepted

TUR-DPO — Topology- & Uncertainty-Aware Direct Preference Optimization

Extends DPO to reward how an answer is derived, not only what it says. Each response is scored through a reasoning-topology graph with a calibrated uncertainty signal that shapes an uncertainty-weighted preference loss.

GSM8K+4.1%mathematical reasoning

MATH+2.6%competition math

BIG-Bench Hard+2.8%base · LLaMA-2 7B

PyTorchDPOAlignmentGraph ReasoningUncertainty

View repository

Plate II

TANSAR logic-constrained associative memory diagram

Fig. 02 · Logic-constrained associative memory as a DAG

AAAI 2027 Under Review

TANSAR — Topology-Aware Neuro-Symbolic Associative Retrieval

Edits knowledge through an explicit correlation-matrix memory organised as a logic-constrained DAG rather than by modifying weights: graph diffusion spreads the query, first-order-logic pruning drops inconsistent paths, and path-utility scoring selects the best valid reasoning chain — enabling transparent multi-hop reasoning and gradient-free updates.

1-hop accuracy> 99.8%

Multi-hop · HotpotQA78.5vs 68.2 baseline

Edit speed1.8 ms≈ 10⁴× faster

Logical contradictions0.0%

Neuro-SymbolicKnowledge EditingAssociative MemoryFirst-Order LogicMulti-Hop QA

View repository

Plate III

Shared safety representations across languages and modalities

Fig. 03 · A shared safety subspace across languages & modalities

EACL 2027 Under Review

Shared Safety Representations Across Languages & Modalities

A three-phase study: multimodal safety alignment with TUR-DPO, sparse-autoencoder feature discovery and cross-modal patching to isolate a shared safety subspace, then causal intervention that transfers refusal behaviour zero-shot across English, Kurdish, Arabic, text, and speech.

Kurdish jailbreak rate78.5% → 4.2%after intervention

Probe accuracy92–96%TUR-DPO layers 16–22

Transferzero-shotEN · KU · AR — text & speech

Sparse AutoencodersActivation PatchingSVCCAMultilingualSafety

View repository

Plate IV

Fig. 04 · Zero-init adapters with meta-reasoning prompting

Applied Soft Computing · Under Review

LLaMA-Adapter+ MRP

Zero-initialised adapters injected into a frozen LLaMA-7B, combined with meta-reasoning prompting: a two-stage selector picks the reasoning strategy per task — Chain-of-Thought, Least-to-Most, Program-Aided, or Self-Refine — with CLIP-ViT image embeddings prepended for vision-language tasks.

GSM8K96.1%

MMLU91.2%

ScienceQA89.7%

Zero-Init AdapterPEFTMulti-ModalCLIP-ViTMeta-Reasoning

View repository

Plate V

Fig. 05 · Uncertainty-gated mixture of prompts

TMLR 2025 Published

UMP-Net — Uncertainty-Aware Mixture of Prompts

An efficient instruction-tuning framework that samples latent noise, adaptively mixes a pool of textual, visual, and multimodal learnable prompts, and routes them through a conformal-prediction uncertainty gate — with heterogeneous KNN clustering for robust cross-modal generalisation under noise.

Prompt pooltextual · visual · multimodal

Routerconformal

VenueTMLR 2025

PEFTConformal PredictionInstruction TuningCross-Modal

View repository

Building trustworthy language models — grounded, interpretable, multilingual.

Selected works.

TUR-DPO — Topology- & Uncertainty-Aware Direct Preference Optimization

TANSAR — Topology-Aware Neuro-Symbolic Associative Retrieval

Shared Safety Representations Across Languages & Modalities

LLaMA-Adapter+ MRP

UMP-Net — Uncertainty-Aware Mixture of Prompts

A working bibliography.

Practice & formation.

Chief Technology Officer

Research Assistant in AI

Instructor

Graduate Researcher · NLP & Speech

Pedagogical Training · 30 ECTS

M.Sc. Artificial Intelligence

B.Sc. Computer Science

First Place · HITEX AI Competition

First Place · Rwanga Award

Instruments & craft.

i. Modeling & Learning

ii. Frameworks

iii. Systems & Cloud

iv. Languages & Tools

Let’s make something worth keeping.