Ahmed Heakl

About me

Selected Publications

Dr.LLM retrofits pretrained LLMs with lightweight per-layer routers that dynamically skip/execute/repeat layers under a compute budget, improving efficiency while preserving accuracy via supervised routing derived from MCTS.

CEPO is a token-level credit assignment method for RLVR that replaces uniform reward signals with a contrastive ratio between correct and wrong answer teachers (drawn from rejected rollouts at no extra cost), focusing gradient updates on the tokens that actually decide correctness. In just 50 training steps, CEPO improves over GRPO by +2.26% at 2B and +3.13% at 4B scale across five multimodal math reasoning benchmarks; distribution matching alternatives (OPSD, SDPO) collapse below the untrained baseline, confirming the information leakage CEPO is designed to avoid.

SVRPBench is an open-source benchmark of 500+ urban-scale stochastic VRP instances featuring time-dependent congestion, probabilistic delays, and heterogeneous time windows, together with a full evaluation suite for learning-based solvers.

Mobile-O is a compact, efficient unified vision–language–diffusion model that performs both multimodal understanding (VQA, OCR, reasoning) and image generation within a single architecture, while running entirely on-device. It is designed specifically for mobile and edge deployment, achieving real-time performance with a small memory footprint.

We introduce Guaranteed Guess (GG), a test-driven CISC-to-RISC transpiler using LLMs. GG achieves 99.4% functional accuracy on x86→ARMv8 translation and improves runtime and memory versus strong baselines through verification-centric generation and refinement.

KITAB-Bench contains 8,809 Arabic pages spanning 9 domains and 36 sub-domains to evaluate OCR and document understanding. The benchmark highlights large gains from VLM-based OCR (≈60% lower character error than traditional OCR) while PDF-to-Markdown conversion peaks around ≈65%.

We propose a curriculum framework and a benchmark with 4k+ reasoning steps across 8 categories for step-by-step visual reasoning. LlamaV-o1 improves over Llava-CoT by 3.8% while delivering ~5× faster inference.

SAHM is the first Arabic financial NLP benchmark, spanning seven tasks, AAOIFI standards QA, fatwa-based QA/MCQ, accounting and business exams, financial sentiment analysis, extractive summarization, and event-cause reasoning, with 14,380 expert-verified instances from authentic regulatory, juristic, and corporate sources. Evaluating 20 LLMs, we show that Arabic fluency does not imply financial reasoning, and that domain-adapted 7–8B models can surpass GPT-5 and match 72B baselines.

CASS introduces the first large-scale dataset and benchmark (70k samples) for transpilation between Nvidia CUDA and AMD HIP/RDNA3 at both source and assembly levels. We train domain-specialized LLMs (up to 7B) that achieve strong functional preservation and substantially improve transpilation accuracy on challenging low-level code.

VideoMolmo is a spatio-temporal grounding framework that combines LLM-driven pointing with mask fusion (via SAM2). We introduce temporal attention and bidirectional mask propagation, curate a large video-caption dataset with dense object points, and propose VPoS-Bench to evaluate grounding, counting, and reasoning in real-world video domains.

CAMEL-Bench is a large-scale Arabic-centric LMM evaluation benchmark spanning diverse domains (OCR/Docs, medical, remote sensing, video, and more) with broad coverage for realistic Arabic multimodal reasoning.

AIN is a bilingual (Arabic/English) large multimodal model designed to close the Arabic LMM gap, targeting broad-domain multimodal understanding and strong Arabic performance across diverse tasks.

Introduces the FoMER benchmark to evaluate step-by-step embodied reasoning for foundation multimodal models, analyzing action validity and reasoning correctness across diverse tasks and embodiments.

News

Academic Service and Awards