Papers by Heshaam Faili

9 papers
Matina: A Large-Scale 73B Token Persian Text Corpus (2025.naacl-long)

Copied to clipboard

Challenge: Existing Persian datasets are small and lack content diversity . lack of high-quality data has slowed development of NLP models and open-source LLMs for Persian.
Approach: They propose a Persian dataset of 72.9B tokens that is preprocessed and deduplicated to ensure high data quality.
Outcome: The proposed model performs well on key Persian NLP tasks.
PerCQA: Persian Community Question Answering Dataset (2022.lrec-1)

Copied to clipboard

Challenge: Community Question Answering (CQA) forums provide answers to many real-life questions.
Approach: They propose to make Persian dataset PerCQA public to encourage more research in Persian CQA.
Outcome: The proposed dataset contains 989 questions and 21,915 annotated answers from the most well-known Persian forum.
MDP-GRPO: Stabilized Group Relative Policy Optimization for Multi-Constraint Instruction Following (2026.acl-long)

Copied to clipboard

Challenge: Large language models (LLMs) can follow many natural-language instructions, yet they remain brittle when a request bundles multiple explicit constraints, such as asking the LLM to respond in a particular structure with an exact ending phrase.
Approach: They propose a method which stabilizes learning through multi-temperature sampling to increase reward dispersion, dual-anchor advantages to restore gradients in homogeneous groups, prospect-theoretic shaping to bound updates and penalize violations based on Kahneman Tversky’s theory and asymmetric KL regularization.
Outcome: The proposed method outperforms standard GRPO on FollowBench, IFEval, and a curated multi-constraint dataset, improving strict constraint satisfaction by up to 5.0% on Llama-3.2-3B.
Esposito: An English-Persian Scientific Parallel Corpus for Machine Translation (2024.lrec-main)

Copied to clipboard

Challenge: Existing scientific corpus for English-Persian language pairs is lacking . supervised neural machine translation requires millions of parallel sentences .
Approach: They propose a parallel corpus called Esposito which contains 3.5 million parallel sentences . they also propose 'test sets' that might serve as a baseline for future studies .
Outcome: The proposed system improves the baseline on average by 7.6 and 8.4 BLEU scores for English-Persian language pairs.
IRUEX: A Study on Large Language Models Problem-Solving Skills in Iran’s University Entrance Exam (2025.coling-main)

Copied to clipboard

Challenge: Recent advances in Large Language Models (LLMs) have profound implications for education.
Approach: They present a novel multiple-choice educational resource specifically designed to evaluate the performance of Large Language Models (LLMs) they use a dataset that contains 868 questions and 36,485 additional questions .
Outcome: The IRUEX dataset contains 868 questions and 36,485 additional questions.
SchemaGraphSQL: Efficient Schema Linking with Pathfinding Graph Algorithms for Text-to-SQL on Large-Scale Databases (2026.findings-eacl)

Copied to clipboard

Challenge: Text-to-SQL systems translate natural language questions into executable SQL queries.
Approach: They propose a schema linking approach that first constructs a graph based on foreign key relations and then uses a single prompt to a lightweight LLM to extract source and destination tables from the user query.
Outcome: The proposed method outperforms specialized, fine-tuned, and complex multi-step approaches on BIRD and Spider 2.0 benchmarks.
PMI-Align: Word Alignment With Point-Wise Mutual Information Without Requiring Parallel Training Data (2023.findings-acl)

Copied to clipboard

Challenge: Recent studies show that using contextualized embeddings from pre-trained multilingual language models could give us high quality word alignments without the need of parallel training data.
Approach: They propose a method which uses contextualized embeddings from pre-trained language models to extract word alignments without parallel training.
Outcome: The proposed method outperforms rival methods on five out of six language pairs.
EPOQUE: An English-Persian Quality Estimation Dataset (2024.lrec-main)

Copied to clipboard

Challenge: Existing human labeled QE datasets are limited to limited language pairs . a small subset of the proposed dataset can improve its performance by 8% .
Approach: They propose to use an English-Persian QE dataset with manually annotated direct assessment labels to evaluate translation quality estimation models.
Outcome: The proposed dataset improves on two state-of-the-art QE models by 8% . the proposed dataset contains 1000 translated sentences from English to Persian .
Stable Evidence, Unstable Decisions: An Empirical Analysis of Model Decision Stability in Vision–Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) provide visual information alongside their predictions, but it remains unclear whether consistency in such information implies consistent decisions.
Approach: They construct configurations that retain lesion content while varying surrounding context and scale and measure decision flips together with consistency in model-reported influential slices.
Outcome: The proposed models show that flip rates reach up to 75% across lesion-containing presentations, despite high overlap in reported evidence.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations