Eran Hirsch

Eran Hirsch

PhD Candidate, Bar-Ilan University

About

I am a PhD student at Bar-Ilan University, advised by Prof. Ido Dagan.

My research addresses the "hallucination" problem in Large Language Models, focusing on making grounded text generation more verifiable[1,2,3,4] and factual[5]. I also work on QA-based semantics[6,7], planning with LLMs[8,9], and text-to-image generation[10,11].

Previously, I spent over a decade in the industry as a Data Scientist, Software Engineer and Team Lead.

References
  1. Hirsch et al. "LAQuer: Localized Attribution Queries in Content-grounded Generation." ACL 2025.
  2. Slobodkin*, Hirsch* et al. "Attribute First, then Generate: Locally-Attributable Grounded Text Generation." ACL 2024.
  3. Wan, Hirsch et al. "GenerationPrograms: Fine-grained Attribution with Executable Programs." COLM 2025.
  4. Alt, Hirsch et al. "User-Centric Evidence Ranking for Attribution and Fact Verification." EACL 2026.
  5. Harary, Hirsch et al. "PrefixNLI: Detecting Factual Inconsistencies as Soon as They Arise." Preprint.
  6. Klein, Hirsch et al. "QASem Parsing: Text-to-text Modelling of QA-based Semantics." EMNLP 2022.
  7. Roit, Slobodkin, Hirsch et al. "Explicating the Implicit: Argument Detection Beyond Sentence Boundaries." ACL 2024.
  8. Hirsch et al. "What's the Plan? Evaluating and Developing Planning-Aware Techniques for LMs." Preprint.
  9. Hirsch et al. "Evaluating Language Models Planning Capabilities on Goal Ordering Challenges." NeurIPS WS 2024.
  10. Rassin, Hirsch et al. "Linguistic Binding in Diffusion Models: Enhancing Attribute Correspondence through Attention Map Alignment." NeurIPS 2023.
  11. Binyamin et al. "Make It Count: Text-to-Image Generation with an Accurate Number of Objects." CVPR 2025.

📌 Selected Research

LAQuer teaser
ACL 2025
LAQuer: Localized Attribution Queries

A method for fine-grained attribution in content-grounded generation, allowing models to cite specific spans rather than full documents.

Paper Code
Attribute First teaser
ACL 2024
Attribute First, then Generate

A multi-step generation process creating localized attribution: select spans from the source, create a plan, then generate the output.

Paper Code
View All Publications →

📢 News

Jan 2026 Evidence Ranking accepted to EACL '26 main conference.
Jan 2026 Released AI Fact Checker, a Chrome extension for verifying claims in AI output against source citations.
Dec 2025 Served on the organizing committee of ISCOL '25.
Nov 2025 PrefixNLI preprint is now available online.
Jun 2025 GenerationPrograms accepted to COLM '25!
Jun 2025 Released CLATTER preprint.
May 2025 LAQuer accepted to ACL '25 main conference.