Md Rizwan Parvez

I am a Scientist at the Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa University, where I lead research on auditable language and coding agents. Before that, I was a Research Scientist at Bosch Research. I received my Ph.D. in Computer Science from the University of California, Los Angeles, where I worked with Kai-Wei Chang, and interned at Google Research, Meta FAIR, Salesforce Research, and Microsoft. My research focuses on natural language processing, machine learning, and software engineering, with the goal of building agents whose claims, code, and actions can be checked against evidence, execution, and the environment they act in, rather than taken on the model’s word. Currently, my research includes three directions: (1) attributed knowledge and retrieval-augmented generation, (2) verifiable code generation and software engineering agents, and (3) grounded agents for maps, the web, GUIs, and charts, together with the safety and security of these agents.

I am actively looking for research interns, and I co-supervise Ph.D. students with university partners. Please note that I currently do not have funding for paid positions, so internships are unpaid research assistantships. If you are interested in AI agents, code generation and software engineering, or retrieval-augmented generation, please check the recruiting information here.

News

2026I will serve as a Senior Area Chair for ACL 2027, an Area Chair for NAACL 2027, and a Senior Program Committee member for AAAI 2027.
2026Our survey on agentic security is published in TMLR.
20262 papers accepted at EMNLP 2026: GUI-Primitives, Reward Engineering for Software Tasks.
2026ChartQAPro is used in the evaluation of Anthropic’s Claude Mythos/Fable 5.
2026Invited talk on language agents for software development at the UNT Data Science Seminar.
05.20262 papers accepted at ICML 2026: Poison with Style, TimeSpot.
04.20263 papers accepted at ACL 2026 Findings: CompassLLM, Mina, CoT-PoT.
01.20261 paper accepted at ICLR 2026: SpatiaLab.
01.20263 papers accepted at EACL 2026 Findings: MapAgent, Conformal Lens, DashboardQA.
12.2025We release WebOperator, an open-source web agent among the top scorers on WebArena.
12.2025I will serve as a Senior Area Chair for ACL 2026.
11.2025We release a survey on agentic security covering 150+ papers.
10.2025I will serve as a Senior Area Chair for EACL 2026.
2025MapEval is selected as a spotlight at ICML 2025.

Research

Large language models have made natural language an interface to computation, yet they fail in ways users cannot easily detect: answers cite unsupported sources, generated code runs but returns wrong results, and agents act on outdated views of their environment. My guiding principle is to never let the model be its own only witness. An answer should be checked against its cited passages, a program against its specification and execution, and an action against observations of the environment it will change. My research builds these checks throughout an agent’s workflow, along three connected threads.

Auditable agents

Attributed

Knowledge

What does the agent know?

Open-RAG FiLCo DelucionQA CoT-PoT Xolver

retrieval seeds code generationRedCoder

execution as a second witnessCoT-PoT

Verifiable

Code

What does the agent produce?

RedCoder MapCoder CodeSim xCodeEval Xolver

simulation and tree searchWebOperator

Grounded

Action

How does the agent act?

MapEval WebOperator SpatiaLab ChartQAPro

Trust and safety, across all three Can the agent be trusted? X-Teaming Poison with Style AI Debate Agentic Security AgentTell

KnowledgeAttributedWhat does the agent know?

Attribution is a pipeline property, not a post-hoc citation. Provenance survives to the output only if valuation, retrieval, filtering, and generation each expose an auditable decision. Open-RAG turns an open LLM into a sparse mixture-of-experts that ignores distractors and decides when to retrieve; its Llama2-7B version outperforms ChatGPT-RAG and Self-RAG. FiLCo learns which context to keep, DelucionQA detects hallucination, and CoT-PoT reaches self-consistency from two samples. The same machinery runs in legal aid (Mina), cyber-threat intelligence (TechniqueRAG), and medical screening (DR-VQA).

Source data valuation

Source valuation Budgeted selection VideoLights

Retrieval and filtering

RedCoder FiLCo ERA DelucionQA Open-RAG

Retrieval-augmented reasoning

Open-RAG Chain of Evidences CoT-PoT

Attribution and audit

DelucionQA Conformal Lens

Applied domains

TechniqueRAG MinaDR-VQA

CodeVerifiableWhat does the agent produce?

The bottleneck is correctness, not fluency. Code can be executed and tested, so it is where agent outputs can be verified. RedCoder was among the first retrieval-augmented code generators; MapCoder extended it to a recall–plan–code–debug team of agents, and CodeSim verifies plans by simulation before code exists (95.1% on HumanEval). Xolver carries this to general problem solving with shared, evolving memory, and xCodeEval evaluates code by execution at scale. At repository scale, SWE-Xplorer reconciles parallel search paths (53.2% to 60.8% on SWE-Bench Verified), and our agents resolve issues through discussion and review real pull requests.

End-to-end code pipeline (figure from my research statement)

Language-to-code pipeline: federated coding agents retrieve, plan, simulate and verify, and execute and debug; a judge agent reconciles feedback into evolving memory; discussion and reviewer agents resolve issues behind safety guardrails.

Experience and skill memory

RedCoder MapCoder Xolver

Foundation models for code

Text and Code LM RedCoder PwS

AI coding agents

MapCoder CodeSim Xolver

Software engineering agents

SWE-XplorerSWE-Discussion

Evaluation and judge agents

xCodeEvalVibeSweBenchPR Reviewer

ActionGroundedHow does the agent act?

Instrument first, then build the agent. On maps, the web, GUIs, and charts, models recite state they cannot operate on. MapEval exposed a gap of more than 20 points to humans in geo-spatial reasoning and grew into MapAgent, SpatiaLab, and TimeSpot. WebOperator searches web actions while ranking them by reversibility, GUI-Primitives isolates screen grounding as the bottleneck for computer-use agents, AgentTell shows that browser-use agents leak information through behavioural side channels, and ChartQAPro and DashboardQA test agents on the charts and dashboards organizations reason with.

Tool-calling agents

MapQaTor MapEval MapAgent

Spatio-temporal grounding

SpatiaLab TimeSpotUrbanAgentBench

Visual data agents

ChartInstruct DataNarrative ChartQAProChartQA-VideoChartOrchestra

Web and computer-use agents

WebOperator GUI-Primitives AgentTell DashboardQA

SafetyTrustworthyCan the agent be trusted?

Trust has to be built in, then audited. Our critical review of LLM evaluation diagnoses why model comparisons mislead. X-Teaming red-teams multi-turn conversations with collaborating planner, attacker, and verifier agents; Poison with Style shows the code supply chain can be poisoned with almost no utility loss; AI Debate helps human judges assess disputed claims; and our agentic security survey and AgentTell map threats to agents, including behavioural side-channel leakage in browser-use agents.

Evaluation

LLM evaluation review

Red-teaming

X-Teaming

Code poisoning

PwS

Security of agents

CTF failure attributionAgentic Security AgentTell

Oversight

AI Debate

Where this is heading

A unified, attributed, and safety-certified knowledge substrate for everything an agent consumes, from documents and memory to skills and tools.
A research platform for trustworthy human–AI software engineering across the full development lifecycle.
Agents that stay grounded as the physical and digital world changes under them.

More on the research page and in my research statement (PDF).

Selected Publications

Full list on the publications page and Google Scholar.

Retrieval and Knowledge

Under review
Xolver: Multi-Agent Generalist Problem Solving with Holistic Experience Learning

Md Tanzib Hosain, Salman Rahman, Md Kishor Morol, and Md Rizwan Parvez

Top-performing results on 14 reasoning, math, coding, and agentic tasks, including AIME, LiveCodeBench, GPQA, HLE, BrowseComp, GAIA, OS-World, and SWE-Bench Verified

paper bibtex
```
@misc{hosain20254xolver,
  title = {{Xolver: Multi-Agent Generalist Problem Solving with Holistic Experience Learning}},
  author = {Hosain, Md Tanzib and Rahman, Salman and Morol, Md Kishor and Parvez, Md Rizwan},
  howpublished = {AI4Math Workshop at ICML 2026; under review},
  year = {2026}
}
```
Details

EMNLP Findings ’24

Open-RAG: Enhanced Retrieval Augmented Reasoning with Open-Source Large Language Models

Shayekh Bin Islam, Md Asib Rahman, K S M Tozammel Hossain, Enamul Hoque, Shafiq Joty, and Md Rizwan Parvez

Press coverage: MarkTechPost

paper bibtex

@inproceedings{shayekh2024openrag,
  title = {{Open-RAG: Enhanced Retrieval Augmented Reasoning with Open-Source Large Language Models}},
  author = {Islam, Shayekh Bin and Rahman, Md Asib and Hossain, K S M Tozammel and Hoque, Enamul and Joty, Shafiq and Parvez, Md Rizwan},
  booktitle = {Findings of EMNLP 2024},
  year = {2024}
}

News More News

Research

Selected Publications

Retrieval and Knowledge

Code Generation and Software Engineering

Grounded Actions: Maps and Spatial Reasoning

Grounded Actions: Charts and Visual Data

Grounded Actions: Web and Computer-Use Agents

Safety, Security, and Trust

Experience

Mentees

Service

News