Manh Nguyen

AI researcher | PhD Student @ Applied Artificial Intelligence Initiative

prof_pic.jpg

Hello, my name is Manh. I am a final-year PhD student at A2I2, Deakin University, advised by Dr. Hung Le and Prof. Sunil Gupta. My research centers on the reliability and efficiency of large language models, with a focus on uncertainty estimation, test-time scaling, and multi-agent systems. Across these threads, I focus on practical NLP research, building robust and scalable systems that turn research ideas into real-world impact.

Prior to my PhD, I spent 5+ years building production ML systems in industry.

Update: I’m open to roles in both academia (Postdoc / Research Fellow) and industry (Data Scientist, Applied Scientist, ML Engineer), available full-time from April 2027 (estimated).

news

Oct 19, 2026 Served as a Reviewer for ICLR 2027 and as a Program Committee Member for SOICT 2026.
Aug 12, 2026 Co-organizing the OE-Agent workshop on “Open and Efficient Agentic AI for Resource-Constrained Environments” at ACML 2026. Submissions are welcome — deadline 18 Oct 2026.
Jun 12, 2026 Won the People’s Choice Award at the Deakin A2I2 Three Minute Thesis (3MT) Heat for my talk “From Confident Lies to Trusted Answers”.
Jun 10, 2026 Co-presented a talk on “Uncertainty Estimation in LLMs” with Dr. Hung Le at the StatML Seminar Series, CSIRO.
May 28, 2026 One paper, titled “Retrieval-augmented Decoding for Improving Truthfulness in Open-ended Generation” has been accepted to ECML PKDD 2026.

selected papers

  1. PRO (AAAI-26)
    pro_logo.svg
    Probabilities Are All You Need: A Probability-Only Approach to Uncertainty Estimation in Large Language Models
    Manh Nguyen, Sunil Gupta, and Hung Le
    The 40th Annual AAAI Conference on Artificial Intelligence, 2026
    TL;DR: Estimates predictive entropy from top-K output-token probabilities with an adaptive K; +2.4% avg AUROC over the best baseline across 3 QA datasets.
    arXiv Bib Code Poster
  2. RAD (ECML 2026)
    rad_logo.svg
    Retrieval-augmented Decoding for Improving Truthfulness in Open-ended Generation
    Manh Nguyen, Sunil Gupta, and Hung Le
    2026
    TL;DR: Builds a tiny reference store of truthful logits from  10 examples to steer decoding; +3.2% avg truthfulness on TruthfulQA, no retraining.
    arXiv Bib Code Poster
  3. RCS
    rcs_logo.svg
    Efficient Best-Of-N with Radial Consensus Score
    Manh Nguyen, Sunil Gupta, and Hung Le
    2026
    TL;DR: Ranks sampled LLM answers by distance to their embeddings’ semantic center; +2-7% selection accuracy over strong baselines incl. majority voting.
    arXiv Bib Code
  4. UAB
    uab_logo.svg
    Uncertainty-Aware Budget Allocation for Adaptive Test-Time Reasoning
    Manh Nguyen, Sunil Gupta, and Hung Le
    2026
    TL;DR: Reallocates a fixed sampling budget toward harder questions using early-round disagreement as a difficulty signal; +2.3% avg accuracy over uniform allocation (up to +5.5% peak).
    arXiv Bib Code
  5. DAR
    darlogo.webp
    Hear Both Sides: Efficient Multi-Agent Debate via Diversity-Aware Message Retention
    Manh Nguyen, Anh Nguyen, Dung Nguyen, and 2 more authors
    2026
    TL;DR: Each round, broadcasts only the most disagreeing agent responses instead of every message, cutting multi-agent debate noise while keeping useful diversity.
    arXiv Bib Code
  6. TB-MAD
    tbt_overview.svg
    Think More or Read More? A Token-Budget Theory of Reasoning with Shared Context
    Manh Nguyen, Sunil Gupta, and Hung Le
    2026
    TL;DR: Predicts the optimal number of debate agents from single-agent measurements alone; self-consistency matches or beats debate in 93% of 980 settings under tight token budgets.
    arXiv Bib