Atta Ul Asad

I am an M.S. student in Artificial Intelligence at the Lahore University of Management Sciences (LUMS). I study how large language models fail as decision support in high-stakes settings, using clinical diagnosis as my main testbed. My current project asks whether showing an LLM a clinician's working diagnosis suppresses its ability to flag that diagnosis as an error. More broadly, I audit failures that aggregate accuracy hides, such as the loss of faithfulness under KV-cache quantization in retrieval-augmented generation.

In summer 2026 I was a research intern at the University of Oklahoma with Dean F. Hougen. I also collaborate with the Stanford AI Lab (SAIL) on KV-cache compression and self-improving coding agents. I received my B.S. in Computer Science from COMSATS University Islamabad.

Email  /  CV  /  Scholar  /  OpenReview  /  GitHub  /  LinkedIn

Photo of Atta Ul Asad

Updates

  • Sep 2026 LORE-KV, our KV-cache eviction paper, is under review at ICLR 2027.
  • Aug 2026 Faithfulness Is Not Free is on arXiv and accepted at the GroundLM workshop at EMNLP 2026.
  • Aug 2026 Started my M.S. in Artificial Intelligence at LUMS.
  • Jun 2026 Joined the University of Oklahoma as a research intern with Dean F. Hougen.

Research

I work on trustworthy, human-centered AI for high-stakes decisions: when LLM assistance improves a decision, when it propagates errors through anchoring or sycophancy, and how to evaluate models by their effect on the final decision rather than by standalone accuracy. * denotes equal contribution.

Schematic of LORE-KV: sampled continuations score prompt tokens for eviction Monte Carlo Future-Query Estimation for KV Cache Eviction
A. Bilal, M. A. Mohsin, M. Umer, W. H. Raza, A. U. Asad, Y. D. Kwon, M. Valko, D. F. Hougen
Under review at ICLR 2027

LORE-KV scores each prompt token by how much deleting it would change the attention output for short continuations sampled from the frozen model, so the cache keeps the tokens that future decoding queries need. At a 128-token budget, LongBench on Qwen2.5-14B goes from 45.49 to 48.24 and 16K RULER on Mistral-7B from 45.20 to 51.05; a single sampled continuation recovers about 89% of the gain.

Schematic: INT4 KV-cache quantization keeps answers correct but less faithful to the documents Faithfulness Is Not Free: Auditing Offline KV-Cache Quantization in Retrieval-Augmented Generation
A. U. Asad*, A. Bilal*, M. Ali, M. Haseeb, D. F. Hougen
GroundLM Workshop at EMNLP, 2026
arXiv

INT8 document KV caches are near-lossless. Under INT4, over 90% of faithfulness changes among answers that stay correct are negative, a regression that accuracy metrics miss, and the harm grows with noisy retrieval and more chunks.

Ongoing Work and Preregistered Studies
Second-Opinion Independence in Clinical LLM Decision Support
In progress

When an LLM serves as a second opinion, does seeing the clinician's working diagnosis suppress its ability to flag that diagnosis as wrong? The study compares documented real misdiagnoses with synthetic errors and measures the change in error detection, not standalone accuracy.

Quantized Self-Correction Audit
Preregistered study
code

3-bit AWQ makes DeepSeek-R1-Distill-Qwen-1.5B loop on MATH-500 and cuts its accuracy from 0.85 to 0.39. The "overthinking"-marker penalty of Lotfi et al. (2026) is statistically indistinguishable from a one-line repetition penalty on loop rate and accuracy.

Re-derivation Leakage in Multi-Agent Unlearning
Preregistered study
code

Shared memory that enforces a retrieval-time ID blocklist and deletion over the derivation closure can still store forgotten TOFU facts: a node written from an agent's parametric answer has a fresh ID and no parent edges, so both checks certify it clean.

Self-Improving Coding Agents
In progress, with M. A. Mohsin and M. Umer

Whether a Qwen2.5-14B coding agent improves through reinforcement learning on its own sampled trajectories (HumanEval+, MBPP+), and when self-training yields real gains rather than reinforcing the agent's own errors.

Education and Experience
LUMS M.S. in Artificial Intelligence, Aug 2026 – present
University of Oklahoma Research Intern with Dean F. Hougen, Jun – Aug 2026
University of Malaya Undergraduate Research Assistant with Uzair Ishtiaq (clinical NLP), Jul – Sep 2025
COMSATS University Islamabad B.S. in Computer Science, 2022 – 2026

Service and Teaching
Reviewer GroundLM @ EMNLP 2026, VLM4RWD 2026
Teaching Assistant Machine Learning (Fall 2025) and Natural Language Processing (Spring 2025), COMSATS University Islamabad

Design and source code from Jon Barron's website. Last updated September 2026.