Saim Rehman

Research profile

Research summary

Saim Rehman is an early-stage researcher whose public bibliographic footprint — resolved from OpenAlex — currently consists of four works, all indexed in 2026. The two preprints reviewed by Evolve, both first-authored and posted to arXiv in 2026, share one research question: when neural networks are compressed in accuracy-preserving ways, does anything besides accuracy change? AERIAL evaluates adversarial robustness in low-precision EEG decoders for brain-computer interfaces and finds that compression alters adversarial transfer rather than direct robustness, in a well-controlled single-dataset design Evolve judges largely supported. GHOST-Q examines quantized vision-language models and shows that headline scores can mask item-level shifts in grounding outcomes — object hallucination measured on POPE and AMBER — with rigorous paired statistics but small effects and a bitsandbytes-only scope. Together the papers position Rehman at the intersection of efficient deep learning and reliability evaluation, spanning neural decoding and multimodal models. Citation indicators sit at a starting point (h-index 0, i10-index 0), and the collaboration footprint is narrow: one distinct coauthor, Muhammad Shafique, with first-author positions on both papers and no corresponding-author role recorded. OpenAlex topic labels additionally list quantum-computing and materials-science themes that do not overlap with the reviewed work, so this profile should be read as an early-stage, name-resolved snapshot.

Research keywords

Highlights

• AERIAL: compression in EEG decoders shifts adversarial transfer, not direct robustness

  First-authored 2026 arXiv study of accuracy-preserving low-precision EEG decoders for brain-computer interfaces. Evolve's verdict: a well-controlled single-dataset study showing that compression alters adversarial transfer rather than direct robustness, with claims largely supported. Reproducibility readiness rated medium.

• GHOST-Q: quantized VLMs keep headline scores but shift grounding outcomes

  First-authored 2026 arXiv preprint showing that post-training quantization preserves aggregate scores on vision-language models while item-level grounding and object-hallucination outcomes (POPE, AMBER) shift. Evolve notes rigorous paired statistics, but small effects and a bitsandbytes-only scope; reproducibility readiness medium.

• A unifying question: does efficiency preserve behavior?

  Across both 2026 first-authored papers, accuracy-preserving compression (pruning, quantization) is stress-tested for side effects — adversarial transfer in EEG decoders and item-level grounding shifts in VLMs — with matched statistical controls applied in each study.

Timeline

1. 2026First reviewed preprint, AERIAL, appears on arXiv: an adversarial evaluation of accuracy-preserving low-precision EEG decoders, first-authored.
2. 2026Second first-authored arXiv preprint, GHOST-Q, studies grounding hallucinations overlooked under same-score trade-offs in quantized vision-language models.
3. 2026OpenAlex records 4 works in 2026 — the only year with indexed output in this snapshot (h-index 0, i10-index 0, 1 distinct coauthor).

Papers in the Evolve catalog

• 2026-09-24first author

  AERIAL: Adversarial Evaluation of Robustness in Accuracy-Preserving Low-Precision EEG Decoders

• 2026-09-24first author

  GHOST-Q: Towards Studying Grounding Hallucinations Overlooked Under Same-score TradeOffs in Quantized VLMS

Frequent co-authors