Prasham
Shah

I led the CS research team on this project and first-authored both the resulting IEEE paper, first presented at the David Miyamoto Scholars Conference, and a distilled poster, awarded Best Poster at IEEE MIT URTC among roughly 700 total conference submissions. My team replicated Udrescu & Tegmark's AI Feynman (2020) from scratch — regression, functional feature expansion, and symmetry detection — to expose its core mechanisms, then extended it to address two problems the original left unresolved: noisy data and variable numerical scales. Testing on synthetic physics datasets with Gaussian noise, the original system rediscovered equations with near-perfect accuracy at zero noise (R² ≈ 1.00 for most physics-derived equations) but degraded sharply as noise increased, falling to R² ≈ 0.49 at σ = 0.2 and collapsing below zero at σ = 0.4, with failures showing up as either underfitting or bloated, spurious equations. My most significant contribution — a novel adaptive-α rule that scales symmetry-detection thresholds to each dataset's distribution — nearly doubled true symmetry detections (60 → 110) while cutting false positives tenfold (29 → 3), with the largest gains at small numerical ranges. This key improvement makes the technology much more useful for cutting-edge science.

I independently studied a question that hadn't yet been examined in the literature: while the noise resistance of classification models and text-only LLMs was known, no one had measured it for multimodal LLMs (MLLMs). I tested two MLLMs (GPT-4o and LLaVA) against textual noise (misspellings) and image noise (Gaussian, salt-and-pepper, and speckle), alongside a text-only baseline (GPT-3.5), across 10 textual and 30 image-based prompts — each run clean, noised, and denoised (via aspell for text, OpenCV's Fast NL Means for images). The hypothesis that MLLMs would show worse noise resistance than a traditional LLM was supported: MLLMs degraded significantly more on noisy text than GPT-3.5 did (p=0.025), with the smaller model (LLaVA) hit hardest (p=0.048). The second hypothesis — that denoising would help — was refuted: classical denoising significantly hurt text performance instead of improving it (p=0.010), and showed no reliable benefit on images either. A secondary, non-central finding suggested lower-parameter models fare worse under noise, consistent with the LLaVA result, though the study wasn't designed to isolate model size as a variable. The practical takeaway: MLLM users should focus on writing clean prompts and providing clear images up front rather than leaning on traditional algorithmic denoisers.