Ayush Anand

Language models are getting eerily good at generating step-by-step reasoning, but actually verifying if those steps are logically sound is a completely different story. In this paper, we explore a strange quirk: why models sometimes scale negatively when forced to separate syntactic text generation (sounding smart) from true logical verification (being right). We map out exactly where these reasoning structures begin to break down as you scale them up.

Predicting crop yields with AI usually fails when the climate shifts, because standard deep learning models just overfit to historical weather patterns. I wrote this paper to show that if you force the AI to respect actual biology - using evolutionary algorithms to pick physical drivers like soil clay and vapor pressure - you can fix this. We achieved predictive parity \((R^{2}\approx0.85)\) with massive CNN-LSTM ensembles while using 99% less data dimensionality. Basically, embedding physics beats brute-force data scaling.

Before building my own architectures for earth observation, I needed to understand everything that had been tried before. I read and systematically reviewed 80 papers spanning a decade (2014-2024). I mapped out where state-of-the-art algorithms (like Random Forests and LSTMs) succeed, assessed the remote sensing data everyone uses, and outlined the persistent evaluation challenges the community still faces.

I am super curious about how features entangle inside LLMs. People like to think of neural networks as having independent "modules" for concepts, but my ongoing virtual ablation experiments on Gemma-2 show that trying to "unlearn" hazardous knowledge usually fails in the middle layers because the model just dynamically routes around the ablation. To measure this, I built a normalized Virtual Ablation Synergy (\(L_{2}\)) metric directly in the residual stream. Turns out, if we want safe AI, we have to target very specific late-layer bottlenecks.