Prakhar Gupta

Machine translation benchmarks have functionally saturated, but pairing human-authored failure cases with deterministic verification rules reveals critical multimodal blind spots that automated metrics consistently miss.

Grounding improvements in language models are largely driven by existing architecture rather than new mechanisms, challenging assumptions about post-training efficacy.

Alignment tuning installs distinct bias directions in LLMs, allowing for targeted debiasing that recovers unbiased answers while maintaining performance.

Hierarchical orchestration can drastically reduce decision-space explosion in LLMs, enhancing both routing accuracy and memory efficiency.

RMCT reveals that you can reduce bias in language models without sacrificing their ability to articulate the very cues you're trying to mitigate.

Forget catastrophic forgetting: sparse memory finetuning, enhanced with a KL-divergence-based update rule, lets LLMs learn continuously without trashing old knowledge.