Applied AI research

Mohit Goyal's research measures things that matter when building AI systems: how few parameters a fine-tune needs, how agent frameworks are put together, and whether a model trusts a verifier it should not. He has 3 publications and preprints, plus one research sprint result.

arXiv preprint · Jan 2026PEFT · LLMs

GRIT: fine-tuning with fewer trainable parameters

arXiv preprint 2601.00231, January 2026. Fourth of nine authors (first author P. Saha). Work done at RAAPID.

46%fewer trainable parameters than LoRA, on average

Read summary

GRIT is a parameter-efficient fine-tuning method. It trains about 46% fewer parameters than LoRA on average (25 to 80% depending on the task), while matching or beating LoRA and QLoRA.

It works by using information about how sensitive the model's loss is to each direction of change, and spending the small update only where it helps most. The update's size also changes during training.

The paper tests Llama-3.2 3B and Llama-3.1 8B on Alpaca, Dolly-15k, BoolQ, QNLI and GSM8K. On GSM8K, the 8B model reached the best accuracy in the comparison (0.662).

His part: experiment design, the reproducible evaluation setup, benchmark comparisons and ablations.

Apart Research sprint · Nov 2026Evals · Verification

False assurance in AI source checkers

Apart Research AI x Epistemics sprint, November 2026, with one teammate. Sprint pilot.

95%of the time Opus still agreed with a wrong "supported" verdict

False assurance is when a checker wrongly says a claim is supported by its sources, and the reader accepts that verdict.

Read summary

In a pilot with 3 chat models (Claude Opus 5.5, Claude Haiku 5.5 and Gemini 3.8 Flash), a wrong "the sources support this" verdict from a source checker was harder to undo than a wrong "they do not". The models read 80 Helpful Community Notes under 8 conditions, for 1,920 answers.

Shown the passage the checker read, the models caught its false alarms. Its false assurances mostly stood: when wrongly told the sources support a claim, Opus still agreed 95% of the time, Gemini 92.5% and Haiku 60%. For no model was a passage from an unrelated page less effective than the real passage, which suggests the models reacted to seeing a passage, not to what it said.

A separate audit of 200 Helpful notes found that 68% of the judgeable cited links do not support the note's claim.

Limits: 3 chat models, 1 prompt, 80 notes. It says nothing yet about agents or people.

Research Square preprintZero trust · Explainability

AoT-BTS: explainable trust scoring for zero-trust networks

Research Square preprint. M. Goyal (first author), A. Singh, S. Soni, S. Khatri and Prof. Mohit Tiwari, at BVCOE.

98.31%accuracy

Read summary

AoT-BTS scores each device's behavior over time to decide, in real time, whether a software-defined network should keep trusting it. It reports 98.31% accuracy, a 0.32% false positive rate and 17.31 ms p99 latency.

Each decision comes with a SHAP explanation, so an operator can see why a device lost trust.

IEEE DELCON 2025Agent frameworks

A review of Microsoft Semantic Kernel

2025 IEEE DELCON, New Delhi. M. Goyal, with A. Gupta.

Read summary

This paper reviews Microsoft Semantic Kernel, an open-source framework for building AI agents into enterprise software. It covers the framework's architecture, capabilities, agent design patterns and orchestration, and open research directions for enterprise agent frameworks.

“A man who is master of patience is master of everything else.”

~ George Savile

© 2026 Mohit Goyal