Shashi Bhushan TN

—

Unclaimed · AI-structured profilename-matched identity

Applied NLP researcher evaluating and fine-tuning LLMs for customer support and agent workflows

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-14) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Publication-volume data for this researcher is not available in the current snapshot, so the output dimension cannot be scored. This axis would measure the number, consistency, and cadence of papers and preprints produced over time.

Impact—

Citation and influence metrics are not present in this record, so impact is unscored. This axis would measure citations to the researcher's work, the works referencing it, and the uptake of its methods by other teams.

Focus0/100

The focus dimension registers 0, reflecting a field distribution split exactly 50/50 between “natural language processing” and “natural language processing (LLM agent evaluation)” across the two cataloged papers. In practice the two labels are adjacent — both papers fine-tune or evaluate LLMs for support and agent workflows — so the scope reads as deliberate applied focus rather than scattering, though a two-paper sample is too small for a firm judgment.

Collaboration56/100

The collaboration dimension scores 56, based on 16 distinct coauthors recorded in the snapshot. The frequent-coauthor list — including Xue-Yong Fu, Md Tahmid Rahman Laskar, Shi Zong, and Gundeep Singh among twelve recurring names — points to stable team-based work, and the researcher sits in the last author position on both cataloged papers, a slot commonly associated with senior or supervisory contributions. The moderate score suggests a real but still compact collaboration footprint.

Activity—

Recent publishing cadence cannot be established from the available snapshot, so activity is not scored. This axis would measure ongoing research output and engagement across recent years.

Reproducibility38/100

Reproducibility scores 38, computed as a mean of 38 across the 2 assessed papers. Evolve's verdicts explain the ceiling: both studies are controlled, well-designed comparisons, but the fine-tuning paper lacks significance tests and roughly half its data is private (reproducibility readiness medium), while the agent-evaluation paper relies on proprietary data with a small evaluation set (readiness low).

Research summary

Shashi Bhushan TN is an applied natural language processing researcher whose cataloged work centers on evaluating and adapting large language models for real production scenarios, most visibly customer support. Evolve's catalog lists two 2026 arXiv papers, in both of which the researcher is the last author. The first, “Can One Adapted Model Do It All?”, runs a controlled comparison of fine-tuning strategies — multi-task full fine-tuning, LoRA and other parameter-efficient variants, model merging — across 13 support LLMs and finds multi-task full fine-tuning strongest overall. The second, “When Better Turns Do Not Make Better Agents”, diagnoses an evaluation gap: supervised fine-tuning lifts next-turn metrics yet yields 0/77 holistic workflow success, arguing that turn-level scores poorly proxy end-to-end task completion. The bibliometric snapshot records 16 distinct coauthors and an even 50/50 field split between natural language processing and LLM agent evaluation; the collaboration dimension scores 56, reproducibility 38, and focus 0. Evolve's reviews consistently flag controlled designs undermined by proprietary data, small evaluation sets, and absent significance testing, capping reproducibility readiness at medium and low. Output, impact, and activity cannot be scored from the available data. Overall, this is an early-stage profile of an industry-oriented researcher working to make LLM adaptation and agent evaluation measurable — and honest about where turn-level metrics fail.

Collaboration pattern

Sixteen distinct coauthors are recorded for this researcher, with a recurring core including Xue-Yong Fu, Md Tahmid Rahman Laskar, Shi Zong, and Gundeep Singh. The twelve-name frequent-coauthor list and the last-author placement on both cataloged papers point to stable, team-style, industry-flavored work on support LLMs rather than loose ad hoc collaborations. The collaboration dimension scores 56, consistent with an established but still compact network centered on one applied research agenda.

Research keywords

LLM agent evaluationnatural language processingfine-tuning strategiesparameter-efficient fine-tuning (LoRA)multi-task learningcatastrophic forgettingmodel mergingsupervised fine-tuningLLM agents and tool callingnext-turn evaluationworkflow success metricscustomer support LLMs

Highlights

Multi-task full fine-tuning wins across 13 customer-support LLMs
In a controlled comparison of fine-tuning strategies spanning 13 support LLMs, the study finds multi-task full fine-tuning the strongest overall option, benchmarked against LoRA and other parameter-efficient variants, model merging, and catastrophic-forgetting trade-offs. Evolve calls the result useful while noting the absence of significance tests and that half the data is private.
Turn-level gains do not translate into workflow success
The paper shows supervised fine-tuning improves next-turn metrics yet yields 0/77 holistic workflow success, diagnosing why turn-level evaluation is a poor proxy for end-to-end agent performance and arguing for gold-history evaluation of tool-calling workflows. Evolve rates the controlled design solid but flags proprietary data and a tiny evaluation set.
A consistent methodological signature: rigorous comparisons, closed data
Across both cataloged papers the pattern repeats — carefully controlled experiments undercut by proprietary training data, small evaluation sets, and no statistical significance testing, yielding reproducibility readiness of only medium and low. This is the profile's clearest open improvement area.

Timeline

2026Publishes “Can One Adapted Model Do It All?”, a controlled comparison of fine-tuning strategies across 13 customer-support LLMs, as last author.
2026Publishes “When Better Turns Do Not Make Better Agents”, showing supervised fine-tuning lifts next-turn metrics while holistic workflow success remains 0/77.
2026Both cataloged papers appear on arXiv within the same year, consolidating a focus on LLM adaptation and agent evaluation for support workflows.

Papers in the Evolve catalog

Frequent co-authors

Xue-Yong Fu×2 Md Tahmid Rahman Laskar×1 Gundeep Singh×1 Karol Chang×1 Kevin Sanders×1 Shi Zong×1 Tania Habib×1 Julien Bouvier Tremblay×1 Shayna Gardiner×1 Harsh Saini×1 Matthias Lee×1 Elena Khasanova×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric dimensions are computed from a limited OpenAlex-based snapshot; institution, country, and ORCID are unresolved in this record.
Output, impact, and activity dimensions are unscored because publication volume and citation data are not available in the snapshot.
The profile rests on two arXiv papers (both 2026) reviewed by Evolve; other publications may exist outside the cataloged sample.
Identity is name-resolved from public bibliographic data; “Shashi Bhushan TN” has not been independently verified against an ORCID or institutional record.