Vineet Kumar

Indian Institute of Technology Madras · India

Unclaimed · AI-structured profileOpenAlex

NLP researcher at IIT Madras working on generative retrieval and knowledge injection for instruction-tuned LLMs

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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—

Not yet scored: the snapshot contains no reliable productivity counts beyond three 2026 arXiv works, and OpenAlex reports h-index and i10-index of 0. This axis would measure total publications, venue tiers, and yearly output volume.

Impact—

Not yet scored: no citation counts, citing works, or venue-level influence are available in this OpenAlex snapshot (h-index 0, i10-index 0). This axis would measure citations, downstream adoption, and reference influence of the work.

Focus0/100

The radar assigns focus a score of 0, the lowest on this profile, and the raw field distribution explains why: work splits exactly 0.5 to information retrieval and 0.5 to LLM adaptation and knowledge injection. OpenAlex topics add breadth — Topic Modeling (3 works), Information Retrieval and Search Behavior (2), Multimodal Machine Learning Applications (2), Natural Language Processing Techniques (1), and Domain Adaptation and Few-Shot Learning (1) — indicating a two-track NLP agenda rather than concentration.

Collaboration50/100

Collaboration scores a mid-range 50. The snapshot records 11 distinct coauthors, matching the frequent-collaborator list: Meghanadh Pulivarthi, Jaydeep Sen, Sachindra Joshi, Kushagra Bhushan, Gaurav Pandey, Yatin Nandwani, Dinesh Raghu, Sai Krishna Reddy Sathi, Sonam Gupta, Vishwajeet Kumar, and Riyaz Ahmad Bhat. Authorship placement varies — first author on STAIR but middle author on DKL and KItCAT — suggesting both led and contributed roles within a sizable team.

Activity40/100

Activity scores 40, a modest reading consistent with the raw counts: recent3yWorks is 3 and the countsByYear series contains a single entry, 2026 with 3 works, with no earlier years in the snapshot. All three catalog papers appeared as 2026 arXiv preprints, so the visible publication cadence is concentrated entirely in one year, and longer-run throughput cannot yet be assessed from this data.

Reproducibility55/100

Reproducibility scores 55, computed as the mean readiness across 2 reviewed papers (reproMean 55, reproPapers 2). Both STAIR and DKL carry a medium reproducibility rating in Evolve's review: each introduces new datasets or merging procedures whose claims — such as STAIR's 82.6 vs 76.9 R@1 over DSI — depend on released data and decoding details that Evolve flags as claim-tempering caveats.

Research summary

Vineet Kumar is a natural language processing researcher affiliated with the Indian Institute of Technology Madras whose work centers on two connected problems: making generative retrieval faithful to document structure, and injecting external knowledge into instruction-tuned language models. The Evolve catalog lists three 2026 arXiv papers. As first author of STAIR, Kumar makes a book's table of contents the retrieval unit of a differentiable search index, reporting 82.6 versus 76.9 Recall@1 against DSI, though Evolve notes synthetic queries and constrained decoding temper the claims. As a middle author on DKL and KItCAT, Kumar contributes model-merging and input-corruption approaches to knowledge injection; Evolve finds DKL's data-efficiency and QA gains strong but its lead over the best RAG baselines often within noise. OpenAlex topic data mirrors this split — Topic Modeling (3 works), Information Retrieval and Search Behavior (2), Multimodal Machine Learning Applications (2), Natural Language Processing Techniques (1), and Domain Adaptation and Few-Shot Learning (1) — and the field distribution divides evenly, 0.5 to LLM adaptation and knowledge injection and 0.5 to information retrieval. With 11 distinct coauthors, and h-index and i10-index both at 0, this is an early-stage profile whose impact cannot yet be scored.

Collaboration pattern

Kumar's OpenAlex snapshot lists 11 distinct coauthors, all recurring in the frequent-collaborator list — including Meghanadh Pulivarthi, Jaydeep Sen, and Sachindra Joshi — indicating a stable, team-based working pattern spanning both information retrieval and LLM adaptation projects. The authorship record shows a first-author lead on STAIR alongside middle-author contributions on DKL and KItCAT, consistent with working within a larger NLP group whose members recur across all three 2026 papers.

Research keywords

retrieval-augmented generationdifferentiable search indexparametric memorylong-context retrievalknowledge injectionmodel mergingtask arithmeticinstruction tuningextended pre-trainingtopic modeling

Highlights

STAIR: table-of-contents units beat DSI on generative retrieval
As first author, Kumar's STAIR makes a book's table of contents the retrieval unit of a generative retriever, reporting 82.6 vs 76.9 Recall@1 over DSI; Evolve notes synthetic queries and constrained decoding temper the claims. Reproducibility: medium.
DKL merges base-model knowledge into instruction-tuned LLMs
A middle-author contribution, DKL merges a base-LLM knowledge vector into an instruction-tuned model; Evolve finds the data-efficiency and QA gains strong, though leads over the best RAG baselines are often within noise. Reproducibility: medium.
KItCAT: knowledge injection via input corruption
KItCAT (2026, middle author) extends the knowledge-injection line with an input-corruption approach for auto-regressive training; it is listed in the catalog without a full Evolve review yet, so field, keywords, and reproducibility remain pending.

Timeline

2026Published STAIR (STructure Aware Information Retriever) on arXiv as first author, introducing a ToC-based dataset and LLM retriever for document structure augmentation.
2026Co-authored DKL: Decoupled Knowledge Learning for Instruction-Tuned Language Models, merging base-LLM knowledge vectors into instruct models.
2026Co-authored KItCAT: Knowledge Injection via Input Corruption for Auto-regressive Training, extending the knowledge-injection research line.

Papers in the Evolve catalog

2026-09-03first author
STAIR (STructure Aware Information Retriever): A novel dataset and LLM based retriever for document structure augmentation
2026-09-02
DKL: Decoupled Knowledge Learning for Instruction-Tuned Language Models

Frequent co-authors

Meghanadh Pulivarthi×3 Jaydeep Sen×3 Sachindra Joshi×3 Kushagra Bhushan×2 Gaurav Pandey×2 Yatin Nandwani×2 Dinesh Raghu×2 Sai Krishna Reddy Sathi×1 Sonam Gupta×1 vishwajeet kumar×1 Riyaz Ahmad Bhat×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.

Bibliometrics come from an OpenAlex snapshot only (identity: openalex); reported h-index and i10-index are both 0, and no citation or venue data is in scope.
'Vineet Kumar' is a common name; this profile is name-resolved from public bibliographic data (OpenAlex) linked to the IIT Madras affiliation, and no ORCID is available for disambiguation.
All three catalog papers are 2026 arXiv preprints; the KItCAT paper has no Evolve review yet (field, verdict, keywords, and reproducibility pending), so assessments rest on two reviewed papers.
Counts by year cover 2026 only; longer-run output, funding, awards, and career history are outside the provided data.