Sachindra Joshi

Research profile

Research summary

Sachindra Joshi is a natural language processing researcher at Indian Institute of Technology Madras whose recent work splits evenly, per the OpenAlex field distribution, between two clusters: NLP (LLM adaptation and knowledge injection) and NLP (information retrieval), at 0.5 each. The OpenAlex-sourced record lists an h-index of 20 and an i10-index of 38, with 22 works in the last three years and 16 distinct recent co-authors. A per-year count series reaching back to 2001 shows a sustained career — peaking at 18 works in 2021 — that historically spans topic modeling (65 works), speech and dialogue systems (47), and NLP techniques (45), before a clear pivot toward LLM-era retrieval. Five 2026 arXiv preprints in the Evolve catalog define the current agenda: STAIR makes a book's table of contents the retrieval unit of a generative retriever, beating DSI at 82.6 vs 76.9 R@1; DKL merges a base-LLM knowledge vector into an instruction-tuned model with strong data-efficiency and QA gains; KItCAT, SearchWiki, and Spatial Matryoshka Training extend knowledge injection, active information seeking, and visual document retrieval. Joshi appears as last author on three of the five catalog papers, consistent with a senior, supervisory role on small teams.

Research keywords

Highlights

• STAIR: table of contents as the unit of generative retrieval

  The 2026 preprint introduces a structure-aware dataset and LLM-based retriever that treats a book's table of contents as the retrieval unit, beating DSI with 82.6 vs 76.9 R@1; Evolve notes that synthetic queries and constrained decoding temper the claims. Reproducibility readiness: medium.

• DKL: merging base-model knowledge into instruction-tuned LLMs

  Decoupled Knowledge Learning merges a base-LLM knowledge vector into an instruct LLM; Evolve finds the data-efficiency and question-answering gains strong, but the RAG lead over the best baselines often within noise. Reproducibility readiness: medium.

• KItCAT: knowledge injection via input corruption (2026 preprint)

  A 2026 arXiv preprint on auto-regressive training with knowledge injection via input corruption. It is listed in the Evolve catalog without a full review, keywords, or a reproducibility rating yet, so it is described at title level only.

• SearchWiki: building and navigating knowledge wikis for information seeking

  A 2026 arXiv preprint, with Joshi in the last author position, on learning to build and navigate knowledge wikis for active information seeking. It has not yet been reviewed by Evolve and is described from its title only.

• Spatial Matryoshka Training for multi-granularity visual document retrieval

  A 2026 arXiv preprint, again with Joshi as last author, targeting multi-granularity visual document retrieval. It carries no Evolve verdict, keywords, or reproducibility rating yet; the description is title-level only.

Timeline

1. 2001Earliest year appearing in the per-year publication counts, with 2 works recorded.
2. 2021Most productive single year in the record, with 18 publications.
3. 2024–2025Recent two-year output of 7 works (2024) and 4 works (2025), within a 22-work count over the last three years.
4. 2026Five arXiv preprints enter the Evolve catalog — STAIR, DKL, KItCAT, SearchWiki, and Spatial Matryoshka Training — with Joshi as last author on three; 11 works counted so far this year.

Papers in the Evolve catalog

• 2026-09-03

  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