Nikhil Reddy Pottanigari

Research profile

Research summary

Nikhil Reddy Pottanigari is an early-stage natural language processing researcher whose visible record — three first-authored arXiv preprints, all dated 2026 — clusters around document AI and evaluation methodology. The first paper asks whether text, page images, or combined inputs serve multimodal document question answering best, reporting that page images can win on accuracy within a 50-page cap while latency grows; Evolve's review calls it controlled and honest but notes the absence of significance testing. The second, MSBD, batches page-boundary decisions into a single LLM call for page stream segmentation, holding accuracy at a model-specific knee to cut inference cost. The third documents LLM-judge saturation in summarization evaluation and proposes semantic scaffold metrics (FPS/QPS/EPS) that preserve score spread, though validity against human preference remains untested. Together the portfolio shows a consistent agenda: making LLM-based document processing and evaluation cheaper, and scrutinizing how evaluation signals behave. Radar scores reflect this profile: focus is 0, with the field distribution split evenly across three adjacent NLP/document-AI subfields; collaboration is 39, built on 6 distinct coauthors; reproducibility averages 55 across 3 papers rated medium. Output, impact, and activity axes lack sufficient data. No institution, country, or ORCID is attached; the profile is name-resolved from public bibliographic data.

Research keywords

Highlights

• Page images beat text on document QA — within a 50-page cap

  First-authored evaluation of text, pixel, and combined input representations for multimodal document QA. Evolve's verdict: page images can lead on accuracy inside a 50-page cap, at a growing latency cost — controlled, honest work, though without significance testing.

• MSBD: one LLM call for many page-boundary decisions

  Introduces Multi-Split Boundary Decisions for LLM-based page stream segmentation, batching boundary decisions into a single call and holding accuracy at a model-specific knee. Plausible cost savings, but evidence is single-run with no significance checks.

• Semantic scaffolds that stay informative after LLM-judge saturation

  Documents saturation in LLM-as-a-judge summarization evaluation and proposes scaffold metrics (FPS/QPS/EPS) that preserve score spread. Evolve views them as plausible diagnostics, but validity against human preference is untested.

Timeline

1. 2026First-authored arXiv preprint evaluates text, pixel, and hybrid inputs for multimodal document question answering.
2. 2026Proposes MSBD, batching page-boundary decisions into one LLM call to cut the cost of page stream segmentation.
3. 2026Documents LLM-judge saturation in summarization evaluation and introduces semantic scaffold metrics FPS/QPS/EPS.

Papers in the Evolve catalog

• 2026-09-18first author

  Text, Pixels, or Both? Evaluating Input Representations for Multimodal Document QA

• 2026-09-18first author

  Splitting Documents at Lower Cost: Multi-Split Boundary Decisions for LLM-Based Page Stream Segmentation

• 2026-09-18first author

  Preserving What Matters: Semantic Scaffolds Beyond Saturation in Summarization Evaluation

Frequent co-authors