Atul Anand

Research profile

Research summary

Atul Anand is a researcher at the University of Edinburgh whose OpenAlex-linked profile centres on natural language processing: topic modelling (5 works), biomedical text mining and ontologies (5), and machine learning in healthcare (5) form the largest clusters, with natural language processing techniques (3) and information retrieval and search behaviour (2) alongside. His newest catalogued output evaluates LLM behaviour in software engineering. In the 2026 arXiv paper 'Impact Is Not Invalidation: Ask About the Claim, Not the Diff' — first-authored and corresponding — an execution-based benchmark shows that diff-level LLM judging fails to detect assertion invalidation while claim-level questioning succeeds; Evolve's review finds the logic rigorous with only minor internal inconsistencies. A second 2026 first-authored arXiv paper reports that sequential knowledge editing breaks a model's ability to tell good evidence from bad without costing accuracy. Bibliometrically the profile is early-stage: h-index 1, i10-index 0, 11 works over the last three years, and annual counts rising to 7 in 2026. Recorded co-authorship is empty — 0 distinct coauthors — which likely reflects disambiguation limits given his first-author position on both papers. Reproducibility evidence covers one paper (mean 55, medium readiness). Output and impact axes lack sufficient data to score.

Research keywords

Highlights

• Claim-level questioning beats diff-level LLM judging for detecting assertion invalidation

  In the 2026 arXiv paper 'Impact Is Not Invalidation: Ask About the Claim, Not the Diff', Anand (first and corresponding author) builds an execution-based benchmark showing that LLM judges asked about a diff fail to detect when changes invalidate assertions, while asking about the claim itself succeeds. Evolve's review finds the logic rigorous with only minor internal inconsistencies.

• Sequential knowledge editing erodes evidence discrimination without costing accuracy

  A second 2026 first-authored, corresponding arXiv paper reports that applying knowledge edits sequentially breaks a model's ability to tell good evidence from bad even though task accuracy is preserved — an evaluation-relevant failure mode invisible to accuracy-only testing. This paper is not yet reviewed by Evolve and has no readiness assessment.

• A shared agenda: test what the model claims, not just what it outputs

  Across both 2026 papers, the first-authored work converges on checking LLM systems against their own claims — whether a code change still satisfies its assertions, and whether an edited model still reasons over valid evidence. Together they sketch an evaluation agenda for LLM behaviour that goes beyond accuracy metrics.

Timeline

1. 2026Published 'Impact Is Not Invalidation: Ask About the Claim, Not the Diff' on arXiv as first and corresponding author — an execution-based benchmark of LLM judgement for assertion invalidation in code changes.
2. 2026Published a first-authored, corresponding arXiv study showing that sequential knowledge editing degrades a model's ability to discriminate good evidence from bad without costing accuracy.
3. 2026Both catalogued papers land in the same year, concentrating on evaluation of LLM behaviour (code-change impact and post-editing reasoning), matching the record-high 7 works counted for 2026.

Papers in the Evolve catalog

• 2026-09-20first authorcorresponding

  Impact Is Not Invalidation: Ask About the Claim, Not the Diff