Archana Vaidheeswaran
I study how to check AI systems. My papers ask whether a model's reasoning holds together, and whether a model used as a judge actually checks what it is asked to check.
I write most of my papers with the early-career researchers I mentor. I am Program Director at Algoverse, an AI safety research fellowship, and I started PRISM, a peer-vetted research fellowship. From 2017 to 2022 I worked in industry as an AI research engineer and data scientist.
I served on the board of Women in Machine Learning (WiML) from 2023 to 2026. Before that I organised the WiML Un-Workshop at ICML 2021 as Virtual Experience Chair.
I have a Master of Engineering by research in Electrical and Computer Engineering from the National University of Singapore, and a Bachelor's in Electrical Engineering from SRM University, Chennai.
- EMNLP '26CAST, on settling contradictions in clinical LLMs, accepted to the Industry Track.
- KDD '26A negative result on logical consistency at the SeT-LLM workshop.
- ICML '26Three workshop papers, one as first author.
- ICLR '26Cross-lingual fairness drift in moral reasoning at the CAO workshop.
- EACL '26FluffInjector at the Student Research Workshop.
- AAAI '26Panellist, Industry Roundtable, AI Governance Workshop, Singapore.
-
ICML 2026FAGEN workshop
An audit of an LLM that explains an autonomous agent's decisions on German energy grid data. None of 30 explanations flagged an injected 600 MW false reading. All three models justified wrong actions 80–95% of the time.
First authorOpenReview -
EACL 2026Student Research Workshop
A benchmark where 25–40% of reasoning steps are swapped for plausible filler. Frontier judges accept these chains far too often. SmartRM, the verifier we trained on it, cuts false positives from 37.4% to 2.7%.
41.6% acceptanceOpenReview -
WACV 2026VALED workshop
Vision-language models proposed for copyright disputes follow the prompt over the image. Primed with a legal accusation, they rate unrelated images as similar. We measure this with a Blindness Index.
-
EMNLP 2026Industry TrackCAST: Contrastive Adjudication via Selective Triggering for Clinical LLMs
When a clinical LLM gives different answers to the same question, CAST puts the competing arguments side by side and asks the model to settle them. On PubMedQA with GPT-4o-mini, accuracy on those questions rises from 28.0% to 57.0%.
Poster -
ICML 2026FoGen workshop
We treat hidden reasoning tokens as a path through representation space. CODI settles like a stable attractor. COCONUT expands like an unstable system.
-
KDD 2026SeT-LLM workshop
We tested whether a layer-level regulariser stops QA models giving the same answer to a question and its negation. Across three backbones and five seeds the gains were small and unstable. Training directly on negated examples works better.
-
ICML 2026Forecasting workshop
Prediction markets work when traders think independently. Preference-optimised LLM traders drift toward a shared prior, and the market prices confidently and wrongly.
-
ICLR 2026CAO workshop
Five models, 50 moral dilemmas and four languages: English, Spanish, Korean and Mandarin. Consequentialist bias grows outside English, and cultural grounding drops by up to 88%.
-
2025Preprint
A market maker as a coordination mechanism that keeps a population of LLM agents calibrated as it grows.
-
2026Preprint
A dataset of realistic, messy queries for testing whether agents pick the right Model Context Protocol tool.
-
ACL 2019BioNLP workshop
BioBERT for medical natural language inference, with attention visualisation so clinicians can see which evidence drove each prediction. 83.45% on MedNLI.
-
PRISMStarted 2025
A 16-week AI safety research fellowship I started. Fellows review each other's plans, so mentor time goes to four decision points. The 2026 cohort drew 795 applications and has 48 fellows in 12 research groups.
-
SOAR 2026EleutherAIA selection rubric that checks claims before scoring them
I wrote it for EleutherAI's Summer of Open AI Research after the first version favoured polished writing. Each claim in an application is checked before it is scored.
-
AlgoverseSince 2025Running research cohorts at scale
I direct Algoverse's AI safety research fellowship, funded by Open Philanthropy. I run selection and paper review for its early-career research teams.
Teaching and service
- Instructor on four LinkedIn Learning courses and Udacity's Intel Edge AI for IoT Developers Nanodegree, with 40,000+ learners.
- Board Director, Women in Machine Learning (WiML), 2023–2026.
- Virtual Experience Chair, WiML Un-Workshop at ICML 2021.
- Panellist, Industry Roundtable, AI Governance Workshop at AAAI-26.
- Author of the Humans of AI Safety interview series, 2024.
- Side project: The Museum of AI Safety, a browser museum with 35 exhibits.