4/5 With Matteo Attimonelli, @aryopg.bsky.social, @rohit-saxena.bsky.social, @dietmarjannach.bsky.social & others: 32.2%-83.6% of queries across four image retrieval benchmarks admit single-modality shortcuts. CIRCUS tests composition on a validated subset. https://arxiv.org/a…
Who's Who of AI
Pasquale Minervini
Why they matter
Researcher with public evidence across Models & releases, AI business, AI research.
- AI signals
- 3 past 30d
- Sources
- 1 distinct domains
- Discussions
- 0 past 30d
- Latest signal
- 6d ago
Researcher in ML/NLP at the University of Edinburgh (faculty at Informatics and EdinburghNLP), Co-Founder/CTO at www.miniml.ai, ELLIS (@ELLIS.eu) Scholar, Generative AI Lab (GAIL, https://gail.ed.ac.uk/) Fellow -- www.neuralnoise.com, he/they
What they're sharing
Do Composed Image Retrieval Benchmarks Require Multimodal Composition? arxiv.org
Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads arxiv.org
SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks arxiv.org
The 20th Conference of the European Chapter of the Association for Computational LinguisticsAthens, GreeceMarch 9-14, 2027 2027.eacl.org
{ICML, ACL} 2026 neuralnoise.com
Articles & links
3/5 Which heads retrieve answers that aren't literally in the context? LOCOS, led by @aryopg.bsky.social, with @beatricealex.bsky.social, measures what each head writes toward the answer token, rather than just where it attends. https://arxiv.org/abs/2607.01002
AI Weekly's analysis
→
- LOCOS scores each attention head by projecting its OV-circuit output onto the answer-token unembedding direction in a single forward pass.
- On Qwen3-8B, ablating the top 50 LOCOS heads drove NoLiMa ROUGE-L from 0.401 to 0.000; the strongest baseline still retained 0.292.
- The same ablation dropped MuSiQue from 0.55 to 0.08 and BABI-Long from 0.62 to 0.20 while parametric recall and arithmetic stayed at baseline.
Read full analysis →
2/5 SCOPE, led by Wai-Chung Kwan, with @aryopg.bsky.social: a Challenger creates document-grounded tasks, a Solver learns through retrieval, and a frozen self-judge rewards answers. Up to +10.4 points on open-ended benchmarks, without curated prompts. https://arxiv.org/abs/260…
Organisations -- sponsorship opportunities are available for #EACL2027, the flagship European conference in computational linguistics, taking place in Athens in March 2027! Support the NLP community and connect with researchers and practitioners: 2027.eacl.org
New blog post on our {ICML, ACL} 2026 papers, plus some new interesting results on open-ended learning and multi-modal retrieval! neuralnoise.com/2026/icml-ac...
Their own posts
Recent commentary
1/5 My colleagues and collaborators will present a few papers at NeurIPS later this year! 🚀 Self-play for open-ended tasks, non-literal retrieval heads, shortcuts in multimodal benchmarks, and a psychophysics-based approach to AI evaluation. A quick tour below. Blog post underway!
Their network
In Pasquale Minervini's orbit
Center = Pasquale Minervini. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Pasquale Minervini? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/neuralnoise-com)