Edinburgh team publishes 10 papers on self-play, multimodal retrieval, VLM robustness

PMinervini · x · 2026-10-27

Pasquale Minervini's group (University of Edinburgh / Miniml.AI) summarized 10 accepted papers at NeurIPS 2026, EMNLP 2026 and LoG.

• NeurIPS: SCOPE co-evolves a Challenger and Solver via self-play over unlabelled documents, with a frozen model as rubric-writing judge, improving Qwen2.5-7B across eight open-ended benchmarks; plus LOCS, composed image retrieval, and AI psychophysics.
• EMNLP: multimodal retrieval heads, VLM instability under image corruption, and interference-aware composition of steering vectors (Findings).
• Preprints: SPSD, FACE-Eval, VLM frequency filtering, and inverse VLM scaling (when larger VLMs perform worse).

Original post →

More from Models

• Tagging @bot on X automatically forwards your post to Grok Bot — EricBuess · 2026-10-27
• Smaug Humanizer goes open-weights: de-AI-slop your text, built on Gemma — bindureddy · 2026-10-27
• Model Lineups Defended: 'One Benchmark Score Can't Capture a Model's Utility' — craigsdennis · 2026-10-27
• Claude research model claims to refute Exponential Time Hypothesis with 14M-token autonomous run — scaling01 · 2026-10-27
• Kimi K3 on AMD MI355X: +38% agentic throughput, +86% peak vs NVIDIA B300 — YouJiacheng · 2026-10-27
• Gemini turns out notably negative on Anthropic's IPO, ex-OpenAI exec jokes — Miles_Brundage · 2026-10-27

Models channel →