QuantPi Research Collaboration with Prof Jilles Vreeken

Aggregate benchmarks hide more than they reveal. We published a paper today at the ICML 2026 Workshop AI4GOOD (https://lnkd.in/ek7Mii98), together with Sascha Xu and Jilles Vreeken from CISPA Helmholtz Center for Information Security: "Where Do Agents Differ? Interpretable Rule Discovery for Performance Differences Across Models and Data" The core finding: when you compare two agentic backends on the same benchmark, the aggregate accuracy gap tells you almost nothing about where the difference actually lives. On SWE-bench, we compared five OpenHands backends. Claude Opus and GPT-5 are separated by about 6 points overall. Unremarkable. But within long-trajectory tasks - the kind where an agent has to sustain coherent reasoning across many steps - Opus resolves 90% and GPT-5 resolves 56%. A 34-point gap that the aggregate number buries completely. And the reverse exists too: a narrow task regime where GPT-5 reaches 89% and Opus drops to 36%. This is not a corner case. Across all ten backend pairs we tested, seven contained at least one regime where the weaker backend outperformed the stronger one. What we see with enterprise customers mirrors this exactly. Model selection decisions are being made on aggregate scores over benchmark distributions that don’t reflect the actual task distribution in production. The result is systems that perform well on paper and fail on the specific slice of work that matters. The method we apply - differential subgroup discovery - finds these structured differences automatically and describes them in interpretable rules. Not “this model is better.” But “this model is better for tasks with these specific characteristics. That distinction is what makes evaluation actually useful. Last but not least, HUGE thanks to Sascha Xu and Jilles Vreeken for their impressive work on this. 

Choosing the wrong AI backend for your use case is easier than it looks - aggregate benchmarks won't tell you where it fails. New research from our team and CISPA Helmholtz Center for Information Security, published today at the [ICML] Int'l Conference on Machine Learning Workshop on AI for Good.

Spotlight on recent work with Malak Sadek and echo wan on the human-centred design of conversational agents. It is fantastic to see Malak now leading her own research group at Cambridge, and expanding the work she started during her PhD at Imperial 👏

Really excited to share our papers from the ACM CUI Conference, written at HIVE CAI at the Centre for Human-Inspired Artificial Intelligence (CHIA), University of Cambridge in collaboration with Céline Mougenot (PhD) and echo wan at the Collective Innovation Lab at Imperial College London. 1️⃣ Using Anti-Values as Conversational AI Design Constraints: Deciding on What We Shouldn't Build -- this provocation flips our usual work of considering human values on its head and asks what happens if we consider anti-values instead, and how that can reframe conversational AI design temporally and relationally: https://lnkd.in/egd5S8Jz ✨ My favourite thing about this work is that it leverages a number of interdisciplinary lenses to consider how we can look at anti-values as helpful constraints: from dark patterns at an interactional level and corporate AI governance approaches like impact assessments, to threat modelling practices in security engineering and broader ethical and socio-technical considerations raised across philosophy, cybernetics and STS. 2️⃣ KNIT-C: Exploring Conversational AI Facilitation of Group Convergence Through Computational Boundary Objects -- this paper extends our previous work on developing the KNIT framework for facilitating convergence across interdisciplinary groups with a conversational interface: https://lnkd.in/eMZfgEQC ✨ My favourite part about this work is that the empirical data collection actually came from an outreach activity with the general public at this year's Cambridge Festival. It led to a number of great discussions around the differences and (and surprisingly many) similarities between facilitating convergence and agreement between a group of professionals at a workplace and a group of strangers in a Cambridge classroom. You can read the original work on KNIT here: https://lnkd.in/edNFgDQ7 3️⃣ Whose Values? Demographic Influences on Perceptions of Conversational AI Value Alignment -- this paper explores how demographic factors impact users' perceptions of chatbots' alignment with a set of values their creators intended to embed within them, and with their own values as chatbot users: https://lnkd.in/eyYt_nYE ✨ My favourite thing about this work is how it highlights subtle nuances in how demographics shape users' perceptions. For me, it was surprising to see how 'country of birth' versus 'country of residence' both independently shaped perceptions, which is incredibly sensical given even my own life experiences, but which I'd never translated into factors to consider during conversational AI design before. 🐝 Learn more about our work and research at: https://lnkd.in/ejC9Xzqn #conversationalAI #research #CUI #humancentredAI #HCAI #values #HCI

Heading to Seoul next week for #𝗜𝗖𝗠𝗟𝟮𝟬𝟮𝟲 to present a finding that surprised us: 𝗺𝘂𝗹𝘁𝗶-𝗮𝗴𝗲𝗻𝘁 𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗰𝗮𝗻 𝗯𝗲 𝘄𝗼𝗿𝘀𝗲 𝘁𝗵𝗮𝗻 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗺𝗼𝗱𝗲𝗹 — not just slower or more expensive, but less accurate. We evaluated 15 frontier LLMs and found multi-agent teams reached 𝗼𝗻𝗹𝘆 𝟯𝟬.𝟭% 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 on tasks a single agent with complete information solves at 80.7%. The setup: we adapted the 𝗛𝗶𝗱𝗱𝗲𝗻 𝗣𝗿𝗼𝗳𝗶𝗹𝗲 𝗽𝗮𝗿𝗮𝗱𝗶𝗴𝗺 from social psychology into HiddenBench, a 65-task benchmark. Each agent gets the same shared facts (which point to the wrong answer) plus one unique private fact. The correct decision only emerges when agents pool what they know. This is exactly the scenario multi-agent systems are built for, and exactly where they fail. What we learned: 🔍 𝗧𝗵𝗲 𝗳𝗮𝗶𝗹𝘂𝗿𝗲 𝗶𝘀𝗻'𝘁 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴, 𝗶𝘁'𝘀 𝗮𝘀𝗸𝗶𝗻𝗴. Forcing agents to disclose everything upfront lifted accuracy to 92–98%. Agents integrate information fine once it's on the table; they just never ask what others might know. 𝗧𝗵𝗲𝘆 𝗰𝗼𝗻𝘃𝗲𝗿𝗴𝗲 𝗽𝗼𝗹𝗶𝘁𝗲𝗹𝘆 𝗼𝗻 𝘀𝗵𝗮𝗿𝗲𝗱 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗹𝗼𝗰𝗸 𝗶𝗻 𝘁𝗵𝗲 𝘄𝗿𝗼𝗻𝗴 𝗮𝗻𝘀𝘄𝗲𝗿. Groupthink, but silicon. 📉 𝗦𝗰𝗮𝗹𝗶𝗻𝗴 𝗺𝗮𝗸𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝘀𝗲. At 7 agents, communication improved accuracy by just +0.006. 🧠 𝗠𝗼𝗱𝗲𝗹 𝘀𝗰𝗮𝗹𝗲 𝗱𝗼𝗲𝘀𝗻'𝘁 𝘀𝗮𝘃𝗲 𝘆𝗼𝘂. GPT-5 variants didn't beat GPT-4.1-Mini in collective settings. Collective reasoning is a distinct capability — one we're not training for. ✅ 𝗔𝗻𝗱 𝗶𝘁'𝘀 𝗳𝗶𝘅𝗮𝗯𝗹𝗲. A lightweight structured communication protocol lifted GPT-4.1 from 3.7% to 80% accuracy. No training, just better coordination structure. As multi-agent systems get deployed for diagnosis support, incident response, and enterprise decisions, the assumption is that more agents means more pooled intelligence. Without explicit coordination mechanisms, the opposite can happen. 𝗜𝗳 𝘆𝗼𝘂'𝗿𝗲 𝗮𝘁 𝗜𝗖𝗠𝗟, 𝗰𝗼𝗺𝗲 𝗳𝗶𝗻𝗱 𝘂𝘀: 🗓 Wednesday, July 8 • 1:30–3:15 PM CST, HALL A #4313 📄 Paper: https://lnkd.in/gT9sWTBQ 💻 Code: https://lnkd.in/gvxcGegM Would love to chat and connect! DM open.

Before you let any AI near your callers, ask one question: where does the data actually live? For Starlight, the answer is Canada. Call audio, recordings, transcripts and consent records sit on a Toronto-region server. Consent is captured at the start of every call, CASL and PIPEDA built in, not bolted on. Every call is recorded and transcribed, so you can audit our work yourself. And the part most vendors won't put in writing: we never sell your callers' information, and we never use it to train third-party models. It's in our Data Processing Agreement. Full transparency, only text-stage AI processing crosses the border to generate a response. No audio, no recordings, no consent records leave Canada, and that step is disclosed in the consent on every call. Your callers hand you the worst day of their life. That data deserves better than a shrug. #PersonalInjuryLaw #PIPEDA #DataPrivacy #LegalTech

Most AI assistants know the internet. Your team needs one that knows your company. Upload your documentation once: support guides, policies, whatever your team leans on. Then just ask. Instead of hallucinating, the assistant searches your knowledge base first. It answers from your own documents, source cited. No switching tabs. No "I think the policy was somewhere..." Try it free: modulex.dev

The European Commission has released a report titled "Enhancing Competitiveness, Sovereignty and Security of the European Union in Frontier AI", capturing key insights from the first meeting of the European Expert Forum on Frontier AI. Experts noted that within roughly three years, frontier models moved from struggling with basic mathematics and elementary programming to matching or surpassing human experts across many reasoning, software engineering and scientific tasks. Offensive cyber capability was singled out as improving particularly fast. Within the next decade or earlier, models may match or exceed human performance across all cognitive tasks, though uncertainty remains. On markets, revenue at the frontier accrues almost entirely to a handful of non European developers, with the two leading firms reaching reported annualised revenue of roughly USD 60 billion in spring 2026. European frontier developers were put at around one percent of that figure, and European AI venture funding at around six percent of the global total. The Union accounts for around 15 percent of global GDP but hosts an estimated 5 percent of global compute. Experts identified four structural barriers: computing infrastructure and energy as an urgent two year priority, growth stage capital as the binding financial constraint, legal clarity on copyright and data protection for training data, and talent, described as both a principal strength and a pressing vulnerability. Experts also recommended selectively backing high return research bets across the current paradigm, alternative approaches such as world models and physical AI, and layers like secure and verifiable AI infrastructure. Instruments discussed included a CERN for AI and a DARPA style funding vehicle. Whichever path prevails, experts stressed the EU must strengthen its sovereign ability to access, choose, control and benefit from any frontier AI model, pointing to Europe's dominant position in advanced semiconductor manufacturing equipment as immediate leverage, and proposing a coalition with trusted partners including the United Kingdom, Canada, Japan, the Republic of Korea, Australia, New Zealand and India. Experts called for frontier AI to become a foremost political priority, with dedicated foresight functions, emergency response structures and ambitious 2030 targets. The next one to two years, they said, could prove decisive. #FrontierAI #EuropeanUnion #AIPolicy #AIGovernance #TechSovereignty #ArtificialIntelligence #EUAIAct #TechLaw P.S. This is for academic discussion only.

❔ Most life science content is now written with AI. But is anyone asking scientists what they actually think about that? Co-Labb is. We've built a short independent survey, just a few minutes, asking researchers how their habits have changed, what they trust, and what they wish vendors would write about but never do. The findings will be published as a free report later this year. Complete it, and you're in with a chance of winning 1 of 5 × $75 prepaid Visa cards. https://lnkd.in/dWbKjhk4

The communications industry is operating under axioms it hasn't examined in thirty years: networks are infrastructure, intelligence lives at the endpoints, value accrues at the application layer. Thomas Kuhn's observation in his 1962 classic, “The Structure of Scientific Revolutions” was that paradigms don't fail loudly, they produce anomalies that get explained away one by one, until the weight becomes undeniable. The anomalies are arriving now. These aren't product questions. They're paradigm questions. The map is no longer describing the territory. What makes this moment harder to read is that the more significant development isn't competitive displacement, it's ontological emergence. It is not a better version of the existing model, but a new category forming that has no prior address and no incumbent to displace. When intelligence stops being a workload running over a network and becomes constitutive of the network itself, the resulting entity isn't a smarter pipe or a more connected AI system. It's new ground. The island is already above the waterline. Most of the industry hasn't looked up yet. That gap between the paradigm cracking and the new terrain being formed is the conversation we're building with industry leaders at Quantumverse in Seattle this September. Not the consensus view of what AI does to networks or the enterprise, but the harder questions: what axioms are you not examining, what anomalies are you explaining away, and what's forming in the space your current framework can't see? Our speakers are globally respected, genuinely candid. We ask the tough questions so your strategy can have tangible data points. Hope to see you in September.