LILT AI

The energy in the room at Beyond the Slop: Global Growth Meet Up in London earlier this week was something else. Angus Cormie from Lenovo Cloud and Software and Xiao Xiao from Mintel shared candid insights on global strategy and how they're using AI to drive real efficiencies. Plus, a global wine tasting to match. The next Global Growth Meet Up is coming to a city near you watch this space. 🌍

We are releasing GAIA-v2-LILT, a re-audited multilingual agent benchmark across Arabic, German, Hindi, Korean, and Portuguese (BR). 165 tasks per language. Built on top of MAPS-GAIA, with three layers of validation added beyond standard MT post-editing: functional alignment, cultural alignment, and difficulty calibration. Reviewed via deterministic filters, granular LLM judges, and specialized human audit with meta-review. 💡 The headline result: measured agent performance jumped an average of +20.7 percentage points after correction. Korean alone gained +28.3 points. Those gains are not from better models. They reflect benchmark measurement error in the original translation pipeline. An uncorrected MT-then-post-edit workflow systematically underestimates how well frontier models perform in non-English contexts, because the benchmark itself penalizes agents for translation artifacts rather than reasoning errors. What we found mattered most: ✅ Functional alignment failures (an answer key in English notation rejecting a correct German answer, for example) drove most of the result-flipping cases. ✅ Cultural anchoring (US-specific transport laws, imperial units, US holiday calendars) shifted task difficulty without surface signal. ✅ Translationese was common but rarely the cause of failure. It slows agents down. It does not break tasks. Word-level edit rates ran from 25 to 55 percent. Hindi required revision to every single task. This is not light-touch cleanup. If you are building English-first models and reporting multilingual numbers, this is worth a look. Translation quality and evaluation validity are the same problem. Read the full report: https://lnkd.in/gy-vxJN4 Dataset: https://lnkd.in/gUQ2x-ZK Technical report: https://lnkd.in/gpK8gvab #multilingual #benchmarks #appliedAI #evaluation

If localization is your final step, you’ve already missed the window. 🌐 Native Intelligence | Edition 04 In enterprise tech, we have been conditioned to treat localization as a post-development task. It is often something you apply to a finished product right before the big launch. That strategy is failing in the age of AI. Because AI is more deeply dependent on language than any software before it, multilinguality must be Core Infrastructure. This means making architectural decisions about regional variation and training data before the first line of code is written. The companies that will define the next decade are not necessarily those with the most features. Instead, they are the ones with the deepest cultural reach. I am curious to see where most organizations actually stand today. Poll Question: When does your team typically start thinking about localization? A) Design Phase B) Development C) Post-Launch (Share your thoughts in the comments on why your team chose that timing!)

We're at DMWF next week, 6th & 7th May. You'll hear a lot about AI in marketing. Less about whether your brand actually survives at scale across languages and cultures. That's what we'll be focused on. Rob Cross is on the panel "Mastering Marketing in the Age of Intelligent Automation" — Wednesday 6th May, 2:30–3:10pm - not one to miss! 📍 Come say hi at booth #190 🚀 #DMWF #DMWFGlobal #DMWF2026 #DigitalMarketing

LILT is headed to Web Summit Vancouver! 🇨🇦 Ready to Make Anything Multilingual? Join the LILT team from May 11-14 to discover how our Agentic AI helps brands scale globally with unprecedented quality and speed. Visit us at Stand #E364 to learn how you can: 🧿 Ship 3-5x faster. 🧿 Reduce multilingual costs by 50%. 🧿 Deliver human-verified, multi-modal communications across text, image, and video. 🎤 Don’t Miss Our CEO's Session: Catch our CEO, Spence Green, on May 12 as he discusses the next frontier of innovation: "From AI assistants to AI coworkers: what's next?" See you in Vancouver! 🗓️ #WebSummit #WebSummitVancouver #LILT#AgenticAI #GlobalScale #AIInnovation #WebSummitVancouver2026

Last few spots remaining. On Tuesday 28th April, we're bringing together senior marketing and localisation leaders in London for an evening that goes Beyond the Slop. Here's what's on: - A panel discussion on The AI Shift in Multilingual Translation — where AI accelerates the lifecycle and where the human touch remains your greatest competitive advantage - A guided global wine tasting led by a professional sommelier - Canapés & networking with global growth leaders from some of London's most exciting companies Tuesday 28th April | 5:30pm – 8:30pm If you're working across markets and looking to grow, this one's for you. 👉 Request your spot here: https://luma.com/i3xugwxj #GlobalGrowth #Localisation #AI #MultilingualMarketing

Stop paying the "fragmentation tax" on your global content strategy. Most enterprises believe they are checking the "AI box" by using legacy platforms with bolted-on features. In reality, they are inheriting a patchwork of disconnected tools that create data silos, static learning, and a massive drain on engineering velocity. The distinction is simple: Is your AI built-in or just bolted-on? For leaders tasked with delivering a strategic AI Win this year, the architecture of your multilingual content stack is the difference between an appreciating technology asset and a growing liability. We have distilled the technical and operational requirements of true AI-native scale into a single framework: The AI-Native Buyer’s Checklist. Use this checklist to audit your stack for: ▪️ Real-Time Adaptation: Does your model learn from feedback instantly, or are you waiting for periodic retraining cycles? ▪️ Infrastructure Security: Is your AI built into the core data layer to protect proprietary assets? ▪️ Operational ROI: Can you prove measurable efficiency gains to the C-suite with hard data? Managing multilingual content should be an autonomous product feature, not a resource-draining bottleneck. Download the Checklist here: https://lnkd.in/gvVisgHt #AINative #MultilingualContent #GenerativeAI #ProductVelocity #LILT

Big news! 🚀 We've launched LILT Assist on Google Cloud Marketplace to help drive real-world impact. Built for Gemini Enterprise, our agent serves as an autonomous operator, managing end-to-end workflows and scaling global production of enterprise multilingual content. Huge thanks to the Google Cloud team as we head into Google Cloud Next 2026. Read the full blog to see our solution in action: https://lnkd.in/gKhEWaTj #GeminiEnterprise #GoogleCloudPartner #AI

Spring events are in full swing, today and tomorrow grab a coffee with Ronan Trayer and Tiago Gouveia Segurado at the B2B eCommerce World event in London. Come and see how we can help scale your manufacturing eCommerce globally - so you can reach new markets in days, not weeks with LILT.

One "stilted" translation is all it takes to lose a customer for good. 🌐 Native Intelligence | Edition 03 Trust is an asymmetric variable. It takes months to build and seconds to evaporate. In the AI world, we often settle for "good enough" translation. But there is a massive gulf between a translated interface and a localized one: ▪️ Translation reproduces content in another language. ▪️ Localization makes that content feel like it belongs there. If a user encounters an AI response that feels "off," the damage is immediate. Whether the tone is too formal, the idiom is wrong, or the cultural reference is mismatched, they don't just see a bug. They see a reason not to return. In new markets where AI adoption is still forming, you don’t get a second chance to make a first impression. The Strategy: Treat language as a substance, not a delivery mechanism. If it doesn't sound human, it won't feel credible. Have you ever abandoned a "global" app because the local experience felt like an afterthought?👇