New paper on Sims (2026) A multi-context memory architecture for precision-aware personalization in language models. We argue that better personalization needs structured context, not just longer memory. Paper: chiragshah.org/papers/Sims_f… #LLM #AI #Personalization #MemorySystems

Something from last year's SIGIR that seems to resonate more with #AgenticAI as we go beyond #GenerativeAI: Shah & White, From To-Do to Ta-Da: Transforming Task-Focused IR with Generative AI. dl.acm.org/doi/pdf/10.1145/3…

Revisiting our IEEE Computer paper on agents that seems to be published ages ago! Shah & White (2025), Agents are not enough We discuss why robust AI ecosystems need more than standalone agents. ieeexplore.ieee.org/stamp/st… #AI #AgenticAI #IEEEComputer #ResponsibleAI

Introducing SimBench (2026): A benchmark for preference-conditioned agentic planning. Designed to evaluate whether agents can produce correct plans for the same task under different user preferences and constraints. Repo: github.com/VersarAI/SimBench #Benchmark #AgenticAI #Evaluation

New paper alert: SimGuide (2026) Procedurally grounded multi-context representations for personalized agent planning. This work focuses on making agent decisions more accurate across competing goals and constraints. Paper: chiragshah.org/papers/SimGui… #AI #AgenticAI #Personalization

Flat user profiles are a shortcut that breaks at scale. SimBench tests agents on users with layered, conflicting preferences — work vs. health vs. family — and scores whether the agent resolves those conflicts correctly. Open benchmark. Runs today. github.com/versarai/simbench

Still relying on context and memory to achieve personalization in your AI systems? Find out why that's not effective or efficient and get solutions for a privacy-driven personalization using Sims! Read more in our Velocity newsletter: versarai.beehiiv.com/p/versa…

We just shipped Velocity #001 — our take on why context windows and memory aren't enough for real personalization, plus SimBench 1.0. Subscribe: versarai.beehiiv.com

AI agents don't fail because they can't plan.They fail because they plan for nobody in particular. SimBench measures whether your agent can tell Jordan from Rafael — and build a different plan for each. github.com/versarai/simbench

Most AI agents treat everyone the same. Same plan, same steps, same assumptions. SimBench tests whether that's changing. Open benchmark for preference-conditioned agentic planning with 47 tasks. 9 users. 4 domains. MIT licensed. github.com/versarai/simbench