Model Releases

  • Hy-SimulMT-30B-A3B-FP8 — Tencent drops a transformer-based MoE model for translation across 20+ languages, with fp8 8-bit quantization and endpoints compatibility for efficient multilingual serving. The sparse MoE activation architecture reduces computational overhead for large-scale deployment without sacrificing translation quality across diverse language pairs. 🤖🌐

Open Source Releases

  • anthropics/claude-code v2.1.312 — Added OIDC user_id_claim to the Claude apps gateway, enabling user identification by an id_token claim other than sub, such as Microsoft Entra ID’s oid, plus click-to-send in fullscreen mode and fixes for managed settings and organization handling. 🛠️
  • whichllm 0.5.22 — Utility that matches available LLMs to user hardware by benchmarking model size, memory requirements, and inference speed, enabling efficient local or edge deployment decisions given specific GPU/CPU constraints. 📊
  • ifcfill 0.5.0 — Transformer module converting raw tabular data into structured IFC (Integer, Float, Categorical) variables with built-in missing-data imputation, designed to standardize preprocessing pipelines for tabular ML models. 📄

Research Worth Reading

  • From Structural Feedback to Prompt Policies: Learning Faithful Text-to-Image Prompt Editors — Proposes StruPO, a structure-aware prompt editor that uses a calibrated scene graph to separate scene-level semantics from stylistic rewriting, ensuring prompt optimization preserves requested objects and relations while improving image fidelity. Evaluates on text-to-image benchmarks showing improvements. 🎨
  • Spiking Neural Networks Are Harder to Steal: A Benchmark Study of Model Stealing Attacks — Systematically evaluates model stealing attacks against spiking neural networks, revealing significantly lower surrogate model accuracy compared to rate-based ANNs due to temporal encoding and sparse activations. Provides a benchmark suite and analysis of query efficiency vs. attack success across spiking models. 🧠
  • Vision Token Pruning via Query-Vision Interaction Decomposition — Decomposes query-vision interaction into separate query and vision relevance scores to guide dynamic token pruning in VLMs, achieving up to 40% FLOPs reduction with minimal accuracy loss across multimodal benchmarks. Introduces a differentiable pruning module that integrates with existing VLM pipelines. ⚡
  • The Trinity: Capability-Driven Recursive Self-Improvement for Multimodal Agents — Introduces a capability-driven framework for recursive self-improvement in multimodal agents, addressing environment construction bottlenecks and harness iteration limits via automated capability assessment and environment generation. Evaluates on simulated robotics and vision-language tasks, showing improved autonomy. 🤖
  • Don’t Pick a Winner: Knowledge Conflicts Need a Prior, Not a Rule — Frames knowledge-retrieval conflicts in RAG systems as a Bayesian prior selection problem rather than a fixed rule, demonstrating improved answer accuracy across incomplete and misleading contexts. Proposes a learnable prior over context trustworthiness that adapts to domain-specific retrieval quality. 📚
  • Skill Dreaming: Evolving Repository Skills through Pull Request Streams — Proposes a method to continuously evolve LLM agent skills by parsing pull request diffs and commit histories, generating synthetic skill-improvement trajectories that adapt to codebase changes. Validated on open-source repositories, showing skill persistence and transfer across PR cycles. 📈

AI Dev Tools

  • expo/skills — Expo Skills is a collection of AI agent skills tailored for Expo projects and Expo Application Services, providing reusable capabilities for building, testing, and deploying React Native applications with AI-assisted workflows. The skills integrate with the Expo SDK to automate common development tasks. 🛠️🚀

Today’s Synthesis

If you’re building multimodal agents that keep tripping over long-horizon tasks, stitch together The Trinity: Capability-Driven Recursive Self-Improvement for Multimodal Agents and Vision Token Pruning via Query-Vision Interaction Decomposition with Skill Dreaming: Evolving Repository Skills through Pull Request Streams . The Trinity treats capability growth as an explicit loop: assess what the agent can do, then generate better environments and iterate harness behavior. But environment and observation streams are expensive, especially when they include full video or image histories. Pair that with query-vision interaction decomposition, which separates query and vision relevance scores to prune tokens dynamically—up to 40% FLOPs with minimal accuracy loss. Instead of feeding every frame into the self-improvement loop, score and prune raw multimodal context before it becomes part of the agent’s state. Then let Skill Dreaming evolve repository-specific skills from PR diffs and commit history, so the pruned, capability-driven loop doesn’t just optimize prompts or environments in isolation. Concrete move: add a differentiable token-pruning module to your VLM pipeline, use it to cap observation tokens per capability assessment step, and route surviving trajectories into skill-evolution data. That gives you cheaper self-improvement cycles without losing the fidelity needed to learn what actually changed.