Monday, June 22, 2026

WorldLines: Benchmarking and Modeling Long-Horizon Stateful Embodied Agents. WorldLines benchmark and modeling approach for long-horizon stateful embodied agents. Source

GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents. GateMem benchmark evaluates memory sharing and governance in multi-principal agent systems. Source

PerceptionDLM: Parallel Region Perception with Multimodal Diffusion Language Models. PerceptionDLM extends diffusion language models to multimodal parallel region perception tasks. Source

Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations. Method for converting examples into task instructions to improve in-context learning for B2B dialog. Source

RRT-355M, Softmax-Free Attention Model. From-scratch, softmax-free attention model at GPT-2 Medium scale with sparse tile-skipping Triton kernels, open weights, and a custom inference engine. Source

Recall, Local Project Memory for Claude Code. Local memory layer for Claude Code that persists project context across short coding sessions. Source

Temporary Cloudflare Accounts for AI Agents. Cloudflare introduces temporary accounts scoped for AI agents. Source

browser-harness, Self-Healing Browser Automation. Self-healing browser automation harness designed to make LLM-driven web agents more robust. Source

Identity Verification on Claude. Anthropic introduces identity verification on Claude access; discussion highlights geopolitical fallout for international users and competitors. Source

There Is Minimal Downside to Switching to Open Models. Practitioner thread arguing open-weight models are already viable replacements for proprietary APIs, with routing-config examples and counterpoints on privacy and capability gaps. Source

When the Trump Administration Cracks Down on Anthropic, Who Benefits?. Analysis of who benefits if the Trump administration targets Anthropic. Source

Beyond Siri: Practical AI Features Coming to Your iPhone in iOS 27. Preview of practical AI features expected in iOS 27 beyond Siri. Source

Fine-Tuning a Local LLM (Qwen 3:0.6B) to Categorize Questions. Show-and-tell of fine-tuning a small Qwen model for question classification, with community pointing out simpler classical ML often suffices. Source

Best Current Methods for Finetuning Whisper on Domain-Specific Vocabulary?. User asks for current best practices to fine-tune Whisper on domain-specific Spanish vocabulary and how much labeled audio is needed. Source

Data-Centric Debugging for Teams Training Neural Nets. Open-source WeightsLab revamp for mid-training data debugging in PyTorch CV pipelines (images, video, LiDAR). Source