Daily digest
12 items · ~12 min · Week 2026-W34
Worth knowing (7)
τ₀-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation
Shanghai Innovation InstituteHierarchical robot foundation model that treats high-level subtask generation as a compute-scalable inference problem solved via world-model-guided beam search with execution memory; trained on 40,115 hours of real-world multimodal data and shows large closed-loop gains on long-horizon manipulation under distribution shift.
4DAnyone: Create Anyone in 4D from a Casual Monocular Video
Ant ResearchReconstructs animatable 4D humans from an uncalibrated monocular video by generating reconstruction-grade, multi-view-consistent videos via a camera-controlled video diffusion model and lifting them to 4D Gaussian Splatting; introduces Reference Context Packing and Target Context Routing plus the MVGameHuman dataset.
EnvHarness: Awakening Static Worlds for Agent Learning
GoogleAn 'ActionableEnv' wrapper plus 'EnvRigger' that programmatically reshape static LLM-agent benchmarks via Setup/Rule/Link plugins, targeting specific agent weaknesses without rebuilding verifiers; across five benchmarks in four domains, lifts SWE-bench Verified resolution from 52.13% to 54.79% and trims average steps per episode by 9.8%.
FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis
University of Science and Technology of ChinaFramework that synthesizes high-quality executable terminal-agent training tasks by reconstructing agent skills into coherent scenarios and grounding all task artifacts in a single shared container state, preserving both source intent and executable state across artifacts.
Claude Code v2.1.241 ships
AnthropicAnthropic released Claude Code v2.1.241 on 2026-08-23 as a maintenance release with bug fixes and reliability improvements; the GitHub release notes body carries only the headline summary.
Claude Code v2.1.240 ships
AnthropicAnthropic released Claude Code v2.1.240 on 2026-08-22 as a maintenance release with bug fixes and reliability improvements; the GitHub release notes body carries only the headline summary.
OpenAI Codex CLI ships two alpha pre-releases on Aug 22
OpenAIOpenAI Codex CLI tagged two pre-release builds on 2026-08-22: 0.150.0-alpha.7 at 21:26 UTC and 0.149.0-alpha.7.2 at 19:48 UTC. The GitHub release pages carry no changelog body for either alpha beyond the tag name.
For reference (5)
Qwen Code v0.22.0 stable release
AlibabaQwen Code v0.22.0 stable released on 2026-08-22 with 50 merged PRs covering Web Shell OOM-prevention, review-loop explanations, autofix auditing, and Python/Java SDK parity for 'auto' permission mode. The release notes report SWE-bench Verified at 77.08% (380/500 resolved) using qwen3.7-plus as the underlying model.
SGLang v0.5.18 lands major model and perf updates
SGLang v0.5.18 shipped on 2026-08-22 with 710 PRs from 212 contributors, adding Muse Glimmer, Intern-S2-Mobius, SANA-Video, LingBot-Video-MoE, LTX-2.5, Cosmos3 Edge & Distilled, and LongCat-Image. Overlapped checkpoint staging makes Qwen3-32B on H100 start 2.38x faster (35.6s vs 84.8s) under --startup-weight-load-mode overlap; TP LMHead All-to-All cuts DeepSeek-V4-Pro B200 decode LMHead time from 320µs to 169µs; FlashInfer MNNVL gives DeepSeek-V4-Flash TP4 decode up to +6.9% at small batches.
Cline v4.1.13 fixes OpenAI-Compatible tool calling and Hub session continuity
ClineCline v4.1.13 shipped 2026-08-22, restoring tool calling for custom OpenAI-Compatible models whose capability list was inferred from flags like supportsReasoning, keeping Hub-backed sessions intact across Hub restart/upgrade via event replay with dedupe, and propagating session/client identity into Langfuse traces.
Cline SDK, Desktop and CLI add hub drain/upgrade commands
ClineAcross 2026-08-22 and 2026-08-23, Cline shipped SDK v0.0.78, Desktop v0.0.16, and CLI v3.0.57. New CLI commands 'cline hub drain' (stop accepting new mutating work, finish in-flight) and 'cline hub upgrade' (drain, wait for idle, restart) make Hub handoffs durable via event replay deduped by event id; Desktop honors server-side feature flags and refreshes them on account change.
llama.cpp rolls up Aug 22 backend and model fixes
Between 2026-08-22 08:13 and 23:40 UTC, llama.cpp tagged ten commit-releases (b10577–b10588) on the master branch. Notable items: PAD_REFLECT_1D op added to the Vulkan backend, ROCm CI restored for Ubuntu, DSpark support for bailingmoe3, mtmd vision+audio support for dots3-note, common/json.h abstraction, concat op optimized via row-level memcpy, mtmd switched to ggml_rope_set_offset, draft-mtp with embeddings fixed.