Daily digest
6 items · ~6 min · Week 2026-W34
Must-read (1)
LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers
University of Illinois at Urbana-ChampaignOpen-source modular framework that unifies LLM routing as a five-component sequential decision process (context encoders, model encoders, scoring functions, decision rules, learning signals) and ships 16+ routers with the xRouteBench benchmark. Learned routers outperform the strongest fixed-model baseline by 14.6% relatively, and user-conditioned routing consistently improves personalization.
Worth knowing (2)
Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
Zhejiang UniversityInteractive video world model that externalizes persistent world state into a camera-indexed bank and uses a linear-scaling teacher (sparse chunk attention + linear-attention global state). On a single H200 at 384x640 each 1.5s chunk is generated in 2.11s, and the model achieves state-of-the-art on WBench while remaining competitive on VBench-Long and VBench-2.0.
OpenAI Codex CLI ships 0.148.0-alpha.20 on Aug 16 after -alpha.19 a day earlier
OpenAIopenai/codex released two tagged pre-releases in window: 0.148.0-alpha.19 on 15 Aug 2026 and 0.148.0-alpha.20 on 16 Aug 2026, both published by github-actions. Per-tag changelogs are not exposed on the GitHub release pages in a loadable form, so specific feature notes for each cut are not verifiable; only version, commit and ship date are confirmed against the official release feed.
For reference (3)
DarwinX: Evolving Agent Harnesses Through Natural Selection
Salesforce AI ResearchTreats agent self-improvement as population-level selection over harnesses (prompts, tools, skills, control flow) with the model frozen, using a preserve-and-extend contract to prevent regression and a verifier-based fitness signal. Reaches 84.7% on Terminal-Bench 2.1, 68.3% on TerminalWorld held-out, and 93.0% pass@1 on WebArena-Infinity; a Terminal-Bench harness transfers unchanged to SWE-bench Verified.
Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
Shanghai AI LabProposes the Mobius-v0 architecture: a globally shared Memory (FFN) stores knowledge vectors while multiple Reasoners (Self-Attention) iteratively query it via hidden states. A 7B model trained from scratch matches a Transformer baseline with 62.6% of the data, and Intern-S2-Mobius continually pretrained from Qwen3.5-35B matches its downstream score while delivering nearly 4x end-to-end inference speedup.
OpenClaw 2026.8.1-beta.2 pre-release ships GPT-5.6 Ultra switching and secret-egress binding
openclaw/openclaw 2026.8.1-beta.2 released 15 Aug 2026: adds GPT-5.6 Ultra across OpenClaw and Codex with atomic model/runtime/thinking switching via /model and fallback, secret egress host binding that fails closed on unbound substitution, channel plugin ingress monitors on IRC/Synology Chat/Google Chat, SQLite snapshot backup commands, macOS app profiles isolating Keychain/prefs/state, and plugin install provenance warnings requiring --force for arbitrary executable sources.