EN RU

Daily digest

6 items · ~6 min · Week 2026-W32

Worth knowing (1)

N0-TWAM: Scaling Tactile-Native World-Action Model for Contact-Rich Manipulation

NeoteAI / Fudan TEAI
Research official 2 src. ~1 min

Presents the first large-scale tactile-native world-action model, pretrained with joint visuo-tactile learning over demonstrations spanning six robot embodiments and 450 tasks, using a unified force-based tactile representation (NeoForce) to condition action generation for contact-rich manipulation.

Why it matters
Extends world-action model scaling beyond vision into touch, a modality largely missing from prior large robotics foundation models; received 48 upvotes on HuggingFace Daily Papers (2026-08-03).
#robotics #world-models #tactile #manipulation #embodied-ai
For reference (5)

Weak-to-Strong On-Policy Distillation

Microsoft Research / University of Maryland / MBZUAI
Research official 2 src. ~1 min

Introduces W2S-OPD, which builds a proxy teacher from the logit difference of a weaker positive/negative model pair to distill capability into a student even when no stronger teacher exists. The student minimizes per-position reverse KL against this proxy teacher on its own rollouts.

Why it matters
Tackles a practical limitation of on-policy distillation at the frontier, where no larger teacher model is available to distill from; received 18 upvotes on HuggingFace Daily Papers (2026-08-03).
#distillation #llm-training #rl #weak-to-strong #on-policy

Scaling Properties of Text Conditioning in Visual Generation

ByteDance Seed
Research official 2 src. ~1 min

Studies how diffusion loss scales with the amount of structured language in text prompts, introducing white-box (GPG) and black-box (ED) metrics to quantify prompt structure, and shows converged loss decreases roughly linearly with GPG and follows a power law with ED.

Why it matters
Fills a gap in scaling-law research for visual generation, where prompt-token scaling had rarely been measured; the resulting system is competitive with both open- and closed-weight models on compositional and reasoning benchmarks.
#diffusion #scaling #text-to-image #multimodal #seed #bytedance

Claude Code v2.1.221 adds Focus view and sandbox credential masking

Anthropic
Tools official 2 src. ~1 min

Claude Code 2.1.221 (Aug 4, 2026) adds a Focus view that hides tool activity behind a per-turn summary, introduces credential-file masking inside the sandbox on Linux/WSL, and fixes several Bash/PowerShell permission-check and MCP connection bugs.

Why it matters
Security hardening (credential masking, permission-check fixes) matters directly for teams running Claude Code as their daily coding agent.
#claude-code #cli #coding-agent #sandbox #security

OpenCode v1.18.12 fixes Azure GPT-5.5+ reasoning failures and composer lag

sst
Tools official 1 src. ~1 min

OpenCode v1.18.12 (Aug 4, 2026) fixes Azure GPT-5.5+ completion failures when reasoning is enabled, reduces composer lag when pasting large images, and improves project search to surface any known recent project instead of only the first five.

Why it matters
Reliability fixes for Azure-hosted reasoning models remove a real blocker for teams running OpenCode against Azure OpenAI deployments.
#opencode #coding-agent #cli #azure #sst

OpenClaw ships v2026.7.1 patch releases fixing Codex integration and plugin updates

OpenClaw
Tools official 1 src. ~1 min

OpenClaw's v2026.7.1-1 and v2026.7.1-2 patches (both Aug 4, 2026) fix Codex progress replies losing app-server turn state, Memory Core startup recovery from legacy-index conflicts, WSL state-permission handling, and managed-plugin updates that were blocked by newer npm's singleton-array lock metadata.

Why it matters
The Codex-integration and plugin-update fixes directly affect OpenClaw's viability as a self-hosted gateway for coding agents.
#openclaw #coding-agent #mcp #plugins #codex