EN RU

Daily digest

9 items · ~9 min · Week 2026-W32

Must-read (3)

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

Research official 2 src. ~1 min

A harness that reformulates long-horizon agent execution as an explicit task-state management problem, keeping verified state outside the model's growing context instead of letting it accumulate unchecked assumptions. It lifts Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench and 2.8% to 8.3% on OSWorld 2.0, and raises Claude Opus 4.7 from 20.0% to 34.3% on an OSWorld 2.0 subset.

Why it matters
HuggingFace Daily Papers top pick for August 4, 2026 with 210 upvotes; large jumps on long-horizon computer-use benchmarks suggest state-management design, not just model scale, is a major lever for agent reliability.
#agents #long-horizon #evals #tool-use

SwanTale: Unified Multi-Speaker Speech and Audio Generation for Instruct and Zero-Shot Tasks

ByteDance
Research official 1 src. ~1 min

ByteDance's model for multi-speaker expressive speech and audio generation supports both instruct (caption-driven) and zero-shot (reference-audio-driven) synthesis, combining reward-conditioned quality control, Engram conditioning, and a unified MoE trained with curriculum learning plus GRPO post-training. Targets applications like dubbing, audio drama, and short-video production.

Why it matters
HuggingFace Daily Papers listing with 142 upvotes; part of ByteDance's SwanAIGC audio research line alongside SwanVoice and SwanBench-Speech.
#speech #audio-generation #multimodal #moe

MiniMax open-sources H3 omni-modal video model weights

MiniMax
Video official + media 4 src. ~1 min

MiniMax released the open weights of its H3 (Hailuo 3.0) omni-modal video model on Hugging Face on August 3, 2026, three days after its API-only debut. The 33B-parameter release includes FL2VA (text/first-last-frame) and Ref2VA (reference-to-video) checkpoints, generating up to 2K, 15-second clips with native stereo audio, and shipped with day-zero native ComfyUI support.

Why it matters
One of the largest open-weight video-generation releases to date with native synchronized audio, though its community license explicitly excludes the US, EU, UK, and South Korea from running the weights locally.
#text-to-video #image-to-video #open-weights #audio-video-generation #video-editing

Worth knowing (3)

DAPD: Dual-Anchored Policy Distillation

Research official 1 src. ~1 min

Identifies a 'privilege illusion' failure mode in on-policy self-distillation, where a student model learns to rely on privileged teacher-time information it cannot access at inference. DAPD fixes this with dual-path anchoring and dual-source anchoring, beating prior on-policy self-distillation by +2.00 points averaged over six benchmarks on Qwen3-4B.

Why it matters
HuggingFace Daily Papers entry with 73 upvotes; names a concrete, previously under-described failure mode in self-distillation pipelines increasingly used to compress reasoning models.
#rl #distillation #reasoning #self-distillation

Progressive Agent Skill Generation via Reinforcement Learning (Skill-α)

The Chinese University of Hong Kong
Research official 1 src. ~1 min

Skill-α frames agent skill generation as a sequential editing process (CREATE, UPDATE, MERGE, PRUNE, NOOP) trained with a rollback reward that compares downstream execution under the original versus edited skill on an anchored query. It supports both document-to-skill and experience-to-skill generation through a unified skill editor.

Why it matters
HuggingFace Daily Papers entry with 52 upvotes; offers a structured, verifiable alternative to ad-hoc agent self-improvement loops.
#agents #rl #skill-learning #self-improvement

Mistral AI releases Shieldstral, an open-weight safety classifier

Mistral AI
Tools official 1 src. ~1 min

Mistral AI released Shieldstral, a 3-billion-parameter open-weight content-safety classifier for text and images, licensed under Apache 2.0. It accepts plain-language safety policies at inference time instead of requiring retraining, runs on a single 16GB GPU, and returns calibrated probability scores rather than binary labels.

Why it matters
Gives developers a lightweight, policy-adaptive moderation model they can self-host, competing with proprietary guardrail APIs from larger labs.
#safety #moderation #open-weights #apache2
For reference (3)

Anthropic names Tino Cuéllar as Chief Global Affairs Officer

Anthropic
Industry official 1 src. ~1 min

Anthropic announced that Mariano-Florentino (Tino) Cuéllar, a former California Supreme Court justice, is joining the company as Chief Global Affairs Officer to lead its policy and government-relations work.

Why it matters
Signals Anthropic's continued investment in AI policy engagement as regulatory scrutiny of frontier labs intensifies globally.
#leadership #policy

Claude Code v2.1.222 fixes worktree destructive-git-command and hook-bypass safety gaps

Anthropic
Tools official 1 src. ~1 min

Following yesterday's v2.1.221 (Focus view, sandbox credential masking), Anthropic shipped Claude Code v2.1.222 on August 4, fixing two safety gaps: worktree-isolated sessions and subagents could run destructive git commands against the main checkout, and PreToolUse auto-allow hooks could be bypassed by tools running in background tasks.

Why it matters
Closes real safety gaps for isolated and background Claude Code sessions running unattended.
#claude-code #coding-agent #cli #sandbox #security #bug-fix