THURSDAY · AUGUST 20, 2026 · ISSUE 2 · YESTERDAY
Daily News — 2026-08-20
263 papers indexed on arXiv·~7,300 packages released on PyPI·33 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
Skills stabilize execution rather than supply knowledge — and retrieval precision is largely irrelevant to whether the task actually succeeds.
Skills in LLM agents don't work by injecting missing facts. They work by stabilizing action sequences the agent already knows how to execute but keeps fumbling.
The deliberate anti-abstraction philosophy—one file per model, no shared building blocks—is what makes transformers the ecosystem's load-bearing definition layer.
The most important sentence in the transformers README is buried in the 'when not to use' section. It's still there in 5.15.1: the model files are deliberately not refactored into modular building blocks, so…
Papers
Skills stabilize execution rather than supply knowledge — and retrieval precision is largely irrelevant to whether the task actually succeeds.
A study finding that skills improve LLM agents by stabilizing execution via procedural anchoring, not by supplying missing knowledge.
When trajectories run long and rewards are sparse, ES avoids RL's horizon-wise variance accumulation — and the Sudoku ordering reversal across three horizon lengths is the cleanest proof of that.
Evolution-strategy fine-tuning for long-horizon LLM agents with full-parameter updates and minimal GPU requirements, using trajectory-level reward weighting.
Letting a VLM click pixels instead of predict 3D coordinates turns out to be the fix that makes embodied navigation both faster and more data-efficient.
TAMP-Nav aligns vision-language models with 2D visual prompting and compressed selective memory to improve embodied navigation efficiency.
Reconstruction fidelity is a poor proxy for generative utility, and V-RAE's frozen-encoder latent spaces prove the point with faster convergence and better generation scores.
V-RAE builds semantically organized video latents from frozen vision representations to improve generation quality, convergence, and predictive modeling.
The first rigorous account of what breaks when a real agent harness owns the RL training loop - and a reproducible fix that gains 14.6 points on SWE-bench.
A framework for reproducible reinforcement learning across arbitrary agent harnesses, targeting improved coding-agent performance with minimal data and compute.
A principled fix for the hallucination problem in two-stage HR editing—sparse attention guided by first-stage correspondence maps, not a generic upscaler.
A diffusion bridge method that translates low-resolution edits into high-resolution outputs while preserving source image details via sparse attention.
7 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A prompt-only, training-free defense that cuts unsafe image generation rates dramatically without touching model weights or requiring API access beyond image queries.
A black-box, zero-shot prompt-level defense that uses distribution-guided suffix expansion and contrastive scoring to reduce harmful image generation.
Lego-RL shows that training coding agents with RL fails silently at the infrastructure layer long before it fails at the model layer.
LEGO-RL pipes coding-agent harnesses directly into policy-gradient training and shows sparse MoE gains across multiple harnesses.
LCA turns a hierarchical model's learned byte segments into a parallel generation unit, gaining throughput without adding prediction heads or sacrificing quality under speculative verification.
Multi-byte prediction cuts inference time in byte-level hierarchical LMs using variable-length windows and causal masking, with minimal quality loss.
Tools & packages
A filesystem abstraction for agent context that trades vector-store opacity for navigable, observable, tiered memory — with benchmark numbers that hold up to scrutiny.
OpenViking unifies agent memory, RAG knowledge, and skills into a single self-evolving context store.
A structured, agent-driven video production system with real quality gates — not just a prompt-to-clip wrapper.
Open-source repo that turns an AI coding assistant into a video production studio via 12 pipelines, 100+ tools, and 700+ skill files.
The deliberate anti-abstraction philosophy—one file per model, no shared building blocks—is what makes transformers the ecosystem's load-bearing definition layer.
A framework for defining and running machine learning models across text, vision, audio, and multimodal tasks, for inference and training.
A protocol-translation proxy that genuinely handles the plumbing between coding agent UIs and arbitrary LLM backends — useful, but read the ToS disclaimer before connecting Anthropic.
A proxy that routes Codex CLI, App, and SDK requests to alternative LLM backends including Claude, Gemini, Grok, DeepSeek, and Ollama.
A zero-dependency Python skeleton for mid-execution agent routing that competes SELF, COLLABORATE, and HANDOFF in one utility function—honest about what the synthetic benchmark can and cannot prove.
State-aware A2A router that selects SELF, COLLABORATE, or HANDOFF mode based on agent state at runtime.
A self-hosted agent multiplexer that makes parallel, cross-device, multi-provider coding workflows scriptable without routing your code through anyone else's servers.
Paseo lets you orchestrate multiple coding agents from both desktop and mobile.
9 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
817 structured security playbooks with progressive token disclosure — the architecture is sound, but skill accuracy rests on community trust, not independent audit.
A collection of 817 structured cybersecurity skills for AI agents, mapped to 6 frameworks across 29 security domains.