TUESDAY · SEPTEMBER 22, 2026 · ISSUE 15 · YESTERDAY No. 15
Daily News — 2026-09-22
341 papers indexed on arXiv·~17,000 packages released on PyPI·31 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
An ontology that evolves per-backbone and cuts total token cost while improving accuracy is a more honest solution to schema discovery than any static prompt layer.
The core problem EvoOntology addresses is embarrassingly common: a data agent querying a large, heterogeneous collection of tables, files, and databases has no prior map of what lives where. It rediscovers schemas from…
A capable RAG and agent framework whose own README now describes it as secondary to the company's commercial document-parsing platform.
LlamaIndex launched in 2022 as a broad orchestration toolkit for RAG and agent applications. Three years later, the README opens with a note that the company's primary focus has shifted—away from the OSS framework and…
Papers
An ontology that evolves per-backbone and cuts total token cost while improving accuracy is a more honest solution to schema discovery than any static prompt layer.
EvoOntology wraps a self-evolving ontology as an MCP server so data agents can query and refine semantic structure at runtime.
Filtered source-code-derived tasks beat a larger unfiltered set — quality of RL environments matters more than raw count.
CodeMidas builds RL training tasks from raw source code alone—no issues or commits—yielding 5,545 tasks and double-digit benchmark gains.
A clean architectural fix to MoE's three-way coupling, validated in production at scale with measurable gains.
IntBMoE decouples MoE participation, execution, and materialization via block-conditioned dense composition, deployed at AMap with a 2.4% UVCTR gain at scale.
Timestep-gated solid-color anchors plus perceptual clustering turn hex strings into a reliable prompt-native color specification—no separate encoder required.
Paint-Anything trains a hex-prompt interface for color-accurate image generation and editing, improving ACBench scores by 85.3% and 28.3% on FLUX.2-4B.
A frozen-model design agent that improves through skill-library evolution rather than weight updates — honest about where prose guidance hits its ceiling.
Procedural memory grown from 1,406 real user briefs—no weight updates, no human labels—lifts design-agent execution success from 72.7% to 99.3%.
Factor-level disentanglement scoring exposes model failures that holistic reference-fidelity metrics structurally cannot see.
OmniVBench fills a gap in R2V evaluation by testing factor-level disentanglement across 18 tasks, paired with a 340K-sample training dataset.
10 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Real BI data breaks LLMs in ways NL2SQL benchmarks never reveal — and specialized data-management tools close most of that gap.
BI-Agent closes a 40-point accuracy gap on BI-Bench, a new end-to-end BI benchmark where frontier LLMs alone score below 50%.
Gricea makes the executable study the artifact of record — yet 28 of the 29 reviewed papers lacked enough detail to replicate their own conditions.
Gricea packages conversational AI studies as reusable artifacts; a replication audit found 96% of surveyed papers lacked enough detail to reproduce faithfully.
A small architectural fix—context-gated memory slots instead of one fixed row per token—that consistently beats its baselines at sub-billion scale with minimal latency cost.
MoME fixes context-blind token memory by gating over hidden state to pick among M slots per token, separating senses like 'python' the language vs. the animal.
Tools & packages
A tightly scoped local automation bridge to Jianying Pro's native engine, honest about what it cannot yet guarantee.
A headless interface for Jianying that supports draft editing, export, and a standalone agent skill.
A capable RAG and agent framework whose own README now describes it as secondary to the company's commercial document-parsing platform.
LlamaIndex bridges LLMs and your own data through index abstractions, query engines, and pluggable data connectors.
Honest cost arithmetic and conservative installation make this a credible Astra-sparing workflow, not a speculative one.
A Codex workflow that uses Astra for planning and review while DeepSeek Flash handles building, with phased tasks and reversible setup.
Structured-state control with a bounded action set and an unusually honest evidence chain — this is what a verifiable AI game agent actually looks like.
A Minecraft agent combining an Astra planner and JEV controller, with native recording, tested routes, and run verification.
A 72-hour live build distilled into failure logs, role files, and agent rules you can drop into any repo today.
Drop-in AGENTS.md with rules, playbooks, bot roles, and a failure log from a 72-hour live xAI Grok Bot build.
A file-by-file Rust port of Claude Code's TUI, backed by its own iocraft fork and a companion reverse-engineering repo — incomplete but architecturally honest.
Rust reimplementation of Claude Code's terminal UI using iocraft, a Rust-native TUI component library.
9 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Reads logits instead of generating text, turning any compatible open model into a structured classifier without a decode loop.
featherless-ai/simple-jev wraps any open model to expose a classifier or JEV scoring endpoint — no fine-tuning required.
A decision-routing layer that measured its way to one honest failure and shipped the better fallback — worth reading for the privacy and fail-open specs alone.
Jev-powered skill layer for Hermes agents — model routing, memory, compaction, computer/browser use — with Claude Code and Codex support too.
A well-documented skill library for typed agent decisions, with honest caveats and real API receipts rather than invented examples.
Community-contributed use cases, workflows, and agent skills for the Jev framework.
A no-generated-answers search front-end that exposes model judgments as editable filters and documents its fallbacks honestly.
Jev is a TypeSafe web search tool covering source selection, query understanding, and relevance ranking via Search1API.