WEDNESDAY · SEPTEMBER 23, 2026 · ISSUE 16 · YESTERDAY No. 16
Daily News — 2026-09-23
309 papers indexed on arXiv·~15,000 packages released on PyPI·34 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
A dual-loop architecture that keeps conversation live during async tool execution, with honest benchmarks showing where argument grounding and delegation routing still fall short.
Realtime-Venus splits a conversational AI system into two concurrent loops that never have to pause for each other. The foreground loop handles live perception and speech; the background loop executes tool calls and…
A cost-driven split: cheap accessibility-text navigation via Jev, expensive visual judgment reserved for Codex—honest about where the speedup comes from and where it doesn't.
The core idea here is a division of labor: Jev, TypeSafe's browser-action model, handles clicks, scrolls, toggles, and navigation, while Codex retains responsibility for text input, visual interpretation, and final…
Papers
A dual-loop architecture that keeps conversation live during async tool execution, with honest benchmarks showing where argument grounding and delegation routing still fall short.
A full-duplex dialogue system using separate audio-visual and audio models for continuous perception, native speech generation, and asynchronous tool execution.
Regularizing the search dynamics rather than the edit space is what finally makes harness evolution transfer out of distribution.
RRSI applies regularization to recursive agent-harness self-improvement, gaining up to 4.7 points on out-of-distribution benchmarks across eight tasks.
A rigorous multi-game benchmark that exposes a clear gap between planning and action execution across every model family tested.
A benchmark suite pairing 5,000 hours of AAA gameplay footage from 21 games with multi-horizon instructions and offline/online evaluation tracks.
Token-level correction cuts annotation time in half while keeping training data on-policy—the benchmark showing top models barely reach 17% F1 on the task is the honest punchline.
An interactive annotation tool that uses token-level correction to steer LLM outputs, reducing median annotation time by 52% over manual post-editing.
Category-specific RL experts consolidated via label-routed distillation beat pooled training on every task category—but expert-gain recovery stays uneven across categories.
A category-aware RL framework that trains per-category software engineering agents and distills them into one policy, reaching 58.04% resolution on Pro-618.
A three-layer persona architecture with ADOS-inspired evaluation that measurably closes the gap between LLM dialogue and human conversational distributions.
A three-layered architecture for building role-playing agents organized into observable expression, latent beliefs, and core motivational drives.
9 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Fitting parametric distributions to your own policy's failures, then sampling from them for corrective training, beats uniform perturbation by over 10 points.
CARE improves VLA robot recovery by learning from real failed rollouts, gaining 14.5 pts in simulation and 15.9 pts in real dual-arm tasks.
Harness-Zero turns optimized agent scaffolding into training data, letting a 9B model internalize harness gains and then run without them.
Harness-Zero distills optimized agent harnesses into model weights, lifting macro-average task success from 23.3% to 44.3% without the harness at deployment.
Gradient reliability, not just token usefulness, determines which 1% of a student rollout is actually worth supervising.
A method that adds an information-efficiency ratio to token selection in on-policy distillation, matching full supervision at budgets as small as 0.1%.
Tools & packages
A cost-driven split: cheap accessibility-text navigation via Jev, expensive visual judgment reserved for Codex—honest about where the speedup comes from and where it doesn't.
Browser automation split between Jev (clicks) and Codex (reasoning/verification), claiming 5–10x faster operations; built at EZCollegeApp.
A local, open-weight implementation of logit-only classification that's honest about where it matches hosted alternatives and where it doesn't.
Open, local alternative to Jev: routes typed LLM decisions without generating a single token.
A sync'd draft-to-film viewer that documents the structural logic of a motion ad, not just its final look — narrow scope, zero-dependency, honest about what it is.
A motion-graphics project showing a finished ad alongside its black-and-white draft frames, with frame-accurate MP4 export.
A loopback proxy that routes each coding-assistant turn to the cheapest model that can handle it — with explainability and sensible fallback rules built in.
jev-router picks the cheapest Claude Code model per task, reducing costs without manual model selection.
A byte-exact Vulkan reimplementation of DLSS 5's neural rendering pass — not an upscaler, a full-resolution re-renderer — with PTX, GLSL, and a WebGPU port, but no weights included.
Vulkan reimplementation of NVIDIA's DLSS 5 Neural Rendering that targets bit-exact parity with the original closed-source network.
Typed, sub-second judgment tools that fail closed and keep policy in your code, not in the model.
Wraps TypeSafe's Jev model as MCP tools for fast, cheap, typed judgments in agent pipelines.
9 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Calibrated, threshold-able decisions from any open LLM's logits — no fine-tuning, no generation, and a 6.8× lift in safely-automatable traffic.
Turns any LLM into a Jev-style decision model: typed decisions with calibrated probabilities, no training needed.
Serious geolocation tooling that keeps the model as judge and lets geometry do the actual work — honest about what it can and cannot yet measure.
Geolocation skill for Claude Code, Codex, Cursor, Gemini CLI and others — uses OSM geometry, elevation, satellite and street view, and shows its reasoning.
A tight active-learning CLI that turns uncertainty sampling, human labels, and GEPA prompt optimization into a repeatable improvement loop for LLM classifiers.
Jev + GEPA repo for turning human feedback into calibrated AI functions with built-in alignment scoring.