Fetching from the wire…
01
02
03
04
05
06
07
08
Infra2026-09-22 · source-backed
The argument is that enterprise agentic systems sending every trajectory step to a frontier model waste 60-80% of their inference budget, because complexity varies inside a single trajectory: the planning step needs frontier reasoning, the formatting step after it needs a 7B model. The classifier trains on 50,000 annotated agent trajectory steps across planning, coding, research and data analysis, using five features available at routing time with under 5ms overhead per step on an A100. This is the step-level counterpart to the query-level routing everyone already ships, and 5ms is small enough to try in a real harness. (arXiv 2609.22951)
Each link below shares sources, entities, or timing with this story.
Counterfactual regret minimization has been one of the few large numerical workloads that ran faster on CPUs, because each iteration issues millions of tiny interdependent gather and scatter steps where kernel launch and framework dispatch dominate. GPU-CFR uses the fact that...
arXiv 2609.19499 fixes N=8 on 500 GSM8K prompts and compares four generation schedules (1x8, 2x4, 4x2, 8x1) on A100s. Eight serial calls consume 4.64 to 4.86 times the gross GPU-device energy and show 5.77 to 6.12 times the P95 latency of one batched call producing the same ei...
arXiv 2609.17863 measured 54 configurations of Qwen2.5-7B-Instruct on vLLM 0.12 across L4, A100 and H100, then calibrated a simulator reproducing them with cross-campaign drift under 1.5%. On the calibrated grid, 18 of 36 configurations reach the cost/quality/latency frontier,...
100 real frontier research tasks across seven scientific domains, full lifecycle, 800 annotated trajectories, 45-pattern failure taxonomy (arXiv 2608.14905). The headline isn't a leaderboard, it's a shared deficit: agents can't check what they produced against what they found,...
arXiv 2609.20301 argues existing observability tools do per-execution debugging but not cross-run profiling, so nobody can answer where failures cluster or which tasks eat the budget at scale. The obstacle is that the responsible entity is a task intent like "diagnose authenti...
arXiv 2609.17598 studies PRs from OpenAI Codex, Devin, GitHub Copilot, Cursor and Claude Code across 2,807 repositories (Dec 2024 to Jul 2025), combining AIDev with 58,792 cached GitHub API responses. Codex PRs were reverted 6.1% of the time against a human baseline of 11.5% (...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.