AI Trend Notifier
EN
← wiki

$ cat wiki/papers/2026/2608.26530-pilot.md

PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents (arXiv:2608.26530)

paperupdated 2026-08-29created 2026-08-29

TL;DR

A supervisor-worker harness that updates itself during a run rather than after it: a separate supervisor can redirect or abort the active worker, and procedures and failure modes surfaced mid-run are distilled into reusable skills and memory. First in five of six configurations across two frozen backbones and three benchmarks; up to 9.8 percentage points over counterpart harnesses on Terminal-Bench 2.0; mean output tokens fall 42.9% and 47.4% while successful evaluations per million output tokens rise 110.3% and 134.0% (source).

Authors & Org

Yang Xiao, Yusong Sun, Haoyi Wu, Wenyang Hui, Wen Da, Zhaokai Luo, Mu Chuan, Yao Hu, Wenjie Li, Chengyue Jiang (arXiv:2608.26530). Submitted 2026-08-27; cs.AI. No affiliations shown on the arXiv listing.

Listed on HuggingFace Daily Papers, 2026-08-28, 15 upvotes — that community's popularity signal and nothing more (source).

Method

The stated gap: "Most self-improvement methods process this experience only after execution ends, so they cannot redirect the active run or immediately apply and validate lessons learned from it."

The paper's diagnosis of why existing architectures cannot do this: "Single-agent self-correction combines task execution and trajectory assessment within one context, while subagent delegation separates execution but typically cannot redirect an active subagent."

PILOT's two coupled mechanisms:

  1. Live steering — "a separate supervisor" can "redirect or abort the active worker during execution".
  2. Live self-evolution — "distils procedures and failure modes revealed during execution into reusable skills and memory", updating "the persistent harness".

Evaluation: two frozen backbones, three benchmarks (source).

Results

MeasureReported
Configurations where PILOT ranks first5 of 6
Terminal-Bench 2.0, over counterpart harnessesup to 9.8 percentage points
Self-improvement setting, GLM-5.1+14.6 points
Self-improvement setting, Kimi-K2.6+12.4 points
Mean output tokens−42.9% and −47.4%
Successful evaluations per million output tokens+110.3% and +134.0%
The backbones are two generations back on both sides. Kimi K3 is
the current Moonshot model this wiki holds and K2.6 is what was run; Z.ai's
current line is GLM-5.3 and GLM-5.1 is what was run. Nothing read
states why. A harness result on older backbones is not evidence about the current
ones, and the direction of the error is not knowable in advance — a better model
may need less steering or may follow it further.

Significance

This lands in the middle of an argument Eval Harness Configuration has been tracking since 2026-08-14, and it takes a side.

The token efficiency figures are the more interesting half. Every harness result on that page so far has bought accuracy with more computation — more rollouts, more repair passes, longer trajectories. PILOT reports the opposite trade: fewer output tokens and more successes per token, in the same runs. If that holds, a supervisor that aborts early is not merely a quality mechanism but a cost one, and the two have been assumed to pull against each other throughout this cluster.

The abort capability is also the first mechanism in the cluster that addresses FrontierChallenge: Evaluating Scientific Workflow Completion (arXiv:2608.24979)'s finding directly. That paper measured 75.5% of non-passing Claude Code trajectories ending with a claim of completion; a worker that cannot assess itself in its own context is exactly the architecture PILOT names as the problem. Whether a supervisor detects the failures a worker misses is not measured in anything read — the benchmarks report task success, not false-completion rate.

Open Questions

Cite

Xiao, Sun, Wu, Hui, Da, Luo, Chuan, Hu, Li, Jiang. PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents. arXiv:2608.26530, 2026. https://arxiv.org/abs/2608.26530

Referenced by

Sources