---
title: "Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.06503v1"
permalink: "https://feed.astralor.com/items/itm_1f1efec6d0651c86ffc4ef86"
published: "2026-07-07T17:03:55.000Z"
score: 78
channel: "tech"
tags: ["LLM Agent","Qwen-2.5-7B","Internal Representation Probe","Reasoning Acceleration","Early Abort","Northwestern University"]
---

# Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade

## AI commentary

> This work cleverly moves failure prediction from behavior to internal representations, using lightweight probes for early abort and saving substantial compute. For teams deploying multi-step agent services, this is a cost-cutting tool. The sample complexity analysis is a practical guide for data requirements.

## At a glance

LLM agents solving multi-step tasks often commit to doomed trajectories while wasting inference compute. This paper proposes an early abort cascade using lightweight probes on hidden activations to predict failure as early as the first round, with recall-controlled gates ensuring successful episodes pass at a user-specified global rate. On TextCraft, at 90% recall target, it saves 47.1% (Qwen-2.5-7B) and 37.2% (Llama-3.2-3B) compute, outperforming single-gate and behavior-only policies by 1.6-1.7x.

## Key points

- Predicts episode failure from LLM agent internal hidden activations as early as the first interaction round.
- Introduces a recall-controlled cascade of gates ensuring successful episodes pass at a global recall target.
- Achieves 47.1% and 37.2% inference compute savings on TextCraft; behavior-only probes save only half as much.
- Characterizes sample complexity for certifying high recall targets, providing practical data requirements.

## Full text

Large language model (LLM) agents solving multi-step tasks frequently commit to trajectories that are doomed to fail, yet continue to consume substantial inference compute before the failure becomes observable. We show that failure is predictable early from the agent's internal representations: lightweight per-round probes on hidden activations anticipate eventual episode failure as early as the first interaction round, where scorers reading only the agent's observable behavior are barely better than chance. We turn this signal into a practical abort cascade: one distribution-free calibrated gate per round, with per-round recall budgets jointly searched so that eventually-successful episodes survive all gates at a user-specified global rate; this episode-level guarantee is the one that matters in deployment, since false-abort risk accumulates across gates. Across two agent models on TextCraft, the cascade meets every recall target from 90% to 97% and, at the 90% target, saves 47.1% +/- 10.3% (Qwen-2.5-7B) and 37.2% +/- 8.8% (Llama-3.2-3B) of inference compute, 1.6--1.7x the best single-gate policy. An otherwise-identical cascade reading only behavior saves roughly half as much, and adding behavioral features to the probe yields no further gain: the hidden states capture what behavior reveals. Finally, we characterize the sample complexity of certifying high recall targets, telling practitioners which recall promises their data can, and provably cannot, back. The code will be released soon.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_1f1efec6d0651c86ffc4ef86