Back to AI Research

AI Research

Doomed from the Start: Early Abort of LLM Agent Epi... | AI Research

Key Takeaways

  • Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade Large language model (LLM) agents often spend significant comp...
  • 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.
  • 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.
  • Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade Large language model (LLM) agents often spend significant computational resources on tasks that are destined to fail.
Paper AbstractExpand

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.

Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade
Large language model (LLM) agents often spend significant computational resources on tasks that are destined to fail. Because these agents operate over many interaction rounds, they continue to generate tokens long after a failure becomes inevitable. This research introduces a method to detect these "doomed" episodes early by analyzing the model's internal hidden states. By using a series of calibrated "gates," the system can safely abort failing episodes early, saving substantial compute while ensuring that successful episodes are allowed to complete.

Predicting Failure from Within

The researchers found that an agent’s internal hidden states provide a much earlier warning of failure than its observable behavior. While traditional methods that look at the agent's output (such as error rates or token probabilities) only become reliable after several rounds, probes placed on the model's internal activations can predict failure as early as the first round. By the time behavioral signals become useful, a significant portion of the episode's computational cost has already been spent. The study shows that internal states capture all the information present in behavioral signals, meaning that looking at the output alone is insufficient for early detection.

The Recall-Controlled Cascade

To manage the risk of accidentally stopping a task that would have eventually succeeded, the authors developed a "recall-controlled cascade." This system places a gate at each of the first few rounds of an episode. Each gate uses a statistical technique (Clopper–Pearson calibration) to ensure that the probability of a successful episode being incorrectly aborted remains within a strict, user-defined budget. Because the risk of a false abort accumulates as an episode passes through multiple gates, the system uses a global recall target to ensure the entire process remains safe. The researchers optimize these gates by searching for the best balance of "budgets" across different rounds to maximize compute savings.

Significant Compute Savings

The cascade proved highly effective across two different agent models (Qwen-2.5-7B and Llama-3.2-3B). At a 90% global success recall target, the system saved between 37% and 47% of inference compute. This performance was 1.6 to 1.7 times better than using a single-gate policy, demonstrating that distributing the decision-making process across multiple rounds is more efficient than relying on one single check. The results also confirmed that the system consistently met its safety targets, proving that the method provides reliable, predictable control over task success.

Data Requirements and Limitations

The study provides a clear account of the data needed to certify these performance guarantees. Because the system relies on statistical confidence bounds, it requires a sufficient number of successful episodes to verify that the recall targets are being met. The researchers note that for very high recall targets (such as 98% or 99%), the amount of data required increases significantly. This transparency allows practitioners to understand exactly which performance promises their available data can support before they deploy the system, ensuring that the safety guarantees are grounded in empirical evidence.

Comments (0)

No comments yet

Be the first to share your thoughts!