LLMs Encode Their Failures: Predicting Success From Pre-Generation Activations

Quick Overview

The research reveals that employing linear probing, which involves drawing a line between success and failure signals in an LLM's internal state, significantly reduces inference costs (by 70% compared to the baseline) and improves accuracy on reasoning tasks, demonstrating that the model's internal state is highly informative about the likelihood of failure, even when the model is deep in thought, thus providing a cost-effective way to monitor model reasoning without incurring full computation.

Key Points: A new technique called Linear Probing successfully predicts LLM failures by analyzing internal state vectors at the end of instruction tokens. The small model (GPT-OS 20B) achieved a significant 70% reduction in inference cost per million tokens by routing difficult problems to a larger, more expensive model only when necessary. The cost reduction was achieved by using the small model as a gatekeeper, rerouting prompts predicted to fail to the larger model, saving computation on easy tasks. The linear probe's accuracy in predicting failure correlated strongly with the human difficulty score for the same problems, suggesting the internal state is highly indicative of reasoning difficulty. The paper suggests decoupling reasoning from human mimicry is crucial, as the model's internal state (when correctly probed) is a better measure of difficulty than human perception of that difficulty. The linear probe works by training a simple linear classifier on the internal state vector generated after the final instruction token to predict success or failure.

Context: This video discusses findings from a research paper titled "LLMs Encode Their Failures: Predicting Success From Pre-Generation Activations," presented on Thursday, February 12, 2026. The core problem addressed is the high and often wasteful cost of deploying large language models (LLMs) by forcing them to compute every token, even for simple queries. The research proposes a novel, cost-effective method using linear probing on the model's internal state immediately after processing the prompt to predict whether the subsequent generation will succeed or fail.