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

2026-02-10

AI Summary

This work addresses the challenge of predicting the success probability of large language models on mathematical and programming tasks without performing full inference, thereby enabling more efficient allocation of computational resources. The authors propose a method that leverages internal activation states from early in the model’s generation process, using lightweight linear probes to extract task difficulty signals encoded in these activations. These signals exhibit strong correlation with actual model performance and diverge significantly from human-perceived difficulty. Building upon this insight, the study introduces a dynamic routing mechanism across multiple models. Evaluated under the E2H-AMC framework on the MATH dataset, the approach achieves up to 70% reduction in inference cost while surpassing the overall performance of the best single model, providing the first empirical validation of internal difficulty signals for effective inference scheduling.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Running LLMs with extended reasoning on every problem is expensive, but determining which inputs actually require additional compute remains challenging. We investigate whether their own likelihood of success is recoverable from their internal representations before generation, and if this signal can guide more efficient inference. We train linear probes on pre-generation activations to predict policy-specific success on math and coding tasks, substantially outperforming surface features such as question length and TF-IDF. Using E2H-AMC, which provides both human and model performance on identical problems, we show that models encode a model-specific notion of difficulty that is distinct from human difficulty, and that this distinction increases with extended reasoning. Leveraging these probes, we demonstrate that routing queries across a pool of models can exceed the best-performing model whilst reducing inference cost by up to 70\% on MATH, showing that internal representations enable practical efficiency gains even when they diverge from human intuitions about difficulty. Our code is available at: https://github.com/KabakaWilliam/llms_know_difficulty

Problem

• LLM efficiency
• difficulty prediction
• pre-generation activation
• inference cost
• model routing

Innovation

• pre-generation activations
• linear probing
• model-specific difficulty
• efficient inference
• query routing