[Paper Review] Large Language Models are Algorithmically Blind

Sohan Venkatesh, Ashish Mahendran Kurapath+1 more|arXiv (Cornell University)|Feb 25, 2026

Artificial Intelligence in Healthcare and Education0 citations

TL;DR

The paper evaluates eight frontier LLMs and finds they largely fail to provide calibrated predictions of algorithm performance in causal discovery, exhibiting algorithmic blindness and often performing no better than random baselines.

ABSTRACT

Large language models (LLMs) demonstrate remarkable breadth of knowledge, yet their ability to reason about computational processes remains poorly understood. Closing this gap matters for practitioners who rely on LLMs to guide algorithm selection and deployment. We address this limitation using causal discovery as a testbed and evaluate eight frontier LLMs against ground truth derived from large-scale algorithm executions and find systematic, near-total failure. Models produce ranges far wider than true confidence intervals yet still fail to contain the true algorithmic mean in the majority of instances; most perform worse than random guessing and the marginal above-random performance of the best model is most consistent with benchmark memorization rather than principled reasoning. We term this failure algorithmic blindness and argue it reflects a fundamental gap between declarative knowledge about algorithms and calibrated procedural prediction.

Motivation & Objective

• Assess whether frontier LLMs can predict algorithm performance in causal discovery tasks with calibrated uncertainty.
• Quantify calibration via interval coverage between LLM-predicted ranges and empirical ground truth.
• Separate memorization effects from genuine reasoning by testing on benchmark and synthetic datasets.

Proposed method

• Compute ground-truth algorithm performance by running 5,200 causal-discovery experiments (13 datasets × 4 algorithms × 100 runs) and derive empirical means with bootstrap CIs.
• Query 8 frontier LLMs with 3 prompt formulations per condition to elicit predicted performance ranges for four metrics.
• Aggregate predictions across prompt formulations and evaluate calibrated coverage against ground truth.
• Compare LLMs to random and heuristic baselines to assess added value.
• Analyze prompt sensitivity via coefficient of variation across prompts and examine dataset type (benchmark vs synthetic) effects.
• Investigate memorization signals through range width, cross-model agreement, and algorithm–metric interactions.

Figure 1: Comparison of LLM estimates and algorithmic ground truth revealing algorithmic blindness.

Experimental results

Research questions

• RQ1 Can frontier LLMs provide calibrated interval estimates of causal-discovery algorithm performance when given problem structure?
• RQ2 Do LLMs rely on principled reasoning or memorized benchmark statistics to predict algorithmic performance?
• RQ3 How do LLM predictions differ between benchmark and synthetic datasets, and across algorithms and metrics?

Key findings

• Mean calibrated coverage across 1,664 comparisons is 15.9%, with seven of eight models below random guessing.
• Claude is the best performer with 39.4% coverage, yet still marginally above random (36.5%).
• Seven models fall below the random baseline; the best model’s marginal advantage is attributed to memorization rather than reasoning.
• Calibration widths (predicted ranges) are 8–27 times wider than true confidence intervals, yet coverage remains poor.
• Synthetic data show substantial coverage degradation and greater cross-model disagreement, indicating memorization effects rather than principled generalization.
• Algorithm–metric interactions and range-width compression provide evidence of retrieval of benchmark statistics rather than structure-conditioned understanding.

Figure 2: Methodology overview. LLMs are prompted with dataset characteristics and algorithmic assumptions to predict performance metric ranges (Precision, Recall, F1, SHD). Ground-truth metrics with bootstrap 95% confidence intervals are computed via large-scale executions and calibration is evalua