When LLMs Stop Following Steps: A Diagnostic Study of Procedural Execution in Language Models

• URL: http://arxiv.org/abs/2605.00817v1
• Date: Fri, 01 May 2026 17:55:47 GMT
• Title: When LLMs Stop Following Steps: A Diagnostic Study of Procedural Execution in Language Models
• Abstract summary: Large language models (LLMs) often achieve strong performance on reasoning benchmarks, but final-answer accuracy alone does not show whether they faithfully execute the procedure specified in a prompt.<n>We study this question through a controlled diagnostic benchmark for procedural execution, where models are given a step-wise arithmetic algorithm and two numeric inputs, and must return the final computed value.<n>Across 14 models and 55 datasets, average first-answer accuracy drops from 61% on 5-step procedures to 20% on 95-step procedures.<n>Generation-level analysis shows that failures often involve missing answers, premature answers, self-executed after an initial error, under-
• Score: 0.8566115760417291
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models (LLMs) often achieve strong performance on reasoning benchmarks, but final-answer accuracy alone does not show whether they faithfully execute the procedure specified in a prompt. We study this question through a controlled diagnostic benchmark for procedural execution, where models are given a step-wise arithmetic algorithm and two numeric inputs, and must return the final computed value. The benchmark uses simple arithmetic operations but increases complexity through algorithm length and look-back dependencies over intermediate variables. Across 14 models and 55 datasets, average first-answer accuracy drops from 61% on 5-step procedures to 20% on 95-step procedures. Generation-level analysis shows that failures often involve missing answers, premature answers, self-correction after an initial error, under-executed traces, and hallucinated extra steps. These findings suggest that apparent reasoning ability can mask substantial weaknesses in faithful instruction execution.

Related papers

• Evaluating Prompting and Execution-Based Methods for Deterministic Computation in LLMs [1.2957535240267326]
  Large Language Models (LLMs) have demonstrated strong capabilities in natural language understanding and reasoning.<n>We systematically evaluate multiple prompting strategies, including Chain-of-Thought (CoT), Least-to-Most decomposition, Program-of-Thought (PoT), and Self-Consistency (SC)<n>Our results show that standard prompting methods achieve only moderate accuracy on sequence-based tasks.
  arXiv  Detail & Related papers  (2026-05-04T23:32:37Z)
• AgentProcessBench: Diagnosing Step-Level Process Quality in Tool-Using Agents [50.481033105867205]
  We introduce AgentProcessBench, the first benchmark dedicated to evaluating step-level effectiveness in realistic, tool-augmented trajectories.<n>The benchmark comprises 1,000 diverse trajectories and 8,509 human-labeled step annotations with 89.1% inter-annotator agreement.<n>It features a ternary labeling scheme to capture exploration and an error propagation rule to reduce labeling ambiguity.
  arXiv  Detail & Related papers  (2026-03-15T16:13:58Z)
• Solver-in-the-Loop: MDP-Based Benchmarks for Self-Correction and Behavioral Rationality in Operations Research [19.31559944205485]
  Operations Research practitioners routinely debug infeasible models through an iterative process.<n>We introduce two benchmarks that place the textbfsolver in the evaluation loop<n>We find that domain-specific RLVR training enables an 8B model to surpass frontier APIs.
  arXiv  Detail & Related papers  (2026-01-28T20:02:44Z)
• Can We Predict Before Executing Machine Learning Agents? [74.39460101251792]
  We formalize the task of Data-centric Solution Preference and construct a comprehensive corpus of 18,438 pairwise comparisons.<n>We demonstrate that LLMs exhibit significant predictive capabilities when primed with a Verified Data Analysis Report.<n>We instantiate this framework in FOREAGENT, an agent that employs a Predict-then-Verify loop, achieving a 6x acceleration in convergence while surpassing execution-based baselines by +6%.
  arXiv  Detail & Related papers  (2026-01-09T16:44:17Z)
• Probing for Arithmetic Errors in Language Models [86.8227317662622]
  Internal activations in language models can be used to detect arithmetic errors.<n>We show that simple probes can accurately decode both the model's predicted output and the correct answer from hidden states.<n>We train lightweight error detectors that predict model correctness with over 90% accuracy.
  arXiv  Detail & Related papers  (2025-07-16T16:27:50Z)
• Dynamic Early Exit in Reasoning Models [21.30793518631921]
  Overthinking in long chain-of-thought (CoT) generation slows down the efficiency of problem solving, but also risks accuracy loss.<n>We propose a simple yet effective method that allows LLMs to self-truncate CoT sequences by early exit during generation.<n>Our method requires no additional training and can be seamlessly integrated into existing o1-like reasoning LLMs.
  arXiv  Detail & Related papers  (2025-04-22T13:36:53Z)
• Do Large Language Models Truly Grasp Addition? A Rule-Focused Diagnostic Using Two-Integer Arithmetic [21.014229380679975]
  Large language models (LLMs) achieve impressive results on advanced mathematics benchmarks but sometimes fail on basic arithmetic tasks.<n>We investigate whether they have truly grasped fundamental arithmetic rules or are merely relying on pattern matching.<n>Our evaluation of 12 leading LLMs reveals a stark disconnect: while models achieve high numeric accuracy, they systematically fail these diagnostics.
  arXiv  Detail & Related papers  (2025-04-07T16:57:10Z)
• EquiBench: Benchmarking Large Language Models' Reasoning about Program Semantics via Equivalence Checking [58.15568681219339]
  We introduce EquiBench, a new benchmark for evaluating large language models (LLMs)<n>This task directly tests a model's ability to reason about program semantics.<n>We evaluate 19 state-of-the-art LLMs and find that in the most challenging categories, the best accuracies are 63.8% and 76.2%, only modestly above the 50% random baseline.
  arXiv  Detail & Related papers  (2025-02-18T02:54:25Z)
• BoostStep: Boosting mathematical capability of Large Language Models via improved single-step reasoning [83.03531832811386]
  BoostStep is a method that enhances reasoning accuracy through step-aligned ICL examples.<n>It integrates seamlessly with chain-of-thought (CoT) and tree search algorithms.<n>It improves DeepSeek-R1-671B's performance on AIME by 2.2%, leveraging simple examples only from the MATH dataset.
  arXiv  Detail & Related papers  (2025-01-06T18:59:13Z)
• Enhancing Mathematical Reasoning in LLMs by Stepwise Correction [39.67266805233599]
  Best-of-N decoding methods instruct large language models (LLMs) to generate multiple solutions, score each using a scoring function, and select the highest scored as the final answer to mathematical reasoning problems. We propose a novel prompting method named Stepwise Correction (StepCo) that helps LLMs identify and revise incorrect steps in their generated reasoning paths. The verify-then-revise process not only improves answer correctness but also reduces token consumption with fewer paths needed to generate.
  arXiv  Detail & Related papers  (2024-10-16T18:18:42Z)
• Improve Mathematical Reasoning in Language Models by Automated Process Supervision [23.807288360423193]
  We propose a novel divide-and-conquer style Monte Carlo Tree Search (MCTS) algorithm named textitOmegaPRM for the efficient collection of high-quality process supervision data.<n>We are able to collect over 1.5 million process supervision annotations to train Process Reward Models (PRMs)<n>This fully automated process supervision alongside the weighted self-consistency algorithm is able to enhance LLMs' math reasoning performances.
  arXiv  Detail & Related papers  (2024-06-05T19:25:40Z)
• Reverse That Number! Decoding Order Matters in Arithmetic Learning [49.5504492920404]
  Our work introduces a novel strategy that reevaluates the digit order by prioritizing output from the least significant digit. Compared to the previous state-of-the-art (SOTA) method, our findings reveal an overall improvement of in accuracy while requiring only a third of the tokens typically used during training.
  arXiv  Detail & Related papers  (2024-03-09T09:04:53Z)
• Evaluating and Improving Tool-Augmented Computation-Intensive Math Reasoning [75.74103236299477]
  Chain-of-thought prompting(CoT) and tool augmentation have been validated as effective practices for improving large language models. We propose a new approach that can deliberate the reasoning steps with tool interfaces, namely textbfDELI. Experimental results on CARP and six other datasets show that the proposed DELI mostly outperforms competitive baselines.
  arXiv  Detail & Related papers  (2023-06-04T17:02:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.