Least-to-Most Prompt Strategy

• The heuristic Least-to-Most (LTM) strategy is an iterative prompting method that decomposes complex problems into ordered subtasks, enhancing reasoning and generalization.
• It employs an answer-sensitivity mechanism with calibrated confidence thresholds to decide timely exits and reduce error propagation.
• LTM consistently outperforms one-shot approaches in tasks like phishing detection and semantic parsing by integrating intermediate subtask solutions.

The heuristic Least-to-Most (LTM) prompt strategy is an iterative prompting technique developed for LLMs, designed to enhance reasoning and generalization in complex, multi-step tasks. LTM operates by decomposing a global problem into a sequence of subtasks ordered from least to most difficult, prompting the model on each subproblem sequentially, and progressively integrating intermediate outputs to reach a robust final prediction. Extensions of LTM, notably the answer-sensitivity mechanism, have further improved practical accuracy on challenging domains such as phishing detection and compositional semantic parsing, consistently outperforming one-shot and even supervised baselines when applied judiciously (Trikilis et al., 28 Jan 2026, Zhou et al., 2022, Arora et al., 2023).

1. Definition and Formal Structure

The core goal of the LTM prompting framework is to solve a complex problem $P$ by decomposing it into an ordered list of subproblems $\{Q_1, \ldots, Q_k\}$, such that $Q_1$ is the most atomic and tractable, while $Q_k$ encapsulates the highest level of abstraction or inference. The strategy proceeds as follows (Zhou et al., 2022, Trikilis et al., 28 Jan 2026):

1. Decomposition: Construct $\{Q_1, \ldots, Q_k\}$ using domain-specific heuristics, generally ordering from low-hanging, easily-resolved checks to deeper, more ambiguous inference steps.
2. Iterative Prompting: For $i = 1$ to $k$, prompt the LLM with $Q_i$, optionally conditioned on answers to previous subtasks $\{A_1, \ldots, A_{i-1}\}$.
3. Stopping Criteria and Aggregation: Employ an answer-sensitivity mechanism (detailed below) to determine, after each subtask, whether the global problem $P$ can already be decisively resolved. If not, continue until all subtasks are completed or an early stop condition is met.
4. Final Synthesis: Aggregate the sequence of sub-answers $\{Q_1, \ldots, Q_k\}$0 into the final solution to $\{Q_1, \ldots, Q_k\}$1.

In mathematical notation, the process can be formalized as: $\{Q_1, \ldots, Q_k\}$2 with the final answer obtained by aggregating $\{Q_1, \ldots, Q_k\}$3 (Zhou et al., 2022).

2. Answer-Sensitivity Mechanism

To enable efficient decision-making at each subtask, the strategy augments sub-answers with calibrated confidence scores $\{Q_1, \ldots, Q_k\}$4. The answer-sensitivity mechanism defines two thresholds, a lower bound $\{Q_1, \ldots, Q_k\}$5 and an upper bound $\{Q_1, \ldots, Q_k\}$6, providing a formal basis for early resolution:

• If $\{Q_1, \ldots, Q_k\}$7, conclude "negative" (e.g., benign URL) with confidence $\{Q_1, \ldots, Q_k\}$8.
• If $\{Q_1, \ldots, Q_k\}$9, conclude "positive" (e.g., phishing) with confidence $Q_1$0.
• If $Q_1$1, proceed to the next subtask.

The mechanism is encapsulated in the following pseudocode (Trikilis et al., 28 Jan 2026):

$Q_1$9

This process mitigates both undecided loops and premature conclusions by setting $Q_1$2 as a strict iteration cap and defaulting to a conservative class if undecided.

3. Heuristics for Decomposition, Granularity, and Prompt Design

Effective application of LTM depends critically on the quality and ordering of subtasks. The following heuristics guide this process (Zhou et al., 2022, Arora et al., 2023, Trikilis et al., 28 Jan 2026):

• Granularity: Each subproblem $Q_1$3 should represent a single atomic attribute or property, avoiding both over-broad and too finely split queries. Typical subtasks number between 5 and 10.
• Sequencing: Order subtasks to eliminate easily classifiable instances early, reserving more resource-intensive or global checks for ambiguous cases.
• Decomposition heuristics: Common criteria include dependency analysis (topological order on a DAG of concepts), compositional structure analysis, symbolic parsing, rule-based templates, and syntactic cue identification.
• Prompt framing: Each prompt restates the top-level goal, summarizes prior answers, and explicitly poses $Q_1$4 in plain language while requesting a structured fixed-format answer (including the confidence score).
• Domain adaptation: Subtasks may be swapped or rephrased to reflect the specificities of the application domain (e.g., replacing "Check for suspicious TLD" with "Check for unusual chemical name").

4. Instantiations Across Domains and Empirical Results

Tabulated Overview of Instantiations

Paper / Task Domain	LTM Variant	Main Result
Phishing URL Detection (Trikilis et al., 28 Jan 2026)	LTM + Answer Sensitivity	Outperforms one-shot, matches supervised, needs less data
Symbolic Manipulation, GSM8K, SCAN (Zhou et al., 2022)	L2M prompting	Solves SCAN at 99% accuracy (vs 16% for CoT, 15k for seq2seq)
Text-to-SQL Parsing (Arora et al., 2023)	LTMP-DA-GP	+6–15 pts over generic prompt; matches/exceeds supervised

LTM = Least-to-Most; CoT = Chain-of-Thought; LTMP-DA-GP = Least-to-Most Prompting with Domain Adapted Generic Prompt.

• In phishing URL detection (Trikilis et al., 28 Jan 2026), LTM with answer sensitivity not only surpasses one-shot baselines but also achieves performance on par with supervised models using a dramatically reduced sample size. The iterative process, guided by confidence thresholds, allowed early exit for easy cases and a detailed drilldown for ambiguous URLs.
• For compositional generalization tasks such as SCAN, L2M prompted GPT-3 models to 99% accuracy across all splits with only 14 exemplars, drastically outperforming both chain-of-thought and fully supervised neural-symbolic models (Zhou et al., 2022).
• In Text-to-SQL parsing, adoption of an "Adapt-and-Decompose" pipeline, which unifies offline domain adaptation and decomposed LTM prompting, yields the highest cross-domain and cross-compositional generalization on the KaggleDBQA benchmark, with up to 38% execution accuracy—surpassing both zero-shot and existing few-shot baselines, and matching state-of-the-art supervised approaches (Arora et al., 2023).

5. Prompt Templates and Practical Implementation

The LTM strategy is supported by generalized prompt templates across varying task modalities. Key components included (Trikilis et al., 28 Jan 2026, Zhou et al., 2022):

• Binary classification: Sequential subquestions such as "Are there any explicit slurs or profanity?" with fixed-format answer plus confidence.
• Reasoning-heavy tasks: Decomposed math word problems, each substep explicitly stated and answered before moving forward.
• Multi-class labeling: Confidence vectors across classes; select the class when confidence exceeds the upper threshold.

The procedure extends generic prompt engineering by emphasizing clear, atomic subquestions, structured answer formats, and failsafes (iteration caps, template-based confidence reporting), which can be rapidly tailored to new domains with minimal required in-domain tuning or additional supervision.

6. Performance Guidelines and Adaptation Best Practices

Empirical and procedural guidelines for deploying LTM include (Zhou et al., 2022, Arora et al., 2023, Trikilis et al., 28 Jan 2026):

• Calibration: Thresholds $Q_1$5 and $Q_1$6 require calibration (e.g., $Q_1$7), typically on held-out data. Confidence normalization may be necessary if LLM outputs are poorly calibrated.
• Prompt length and latency: Subtask count should be limited (usually $Q_1$8) to balance latency and completeness.
• Hybrid fallback: If oscillatory or indecisive answers are detected, the pipeline can be configured to fallback to a one-shot or simpler classifier.
• Monitoring: Logging complete subtask histories is essential for error analysis, prompt refinement, and debugging.
• Domain adaptation: Prompt exemplars and subtasks should be customized for domain specifics; schema and data-type descriptions increase generalizability in structured tasks.
• Offline adaptation and universality: Techniques such as submodular set-cover for exemplar selection and domain adaptation increase scalability and universality, enabling strong performance under token or exemplarity constraints (Arora et al., 2023).

7. Relation to Other Prompting Strategies and Impact

LTM generalizes and extends chain-of-thought (CoT) prompting by ensuring that each step is directly supported by previous outputs and by enforcing a curriculum from simple to complex. In empirical studies, LTM consistently closes gaps that exist with standard CoT on compositional and cross-domain generalization, limits error propagation by atomic inspection at each step, and offers robust early-exit mechanisms through answer sensitivity (Zhou et al., 2022, Trikilis et al., 28 Jan 2026).

This suggests that LTM and its heuristic instantiations provide a practical, theoretically sound approach to leveraging LLMs for systematic multi-step reasoning, achieving high accuracy and generalizability even in domains with severely limited annotation resources.