Research map / Large language models / Reasoning and chain of thought

Strategy Executability in Mathematical Reasoning: Leveraging Human–Model Differences for Effective Guidance

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that strategy usage in correct solutions does not predict whether a strategy helps as guidance (executability), and proposes Selective Strategy Retrieval (SSR), a test-time, multi-route, source-aware retrieval method that improves compact reasoning models on math benchmarks.

Problem

Example-based guidance for mathematical reasoning is unstable across problems and models even when guidance is correct and relevant; strategies that appear in successful solutions may not be executable when transferred as guidance to a target model.

Approach

Distinguishes strategy usage from strategy executability (model-relative success probability when a strategy is given as guidance), finds human and model solutions differ in executable strategies, and selects strategies via a learned executability predictor over a strategy knowledge graph with three retrieval routes.

Key results

• SSR improves Qwen3-8B on AIME25 from 67.33% (direct solving) to 74.00%, and on Apex from 8.16% to 13.06%.
• SSR yields the best accuracy across all three models and benchmarks in the main table, e.g., Qwen3-14B on HM-ReasoningBench 67.40 to 70.20.
• Removing the problem-transfer route causes the largest ablation drop, e.g., Qwen3-14B on Apex from 14.69 to 11.84.

Surprising findings

• Strategies frequently used in correct solutions do not necessarily yield higher accuracy when transferred as guidance (usage and executability dissociate).
• Providing more in-context examples did not help and sometimes degraded performance.

Limitations

• Evaluated only on compact reasoning models (Qwen3-8B, Qwen3-14B, R1-Distill-7B) in the main experiments with closed-source models only in the gain figure.
• Relies on GPT-5.1 for difficulty annotation, strategy extraction, and answer verification, introducing judge dependence.
• Executability predictor requires per-model empirical executions on a training split, so it must be re-estimated for each target model and protocol.

Experimental setup

Datasets: HM-ReasoningBench Omni-Math HARP

Benchmarks: HM-ReasoningBench AIME 2025 MathArena Apex

Metrics: accuracy pass@1

Baselines: Direct Solving In-Context Learning Human-Only Guidance Model-Only Guidance Self-Consistency Least-to-Most Prompting Tree-of-Thoughts

Builds on: in-context learning self-consistency least-to-most prompting tree-of-thoughts

Closest papers by meaning

• MARGE: Improving Math Reasoning with Guided Exploration (ICML 2025 · similarity 0.87)
• StrategyLLM: Large Language Models as Strategy Generators, Executors, Optimizers, and Evaluators for Problem Solving (NeurIPS 2024 · similarity 0.87)
• RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation (ICML 2025 · similarity 0.87)
• Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
• Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.87)
• Modeling Hierarchical Thinking in Large Reasoning Models (ICML 2026 · similarity 0.86)