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

R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 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.

Proposes R-Horizon, which composes single-horizon problems into dependent multi-step queries to both benchmark and RL-train large reasoning models for long-horizon reasoning.

Problem

Existing benchmarks and RL training data focus on isolated single-horizon problems, so the long-horizon reasoning ability of large reasoning models (LRMs) is under-evaluated and under-trained.

Approach

Query composition chains existing problems by making each problem's input depend on the previous answer (sequential dependency), yielding a scalable, controllable benchmark and RLVR training data with all-or-nothing scoring and theoretical accuracy estimated as the product of atomic pass rates.

Key results

• DeepSeek-R1 drops from 87.3% (n=1) to 24.6% (n=5) on composed AIME25.
• RLVR on 2-query composed data improves R1-Qwen-7B's composed AIME24 (n=2) by +17.4 points and original AIME24 by +7.5 points.
• R1-Qwen-7B drops from 93.6% (n=1) to 0% (n=16) on composed MATH500.

Surprising findings

• Models often answer later composed problems correctly despite wrong earlier answers, suggesting data contamination or shortcut behavior.
• Training with composed data improves performance on original single-problem benchmarks as well (+7.5 on AIME24).

Limitations

• Composed problems are synthetic and sequentially dependent by construction, which may not reflect real-world long-horizon tasks.
• Theoretical accuracy assumes independence between atomic problem pass rates.
• Evidence of anomalous correct answers after wrong predecessors raises concerns about contamination and metric validity.

Experimental setup

Datasets: Skywork-OR1-RL math data MATH500 AIME24 AIME25 AMC23 LiveCodeBench WebShaper

Benchmarks: R-Horizon MATH500 R-Horizon AIME24 R-Horizon AIME25 R-Horizon AMC23 R-Horizon LiveCodeBench R-Horizon WebShaper

Metrics: all-or-nothing accuracy avg@k accuracy theoretical expected accuracy pass@1

Baselines: DeepSeek-R1 Qwen3-235B-A22B-Thinking-2507 o4-mini R1-Qwen-7B Skywork-OR1-7B

Builds on: GRPO Skywork-OR1 WebShaper LiveCodeBench

Closest papers by meaning

• How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.89)
• h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning (ICML 2026 · similarity 0.89)
• Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (NeurIPS 2025 · similarity 0.89)
• RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.89)
• LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts (ICLR 2026 · similarity 0.88)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.88)