r-horizon-long-horizon-reasoning

Construct multi-step reasoning benchmarks with interdependent problems to evaluate and improve long-horizon reasoning in large reasoning models. Enables evaluation of reasoning depth and breadth beyon

SKILL.md

GitHub

R-Horizon: Evaluating and Improving Long-Horizon Reasoning

Large reasoning models show surprising limitations in extended reasoning chains. Current benchmarks focus on single-horizon tasks, missing critical evaluation of how well models handle multi-step interdependent problems or allocate thinking across multiple sub-tasks.

R-Horizon addresses this by constructing reasoning tasks where multiple problems depend on each other, creating true long-horizon reasoning scenarios. This reveals whether models can sustain reasoning quality across deep chains or whether they degrade as reasoning depth increases.

Core Concept

The framework constructs multi-step reasoning problems through query composition: building dependent sub-problems where solutions to earlier steps inform later steps. This naturally extends single-horizon reasoning benchmarks into long-horizon evaluation suites.

Key insight: models need to not only reason deeply but allocate their reasoning budget efficiently across multiple interconnected sub-problems.

Architecture Overview

  • Query Composition Engine: Takes base reasoning problems and creates dependency chains
  • Controlled Depth Variation: Generate problems at varying reasoning depths to test breadth and depth separately
  • Reward Integration: Connects with Reinforcement Learning with Verified Rewards (RLVR) for both evaluation and training
  • Benchmark Reusability: Creates persistent evaluation benchmarks across multiple reasoning models

Implementation Steps

The core workflow involves three stages: problem generation, composition, and evaluation.

Stage 1: Base Problem Selection

Start with high-quality reasoning problems. For mathematical reasoning, extract problems from benchmarks like AIME or Math competition datasets:

# Load base reasoning problems
base_problems = load_benchmark_problems(
    source='aime_2024',
    difficulty='high'
)

# Filter for composability
composable = [
    p for p in base_problems
    if can_create_dependency(p)
]

Stage 2: Dependency Graph Construction

Create chains where problem N+1 depends on solution to problem N. Build a dependency graph that represents the reasoning structure:

def create_composition_chain(problems, chain_length=5):
    """
    Build a chain of problems where each depends on previous solution.
    """
    chain = []
    for i in range(chain_length):
        base = problems[i]
        if i == 0:
            chain.append(base)
        else:
            # Modify problem to require solution from problem i-1
            dependent = inject_dependency(base, chain[i-1])
            chain.append(dependent)
    return chain

# Create multiple chain depth levels
depth_3_chains = create_composition_chain(problems, chain_length=3)
depth_5_chains = create_composition_chain(problems, chain_length=5)
depth_8_chains = create_composition_chain(problems, chain_length=8)

Stage 3: Evaluation with RLVR

Evaluate models on composed problems and use results to train reasoning improvement:

def evaluate_reasoning_chains(model, chains):
    """
    Test model on multi-step reasoning chains.
    Returns per-step accuracy to track reasoning degradation.
    """
    results = {
        'step_accuracy': [],
        'chain_completion': 0,
        'thinking_allocation': []
    }

    for chain in chains:
        step_results = []
        for step_idx, problem in enumerate(chain):
            response = model.generate(
                prompt=problem,
                max_thinking_tokens=8000
            )
            is_correct = verify_solution(response)
            step_results.append(is_correct)

        results['step_accuracy'].append(step_results)
        if all(step_results):
            results['chain_completion'] += 1

    return results

Practical Guidance

When to Use R-Horizon:

  • Evaluating whether large reasoning models truly handle long-horizon tasks
  • Creating curriculum-based training where models progress from single to multi-step reasoning
  • Identifying where reasoning models degrade in extended chains

When NOT to Use:

  • Single-step problem solving where long-horizon complexity adds no value
  • Domains without clear problem interdependencies (unrelated tasks)

Hyperparameter Considerations:

Parameter Typical Range Guidance
Chain Depth 3-10 Deeper chains reveal more limitations; 5-6 balances evaluation informativeness
Problem Difficulty Easy to Hard Mix difficulties to avoid ceiling/floor effects
Thinking Budget 4000-16000 tokens Allow sufficient reasoning tokens; 8000 is baseline
Composition Type Sequential/Graph Sequential most straightforward; graph captures realistic dependencies

Common Pitfalls:

  • Creating false dependencies that don't require genuine reasoning across steps
  • Using problems that are too easy, masking actual reasoning limitations
  • Insufficient thinking budget that prevents models from demonstrating capability
  • Not validating intermediate solutions, missing cascading errors

Reference

Based on the research at: https://arxiv.org/abs/2510.08189