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

ReJump: A Tree-Jump Representation for Analyzing and Improving LLM Reasoning

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.

Proposes ReJump, a tree-jump representation of LLM reasoning traces (tree of partial solutions plus visitation order with calc/verify/backtrack transitions), an LLM-based extractor (TreeExtractor, Gemini 2.5 Pro) and metrics for exploration, overthinking, forgetting and verification, applied to analyze LRMs and to improve test-time selection.

Problem

Existing comparisons of large reasoning models focus on final accuracy, missing differences in reasoning behavior such as exploration, exploitation, overthinking, forgetting and verification. A structured representation is needed to analyze and compare reasoning traces.

Approach

Represent a reasoning trace as a visitation order over nodes of a tree of partial solutions, where transitions ("jumps") may be between non-adjacent nodes (backtrack, verify), and define tree- and jump-level metrics; an LLM agent extracts this representation from raw CoTs and the metrics are used both for analysis and for selecting among sampled outputs or prompts.

Key results

• TreeExtractor with Gemini 2.5 Pro reaches tree similarity 0.943 and jump similarity 0.940 against 82 ground-truth ReJump representations on Game of 24.
• Human evaluation of extracted ReJump on MATH500 gives 81% (DeepSeek-R1) and 80% (QwQ-32B) accuracy at pass@1, above 90% at pass@3.
• On Game of 24, Best-of-N selection by highest jump distance raises QwQ-32B pass@1 from 0.76 (majority vote) to 0.82 and Phi-4-reasoning-plus from 0.77 to 0.84.

Surprising findings

• LRMs achieve higher pass@1 than CoT-prompted LLMs mainly by generating more distinct solutions, with similar per-attempt success rates.
• Increasing in-context reasoning examples raises jump similarity to the teacher but does not consistently improve accuracy, and tree similarity shows no clear trend.

Limitations

• Extraction requires a separate capable LLM per trace, making it costly and slow for training-time analysis.
• Tree and jump similarity ignore semantic content and temporal ordering of actions.
• Requires task-specific definition of partial solutions in prompts.

Experimental setup

Datasets: MATH-500 Game of 24

Benchmarks: MATH-500 Game of 24

Metrics: pass@1 tree similarity jump similarity solution count jump distance success rate verification rate overthinking rate forgetting rate human accuracy of extracted ReJump

Baselines: Landscape of Thoughts Majority vote CoT-prompted LLMs

Builds on: Chain-of-Thought prompting Tree-of-Thought Zhang-Shasha tree edit distance DAPO RLVR

Closest papers by meaning

• Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.91)
• Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.90)
• GLoRe: When, Where, and How to Improve LLM Reasoning via Global and Local Refinements (ICML 2024 · similarity 0.89)
• RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.89)
• RE-IMAGINE: Symbolic Benchmark Synthesis for Reasoning Evaluation (ICML 2025 · similarity 0.89)
• $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.89)