Research map / Large language models / Retrieval and knowledge

Information Re-Organization Improves Reasoning in Large Language Models

NeurIPS 2024 · Large language models › Retrieval and knowledge · approach: Knowledge-augmented question answering · 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 InfoRE, which extracts a MindMap of logical relations from context, prunes noisy relations with an RL-trained BERT policy (PPO), and reasons over the re-organized context in a zero-shot setting.

Problem

LLMs often fail at context-aware multi-hop reasoning because they do not first identify logical relationships in the context; existing methods focus on refining the reasoning chain rather than the input context.

Approach

Before reasoning, re-organize the context into a MindMap-style structure of logical relations and multi-hop connections using an LLM, then prune irrelevant relations with an RL-trained BERT pruner (PPO, F1 reward), and reason over the pruned structure zero-shot.

Key results

• InfoRE with GPT-4 raises 2WikiMultiHopQA F1 from 72.69 to 76.52 zero-shot, and FEVEROUS accuracy from 92.33 to 95.62.
• Average absolute improvement of about 4% across tasks in zero-shot setting, per the abstract.
• Ablation on 2WikiMultiHopQA with GPT-3.5: removing extraction drops F1 from 64.58 to 61.64; removing pruning drops it to 63.05.

Measured effects reported in the paper

Surprising findings

• Using GPT-3.5 to re-organize context for GPT-4 reasoning reduced performance, suggesting re-organization quality matters more for stronger reasoners.

Limitations

• Limited to MindMap structure; other structures such as tables or timelines are not explored.
• Re-organization relies on large language models, so smaller models are not tested.
• No error bars reported due to computational cost.

Experimental setup

Datasets: HOVER FEVEROUS SCIFACT 2WikiMultiHopQA StrategyQA MuSiQue HotpotQA WikiHop

Benchmarks: HOVER FEVEROUS SCIFACT 2WikiMultiHopQA StrategyQA MuSiQue HotpotQA WikiHop

Metrics: F1 accuracy average absolute improvement

Baselines: Standard prompting CoT

Builds on: chain-of-thought proximal policy optimization BERT MindMap

Closest papers by meaning

• A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.89)
• Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models (ICML 2026 · similarity 0.89)
• Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models (ICLR 2026 · similarity 0.89)
• Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.89)
• Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.88)
• BABILong: Testing the Limits of LLMs with Long Context Reasoning-in-a-Haystack (NeurIPS 2024 · similarity 0.88)