From Reasoning Depth to Reasoning Breadth: Evaluating Multi-Point Associative Reasoning in Large Language Models

2026-08-10

AI Summary

This work addresses a critical gap in evaluating large language models: while recent advances have enhanced reasoning depth, systematic assessment of “reasoning breadth”—the ability to concurrently explore multiple semantic directions and integrate diverse cues—remains absent. The study formalizes the concept of reasoning formulate a novel multi-clue target recovery task inspired by cooperative games and introduces MPAR-Bench, a bilingual benchmark comprising 1,000 newly constructed problems generated via multi-agent collaboration, diversity filtering, and human verification. Evaluation employs ANLS scores, embedding similarity, reasoning trajectory validation, and four perturbation types (clue masking, sequence shuffling, distractor injection, and multi-step clues). Results reveal that current models suffer accuracy drops of 5–18 percentage points under perturbations; chain-of-thought improves baseline performance but not robustness, and excessive depth can even overturn correct initial hypotheses, demonstrating that reasoning depth does not equate to reasoning breadth.

Abstract

Large language models (LLMs) have made substantial progress on reasoning tasks that require increasingly long and complex inferential chains. This progress primarily reflects reasoning depth. A complementary and comparatively unexamined capability is reasoning breadth: exploring multiple semantic directions in parallel and integrating the resulting clues into one coherent answer. We introduce MPAR-Bench, a bilingual English-Chinese benchmark that isolates reasoning breadth through multi-point associative reasoning. Inspired by the cooperative game Just One, each item asks a model to recover a hidden target from several independently generated, semantically diverse clues. We construct 1,000 items using a multi-agent clue-generation pipeline, embedding-based diversity filtering, and human verification. Only the answer space is drawn from public word lists, whereas every clue set is generated from scratch. Beyond exact-match accuracy, we evaluate models using accuracy, ANLS, embedding similarity, reasoning-trace verification, and four perturbations: clue masking, order shuffling, distractor injection, and multi-step clues. Across evaluated models, perturbations reduce accuracy by 9-18 percentage points in English and 5-12 percentage points in Chinese. Thinking mode improves standard-setting accuracy, especially in English, but does not consistently reduce sensitivity to perturbations. Case-level analysis also shows that extended reasoning can overturn an initially correct hypothesis. These results indicate that greater reasoning depth does not automatically confer robust reasoning breadth, and that reasoning breadth remains largely uncovered by current benchmarks.

Problem

• reasoning breadth
• multi-point associative reasoning
• large language models
• reasoning evaluation
• semantic integration

Innovation

• reasoning breadth
• multi-point associative reasoning
• MPAR-Bench
• perturbation robustness
• large language models