IC-275Mistral 7B Instruct exhibits a reasoning-type-dependent failure mode where certain problems are exclusively solvable by one non-deductive reasoning type

Danqing Wang, Jianxin Ma, Fei Fang, Lei Li

SourceTypedThinker: Diversify Large Language Model Reasoning with Typed Thinking

The paper prompts Mistral 7B Instruct with each of four reasoning types (deductive, inductive, abductive, analogical) on four benchmarks and samples up to 10 solutions per type at temperature 1. A problem is counted as exclusively solvable by a type if at least one solution under that type is correct but no solution under any other type is. Figure 1 shows that on every benchmark, a non-zero percentage of problems fall into this category for at least one non-deductive type, meaning the model cannot solve them regardless of how many times it is sampled with the wrong reasoning type. The authors note that the problems solvable by different types do not completely overlap, and that repeated sampling with high temperature does not overcome the limitation when the wrong type is used.

Evidence
correlational
Key metric
Figure 1: percentage of problems solvable exclusively by one reasoning type, non-zero for inductive, analogical, and abductive on LogiQA, BBH, GSM8K, and MATH; e.g. on MATH a visible share is exclusively inductive
Caveat
The analysis is performed only on Mistral 7B Instruct; the same exclusive-solvability analysis is not repeated for Llama3 8B, Qwen 2 7B, GPT-4o, or MetaMath. The sampling budget is limited to 10 solutions per type.
Model
Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral-7B-Instruct
Concepts
Failure mode
Datasets
LogiQA [eval], GSM8K [eval]
Extraction
automatic-extraction