TLDR: A new research paper introduces a framework using logic programming to evaluate the efficiency of language models (LMs) in deductive reasoning, beyond just correctness. By injecting irrelevant information into math word problems, the study found that LMs exhibit significant accuracy declines and generate proofs with frequent detours through unnecessary inferences. This inefficiency is particularly pronounced when irrelevant information semantically overlaps with the query, highlighting a need for LMs to better discern and ignore distractions.
Large language models (LMs) have shown impressive capabilities in deductive reasoning, solving a wide array of complex tasks. However, a new study suggests that while these models might often get to the correct answer, they are far from efficient in their reasoning process, especially when faced with irrelevant information.
Human-like reasoning often involves sifting through vast amounts of data and skillfully ignoring distractions to arrive at a conclusion. This efficiency, a crucial aspect of intelligence, has largely been overlooked in standard evaluations of LMs, which primarily focus on correctness.
The Efficiency Gap in AI Reasoning
Researchers propose a novel framework to assess how efficiently LMs reason, drawing insights from logic programming. Imagine a proof as a path in a complex network of information. The most efficient proof is simply the shortest path to the goal. The new method aligns the natural language proofs generated by LMs with these shortest, most direct proofs found through logic programming. This allows for a precise quantification of efficiency, specifically by measuring how well a model avoids unnecessary inferences.
To test this, a unique dataset was created using grade school math word problems. These problems were deliberately injected with varying amounts of irrelevant information, or ‘axioms,’ which also differed in how much they semantically overlapped with the problem’s actual goal. For instance, a problem asking about ‘Ryan’s cats’ might include irrelevant information also mentioning ‘Ryan’ or ‘cats,’ making it harder for the LM to distinguish relevant from irrelevant data.
Key Findings: Distractions Degrade Performance
The empirical results were striking. Current LMs experienced noticeable declines in accuracy when irrelevant axioms were introduced. This performance drop occurred even with minimal, domain-consistent distractions and worsened as more irrelevant information was added. The proofs generated by these models frequently took ‘detours’ through inferences that were not necessary to solve the problem.
One significant observation was that LMs were particularly inefficient when the irrelevant axioms had semantic overlap with the query. This suggests that while LMs might use lexical cues (like matching names or entities) as a heuristic to guide their search, this strategy can backfire when those cues lead to irrelevant paths.
Furthermore, the study found that LMs often generated more tokens than required to solve problems correctly. This indicates that inefficiency isn’t just about making irrelevant logical steps, but also about verbosity in expressing those steps in natural language. The efficiency scores, which measure the ratio of the shortest proof length to the LM’s proof length, were consistently far from 100%, confirming that models frequently produced irrelevant theorems even when they reached the correct final answer.
The research also compared performance on ‘non-ground’ queries (e.g., “How many drones does Yanick have?”) versus ‘ground’ queries (e.g., “Show that Yanick has 5 drones.”). LMs generally performed better on non-ground queries, which are more common in typical math word problems, suggesting a potential influence from their training data.
Also Read:
- New Benchmark Reveals LLMs’ Ongoing Struggle with Advanced High School Math
- Unpacking How AI Models Reason About Obligations and Permissions
A Call for More Efficient AI Reasoning
This work, titled “Are Language Models Efficient Reasoners? A Perspective from Logic Programming”, by Andreas Opedal, Yanick Zengaffinen, Haruki Shirakami, Clemente Pasti, Mrinmaya Sachan, Abulhair Saparov, Ryan Cotterell, and Bernhard Schölkopf, highlights a critical area for improvement in language models. It underscores the need for models that not only provide correct answers but also do so efficiently, by effectively identifying and ignoring irrelevant information. The framework presented offers a valuable tool for future research aimed at developing more human-like and resource-efficient AI reasoning systems.
