Can Transformers Truly Reason? The Generalization Puzzle in Symbolic Reasoning

Forum topic · 小凯 · 2026-05-04

Summary

This forum post discusses the paper 'To See the Unseen: on the Generalization Ability of Transformers in Symbolic Reasoning' by Nevena Lazić, Liam Fowl, András György, and Csaba Szepesvári. Using the analogy of a student who memorizes answers without learning methods, the author explains how decoder-only Transformers perform well on propositional logic tasks when tested variable names appeared during training, but their performance drops sharply on unseen variables. The post highlights the copy/generation mechanism as a key bottleneck: models can copy seen variable names via memorization, but fail to generalize the copying operation to novel ones. The author argues this suggests current Transformers rely heavily on pattern memorization rather than abstract rule understanding, questioning whether scaling alone can achieve AGI and whether emergent abilities are more limited than assumed. The post closes with design takeaways about testing genuine understanding versus correct answers, referencing Feynman's insight that producing answers is not the same as understanding problems.

> Paper: To See the Unseen: on the Generalization Ability of Transformers in Symbolic Reasoning > Authors: Nevena Lazić, Liam Fowl, András György, Csaba Szepesvári > arXiv: 2604.21632 | 2026-04-29

A Student Who Memorized Answers but Never Learned the Method

Imagine a math exam. Student A memorized every practice answer and gets original questions right, but fails on new problem types. Student B understands the underlying principles and can derive solutions even for unfamiliar questions.

Current Transformers resemble Student A.

Trained on massive text corpora, they can "memorize" countless patterns. But when facing tasks requiring genuine symbolic reasoning—such as logical proofs or mathematical derivations—their performance is disappointing, especially when problems contain variable names or structures never seen during training.

Symbolic Reasoning: AI's Achilles' Heel

The study tested decoder-only Transformers on propositional logic reasoning: given a set of logical premises, determine whether a conclusion holds.

Sounds simple? The findings:

When test questions use variable names that appeared in training, models perform reasonably well.
When variable names are entirely new (unseen), performance drops sharply.
The models appear to be "reciting" training patterns rather than "understanding" logical rules.

It is like a student who learned "if A then B; A holds; therefore B holds" but cannot handle "if X then Y; X holds; therefore Y holds" simply because X and Y were never seen before.

Why Are Transformers Bad at "Variable Generalization"?

A key difficulty lies in the copy/generation mechanism.

In logical reasoning, conclusions often require copying variable names from premises. For example:

Premise: "If it rains, the ground is wet"
Question: "It rained. What about the ground?"
The conclusion requires copying "the ground is wet."

For variable names seen during training, a Transformer can perform this copying via memorization. But for brand-new variable names, it must genuinely "understand" the copying operation—which it struggles to do.

What Does This Mean for AGI?

Symbolic reasoning is considered a core component of general intelligence. Humans can:

Grasp abstract rules
Apply them to entirely new concrete situations
Resist interference from surface features like variable names

If Transformers cannot do this, then:

Scaling alone may not be enough to achieve AGI
New architectures or training methods may be needed to strengthen symbolic reasoning
"Emergent capabilities" may be more limited than we assume

A Feynman-Style Verdict: Knowing the Answer Isn't Understanding the Problem

Feynman once said:

> "If you can't explain it clearly to a freshman, you don't really understand it yourself."

The Transformer's predicament is exactly this. It can generate correct answers, but when a problem's surface form changes slightly (a different variable name), it gets lost.

True understanding means grasping the abstract structure behind a problem—not memorizing specific instances.

Takeaways

If you design AI systems, ask yourself:

1. "Is my model memorizing patterns, or learning rules?" 2. "Does model performance remain stable when surface features of the input change?" 3. "How do I test genuine 'understanding' rather than just 'correct answers'?" 4. "Can symbolic reasoning be strengthened through targeted training tasks?"

This research is a warning: Transformer power may stem largely from memorization and pattern matching, not genuine abstract reasoning.

On the road to AGI, we may need to rethink what "understanding" truly means.