Nothing from Something: Can a Language Model Discover 0?

Technical Summary: Nothing from Something: Can a Language Model Discover 0?

Problem Statement
The paper investigates the capacity of artificial neural networks to perform "out-of-distribution" generalization, specifically the ability to hypothesize and learn genuinely new mathematical structures not present in their training data. While recent large language models (LLMs) have achieved high performance on complex mathematical benchmarks (e.g., IMO, Putnam), these successes largely rely on training data that mirrors the complexity and content of the test problems. The authors posit that true mathematical discovery requires a system to bridge a conceptual gap—moving from a universe of concepts to one containing more powerful concepts. Using the introduction of the number "zero" as a case study, the paper asks: Can a model trained exclusively on positive single-digit arithmetic generalize to include zero, and does language pretraining facilitate this leap?

Methodology
The authors employ a controlled experimental setup using GPT-style decoder-only transformers (specifically GPT-2 sized architectures).

• Models: Two primary configurations were tested: a small model (4 layers, 128 embedding dimension, ~800K parameters) and a larger model (12 layers, 768 embedding dimension, ~124M parameters). These were evaluated under four conditions: (1) pre-trained on unfiltered OpenWebText, (2) pre-trained on a filtered OpenWebText corpus (stripped of numbers and mathematical symbols), (3) pre-trained on unfiltered data but without the specific math training, and (4) no pretraining at all.
• Data Construction:
  • Arithmetic Task: Models were trained on variable-length single-digit addition and subtraction problems (2–4 operands).
  • Zero-Holdout Strategy: The training and validation sets contained only problems where zero did not appear as an operand or a result (except in the ones place of multi-digit answers like "10"). The test set consisted entirely of problems requiring zero (e.g., "3-3=0" or "0+5=5").
  • Tokenization: A custom manual tokenization scheme was used where each digit and symbol is a distinct token. This ensures the model treats "0" as a specific token rather than part of a larger integer token.
  • Language Pretraining: To ensure the "filtered" models had no prior exposure to arithmetic, the authors curated a version of OpenWebText by removing all non-natural language characters, converting numbers to words, and filtering out mathematical symbols. Perplexity checks confirmed these models were unfamiliar with arithmetic tokens prior to the specific arithmetic training.
• Experimental Design:
  1. Zero-Shot Generalization: Models trained on positive-only arithmetic were tested on zero-inclusive problems.
  2. Few-Shot Learning: A small fraction of the training data (ranging from 1 to 1024 examples) was replaced with zero-inclusive problems to measure sample efficiency.
  3. Control Experiments: The zero-holdout procedure was repeated for digits 1–9 (and 1–7 in a base-8 setting) to determine if zero is uniquely difficult to generalize.

Key Results

1. Failure of Zero-Shot Generalization: Regardless of language pretraining or model size, GPT-2 sized models failed to generalize to zero when it was entirely absent from the training distribution. Even as training loss decreased and validation accuracy saturated, test accuracy on zero-inclusive problems remained at chance levels.
2. Impact of Few-Shot Data: When a small number of zero-inclusive examples were introduced into the training set, model performance improved significantly.
   • For the language-pretrained model, just 64 examples (0.64% of training data) yielded >60% accuracy, and 1024 examples (>90% accuracy).
   • Sample Efficiency: Language pretraining reduced the number of required few-shot examples by approximately 50% compared to models without pretraining. This suggests language abilities act as a scaffold for mathematical discovery.
3. Zero is Special: Generalization difficulty varied by digit. Digits in the middle of the range (e.g., 4, 5) were easier to generalize to than zero or the "carry" digit (9 in base-10, 7 in base-8).
   • Analysis of embedding vectors revealed that representations for 0 and 9 have fewer "near neighbors" (digits with high cosine similarity) compared to middle digits. This supports a hypothesis that models struggle to extrapolate to digits with sparse representational neighbors.
4. Scaling Limits: Experiments with larger, open-source 1-billion parameter models (included in the Appendix) also failed to discover zero at test time without explicit exposure, suggesting that scaling alone does not solve this specific generalization problem.

Significance and Claims
The paper claims to provide a rigorous benchmark for "conceptual generalization" in neural models, distinct from compositional generalization. The authors conclude that:

• Current GPT-2 sized models cannot spontaneously discover the concept of zero through arithmetic training alone; they require explicit exposure.
• Language pretraining does not enable zero-shot discovery but significantly accelerates the learning process once the concept is introduced, reducing the data requirement by roughly half.
• This finding parallels cognitive science theories (e.g., Carey, 2009) suggesting that language abilities scaffold the evolution of formal concepts in human cognition.
• The difficulty in generalizing to zero (and the carry digit) highlights a limitation in the models' ability to extrapolate beyond the "interpolation" of existing token representations.

The authors maintain a modest stance, noting that while these results illuminate the mechanisms of generalization in simple arithmetic, they do not yet demonstrate the ability of AI to independently generate complex, novel mathematical proofs. The work serves as a foundational step in understanding how to measure and potentially enable genuine mathematical discovery in AI systems.