On the Bias of Next-Token Predictors Toward Systematically Inefficient Reasoning: A Shortest-Path Case Study
By: Riccardo Alberghi , Elizaveta Demyanenko , Luca Biggio and more
Potential Business Impact:
Makes AI better at solving problems by learning from mistakes.
Recent advances in natural language processing highlight two key factors for improving reasoning in large language models (LLMs): (i) allocating more test-time compute tends to help on harder problems but often introduces redundancy in the reasoning trace, and (ii) compute is most effective when reasoning is systematic and incremental, forming structured chains of thought (CoTs) akin to human problem-solving. To study these factors in isolation, we introduce a controlled setting based on shortest-path tasks in layered graphs. We train decoder-only transformers on question-trace-answer triples using a custom tokenizer, comparing models trained on optimal bottom-up dynamic programming traces with those trained on longer, valid traces involving backtracking. Surprisingly, with the same training-token budget, models trained on inefficient traces generalize better to unseen graphs. This benefit is not due to length alone-injecting arbitrary redundancy into reasoning traces fails to help and can even hurt performance. Instead, we find that generalization correlates with the model's confidence in next-token prediction, suggesting that long, coherent, and locally incremental traces make the training signal easier to optimize.
Similar Papers
Short-Path Prompting in LLMs: Analyzing Reasoning Instability and Solutions for Robust Performance
Computation and Language
Makes AI think better even with short questions.
Reasoning Bias of Next Token Prediction Training
Computation and Language
Makes computers better at thinking by training them differently.
Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens
Machine Learning (CS)
Teaches computers to think better, not just copy.