Research map / Large language models / Reasoning and chain of thought

Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Trace

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces PITA, a 23M-statement Lean-formalized propositional logic dataset, and a transitive inference toy task, showing reasoning-trace models generalize better on broad/shallow splits but worse on narrow/deep splits than direct-prediction models.

Problem

Understanding how reasoning traces support reasoning and where they fail, specifically length generalization as a function of task depth and breadth.

Approach

Defines task depth (number of unique proof states) and breadth (number of unique statements per size) as topology measures; compares RT (proof-trained) vs direct prediction models on PITA splits and a minimal transitive inference task with a max-margin theoretical analysis.

Key results

• RT models outperform DP models on breadth-dominated splits (Full, Imply), particularly for larger models.
• DP models outperform RT models on depth-dominated PHP split by as much as 50 percentage points for the 7B model.
• In the transitive inference task, DP generalization accuracy decays as B^-2 while RT generalization remains flat.

Surprising findings

• RT models underperform direct prediction on deep, narrow splits despite step-wise structure favoring reasoning traces.
• Failure to terminate accounts for a large proportion of RT errors.

Limitations

• Largest trainable model is 7B for PHP and 32B for other splits due to compute constraints.
• Transitive inference analysis uses a simplified single-layer model and ignores autoregressive failures.
• Synthetic propositional logic may not transfer to natural language reasoning.

Experimental setup

Datasets: PITA

Benchmarks: PITA splits (Full, Imply, Or, PHP) transitive inference task

Metrics: generalization accuracy

Baselines: direct prediction models

Builds on: chain-of-thought prompting QLoRA focused proof search muP

Closest papers by meaning

• Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens (NeurIPS 2026 · similarity 0.91)
• On the Bias of Next-Token Predictors Toward Systematically Inefficient Reasoning: A Shortest-Path Case Study (NeurIPS 2025 · similarity 0.90)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.90)
• Reasoning with Latent Thoughts: On the Power of Looped Transformers (ICLR 2025 · similarity 0.89)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.89)
• Decoding the Critique Mechanism in Large Reasoning Models (NeurIPS 2026 · similarity 0.89)