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

Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

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.

Mechanistically analyzes reasoning hop generalization failures in LLMs, finds errors concentrate on a few key error types driven by competition between erroneous and correct attention heads, and proposes TCR, a test-time intervention that knocks out predicted ep heads when entropy is high.

Problem

LLMs' CoT performance drops sharply when the number of reasoning hops at test time exceeds the training distribution, and the internal mechanisms behind this failure are poorly understood.

Approach

Errors concentrate on a few key error types; at those token positions, erroneous processing heads (ep heads) amplify incorrect trajectories and suppress correct ones, so knocking out ep heads at test time (chosen by a trained head selector and triggered by an entropy threshold) rectifies errors.

Key results

• TCR improves average accuracy across seven tasks with Qwen2.5-7B-Instruct from 41.7% to 48.5% (+6.8%).
• TCR-gold with an oracle detector raises Qwen2.5-7B-Instruct average accuracy from 41.7% to 61.3%.
• Trained head selector achieves 79.6% in-distribution Hit@1 for Qwen2.5-7B-Instruct versus 42.3% random guessing.

Surprising findings

• Knocking out individual ep heads can restore correct predictions while preserving originally correct ones.
• DoLa yields marginal or negative effects on these reasoning hop generalization tasks.

Limitations

• Experiments limited to open-source models within 10B parameters.
• Only classical CoT studied; long-CoT with reflection not explored.
• Logit Lens may suffer from semantic-space misalignment; in-distribution task set is small.

Experimental setup

Datasets: Parity-NL LLC MDM MOAS CLF ObjC NumS

Benchmarks: Parity-NL LLC MDM MOAS CLF ObjC NumS

Metrics: accuracy Hit@1

Baselines: DoLa Removing-and-Resampling Random head selector Original model output

Builds on: Chain-of-thought prompting Logit Lens DoLa LoRA

Closest papers by meaning

• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.90)
• Reasoning with Latent Thoughts: On the Power of Looped Transformers (ICLR 2025 · similarity 0.90)
• Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models (ICML 2026 · similarity 0.90)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.89)
• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.89)
• Ada-R1: Hybrid-CoT via Bi-Level Adaptive Reasoning Optimization (NeurIPS 2025 · similarity 0.89)