[Paper] Do explanations generalize across large reasoning models?

Source: arXiv - 2601.11517v1

Overview

Large reasoning models (LRMs) such as GPT‑4 or Claude often emit a chain‑of‑thought (CoT)—a step‑by‑step natural‑language explanation that leads to the final answer. This paper asks a surprisingly practical question: Do those explanations actually capture general problem‑solving knowledge, or are they just model‑specific quirks? By testing whether a CoT generated by one LRM can steer the behavior of other LRMs, the authors uncover when and how explanations transfer across models—a finding that matters for anyone building AI‑augmented tools, debugging model outputs, or trying to extract scientific insight from LLMs.

Key Contributions

• Definition of explanation generalization: Introduces a concrete metric—cross‑model consistency—that measures whether a CoT from Model A improves the answer quality of Model B.
• Empirical evidence of transfer: Shows that CoTs frequently boost consistency across a suite of LRMs (GPT‑3.5, GPT‑4, Claude, LLaMA‑2, etc.).
• Correlation with human preferences: Demonstrates that explanations that generalize better also rank higher in human preference studies and align with reinforcement‑learning‑from‑human‑feedback (RLHF) fine‑tuning.
• Analysis of success factors: Identifies linguistic and structural cues (e.g., explicit reasoning steps, low‑entropy phrasing) that make a CoT more portable.
• Simple ensembling technique: Proposes a sentence‑level voting scheme that aggregates multiple CoTs to further raise cross‑model agreement.
• Framework for caution: Provides a checklist for practitioners to assess when LRM explanations are safe to trust for downstream insights.

Methodology

1. Model pool: The authors selected several state‑of‑the‑art LRMs spanning different architectures and training regimes.
2. Task suite: A diverse set of reasoning benchmarks (math word problems, logical puzzles, commonsense QA) was used to ensure coverage of both symbolic and open‑ended reasoning.
3. Explanation extraction: For each input, Model A generated a CoT and a final answer. The CoT was then re‑fed to Model B as part of the prompt (e.g., “Here is a reasoning chain: … What is the answer?”).
4. Cross‑model consistency metric: Measured the proportion of cases where Model B’s answer matched Model A’s answer and the ground‑truth answer, compared to a baseline where no CoT was supplied.
5. Human evaluation: Crowdsourced workers ranked pairs of CoTs on clarity, plausibility, and helpfulness. Rankings were correlated with the consistency scores.
6. Analysis & ensembling: Linguistic features of high‑generalizing CoTs were extracted, and a sentence‑level voting ensemble was built by concatenating the most agreed‑upon reasoning steps from multiple CoTs.

Results & Findings

• Generalization is common: Across all model pairs, providing a CoT raised cross‑model consistency by 12–28 % relative to the no‑explanation baseline.
• Human‑preferred explanations generalize best: CoTs that received higher human preference scores showed a strong positive correlation (ρ≈0.68) with consistency gains.
• RL‑fine‑tuned models excel: Models that had undergone RLHF (e.g., ChatGPT) produced CoTs that transferred more effectively than purely supervised‑trained models.
• Structure matters: Explanations that explicitly enumerate steps, use concrete numbers, and avoid ambiguous pronouns yielded the highest transfer rates.
• Ensembling wins: The sentence‑level voting ensemble improved consistency by an additional 5–9 % over the best single CoT, with minimal extra compute.

Practical Implications

• Prompt engineering: When building pipelines that rely on LLM reasoning (e.g., code generation assistants, data‑analysis bots), injecting a well‑structured CoT from a strong LRM can make downstream models more reliable.
• Model‑agnostic debugging: Developers can use a “debug CoT” from a trusted model to surface hidden reasoning errors in a weaker or specialized model without retraining it.
• Scientific discovery workflows: Researchers can treat CoTs as hypothesis drafts—if a chain of reasoning persists across multiple LRMs, it’s more likely to reflect a genuine pattern rather than a model artifact.
• Ensemble services: SaaS platforms can cheaply boost answer consistency by aggregating a few short CoTs (e.g., three sentences from different models) instead of running large ensembles of full model generations.
• Human‑in‑the‑loop tools: UI designs that surface the CoT to users can double as a quality filter; users can accept or reject the reasoning, and the system can fall back to a baseline if the CoT fails to improve consistency.

Limitations & Future Work

• Scope of tasks: The study focuses on benchmark reasoning problems; real‑world domains (legal reasoning, medical diagnosis) may exhibit different transfer dynamics.
• Model diversity: While several major LRMs were tested, the findings may not hold for smaller, domain‑specific models or future multimodal architectures.
• Prompt sensitivity: The exact phrasing used to feed the CoT to the second model influences outcomes, and the paper does not exhaustively map this space.
• Explainability vs. performance trade‑off: Some high‑performing models may generate terse answers that are less transferable; balancing brevity and explainability remains open.
• Future directions: Extending the framework to cross‑modal explanations (e.g., visual reasoning), exploring automated CoT quality scoring, and integrating formal verification of reasoning steps are promising next steps.

Authors

• Koyena Pal
• David Bau
• Chandan Singh

Paper Information

• arXiv ID: 2601.11517v1
• Categories: cs.CL, cs.AI
• Published: January 16, 2026
• PDF: Download PDF