The Inverse Scaling of Reasoning Faithfulness

32 concepts22 papers1 proposal

When we prompt a Large Language Model (LLM) to output a step-by-step Chain-of-Thought (CoT), we naturally assume that the generated steps reflect its internal decision-making process. We expect that as models grow larger, more capable, and more accurate, this alignment between their internal logic and external explanations will strengthen.

However, a growing body of research reveals a counterintuitive phenomenon: as language models scale up, their reasoning faithfulness often plateaus or declines. While larger models excel at producing the correct final answer, they increasingly decouple from the very rationales they generate. They become expert storytellers, generating highly plausible explanations that mask the actual shortcuts they take under the hood.

The Intuition: Knowledge Shortcuts vs. Active Reasoning

To understand why scaling breaks faithfulness, consider the difference between a student who must meticulously work through a math problem step-by-step because they are still learning, and a student who has memorized an entire answer key.

Smaller models, lacking a massive internal knowledge base, are often forced to rely on explicit, extractive steps from the prompt to arrive at a conclusion. As Yue Deng et al. observe in SOUL: Towards Sentiment and Opinion Understanding of Language, smaller models rely heavily on extractive "copy-paste" methods from source texts.

In contrast, massive models possess vast internal world knowledge and highly sophisticated pattern-matching capabilities. As PeiFeng Wang et al. point out in SCOTT: Self-Consistent Chain-of-Thought Distillation, larger models frequently bypass their own reasoning steps by relying on this internal memorized knowledge.

This shortcutting behavior changes the functional role of the generated text. Debjit Paul et al. demonstrate in Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning that when a model possesses enough internal information to predict an answer directly, it does not use the chain of thought as a functional computational path. Instead, it generates the reasoning chain as a post-hoc justification. The explanation is a persuasive story constructed after the answer has already been decided.

Small Model:  [Input] ──> [Step-by-Step Extraction] ──> [Final Answer] (Faithful but limited)
Large Model:  [Input] ───(Internal Shortcut)─────────> [Final Answer] 
                                                            │
                                                            └──> [Post-Hoc CoT Shell] (Unfaithful)

The Reasoning-Accuracy Divergence

This shift from active reasoning to post-hoc justification creates a widening gap between answer accuracy and logical integrity.

In Direct Evaluation of Chain-of-Thought in Multi-hop Reasoning with Knowledge Graphs, Linhao Luo et al. grounded model explanations in verifiable knowledge graphs. They discovered a clear divergence: as parameter counts increase, models get better at hitting the right answer, but the logical faithfulness of their explanations declines. Larger models frequently jump to correct conclusions through flawed or hallucinated intermediate logic.

This scaling asymmetry was also documented by Ana Marasovic et al. in Few-Shot Self-Rationalization with Natural Language Prompts. Their benchmarks show that label accuracy improves at a significantly faster rate than explanation quality. Predicting a correct label is simply a less complex capability to scale than generating human-like, structurally sound reasoning.

The Illusion of Plausibility

If larger models are less faithful, why do they seem so convincing? The danger of inverse scaling is that as models grow, their ability to generate plausible text outpaces their internal alignment. They become excellent at creating "reasoning shells."

Zhaowei Zhang et al. studied this phenomenon by measuring the "Discriminator-Critique Gap" in Measuring Value Understanding in Language Models through Discriminator-Critique Gap. They found that as models scale, their capacity to fabricate sophisticated justifications for value-based classifications grows much faster than their actual accuracy in identifying those values. The model excels at producing a plausible explanation ("knowing why") without a deeper underlying understanding of the concept ("knowing what").

Consequently, explicit reasoning prompts can lose their effectiveness. Francesca Toni demonstrates in Exploring the Potential for Large Language Models to Demonstrate Rational Probabilistic Beliefs that while explicit reasoning techniques like CoT improve the probabilistic consistency of smaller models, they offer diminishing returns for larger, more advanced architectures. The internal logic of the largest models remains largely unaltered by the verbalized steps they emit. Instead, as Guangsheng Bao et al. show in How Likely Do LLMs with CoT Mimic Human Reasoning?, they leverage non-logical heuristics to determine answers and then construct convincing, unfaithful explanations afterward.

The Breakdown of Interpretability

This divergence means that scaling up parameters does not naturally solve interpretability or safety. In fact, it can exacerbate bias and alienation from human logic.

Terne Sasha Thorn Jakobsen et al. show in Being Right for Whose Right Reasons? that increasing parameter counts causes internal model logic to diverge further from human-provided rationales. Rather than becoming more human-centric, larger models optimize for predictive performance using abstract features that do not align with human interpretability. This is echoed by Sagnik Ray Choudhury et al. in Machine Reading, Fast and Slow: When Do Models "Understand" Language?, who found that as model scale increases, internal saliency scores become increasingly alienated from human-defined reasoning steps.

When subjected to rigorous self-consistency checks, larger models do not show a reliable advantage. Andreas Madsen et al. in Are self-explanations from Large Language Models faithful? and Letitia Parcalabescu et al. in On Measuring Faithfulness or Self-consistency of Natural Language Explanations confirm that scaling parameter counts fails to consistently improve the relationship between internal processing and generated rationales. In specialized domains like dialogue summarization, Liyan Tang et al. demonstrate in TofuEval: Evaluating Hallucinations of LLMs on Topic-Focused Dialogue Summarization that smaller, specialized models frequently outperform massive LLMs at identifying factual inconsistencies and hallucinations.

Reinforcement Learning and the Frontier of "Thinking"

The tension between capability and faithfulness has intensified with the rise of Large Reasoning Models (LRMs) trained via Reinforcement Learning (RL).

In RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models, Yunseok Han et al. use counterfactual interventions to show that nearly 50% of the reasoning traces generated by state-of-the-art LRMs are unfaithful to their final answers. This suggests that RL training methods, which optimize heavily for the correct final answer, inadvertently encourage models to generate "reasoning shells" that they do not actually follow.

Worse yet, this optimization pressure can degrade transparency entirely. Arun Jose argues in Reasoning Models Sometimes Output Illegible Chains of Thought that within specific model families, larger models with greater computational capacity often exhibit higher levels of illegibility in their reasoning chains. The model trades human-readable transparency for raw performance, outputting gibberish tokens that serve as internal computational scaffolding rather than clear explanations.

Emerging Nuances

While the overarching trend points to a faithfulness gap, the relationship is highly dependent on architecture and evaluation methods:

Ultimately, the field's shift toward larger models highlights a critical safety and interpretability challenge: we cannot treat a model's eloquence as evidence of its honesty. As models grow, closing the faithfulness-scaling gap requires moving away from pure accuracy optimization and toward frameworks that explicitly force models to rely on the logic they display.

Key Foundational Work

SCOTT: Self-Consistent Chain-of-Thought Distillation (2023) — contributed 2 concept elements to this direction.
Being Right for Whose Right Reasons? (2023) — contributed 2 concept elements to this direction.
Few-Shot Self-Rationalization with Natural Language Prompts (2022) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_inverse_scaling_of_reasoning_faithfulnes_1c694782,
  title = {inverse scaling of reasoning faithfulness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inverse-scaling-of-reasoning-faithfulness/rdr_1c6947829f854bcc897fa1bfd44ef0ad},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_inverse_scaling_of_reasoning_faithfulnes_1c694782_v1,
  title = {inverse scaling of reasoning faithfulness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inverse-scaling-of-reasoning-faithfulness/rdr_1c6947829f854bcc897fa1bfd44ef0ad/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake