Scaling behavior of perturbation-induced reasoning failures in larger LLMs

Determine the scaling behavior of robustness to meaning-preserving perturbations and associated mechanistic failure signatures in transformer-based large language models as parameter counts increase beyond 8B parameters, specifically assessing how answer flip rates under name substitution and number-format paraphrasing, first divergence layers under logit lens, activation patching recoverability, attention versus MLP component ablation effects, and the Cascading Amplification Index evolve when applying the Mechanistic Perturbation Diagnostics framework to models larger than the 7–8B parameter range evaluated here.

Background

The study evaluates three instruction-tuned transformer LLMs—Mistral-7B-Instruct-v0.2, Llama-3-8B-Instruct, and Qwen2.5-7B-Instruct—on 677 GSM8K problems paired with meaning-preserving perturbations (name substitution and number-format paraphrasing). All models exhibit substantial answer-flip rates, with number paraphrasing more disruptive than name swaps.

To trace the mechanistic basis of failures, the authors introduce the Mechanistic Perturbation Diagnostics (MPD) framework combining logit lens analysis, activation patching, component ablation, and the Cascading Amplification Index (CAI). They identify architecture-specific failure modes: distributed (Mistral), localized (Llama-3), and entangled (Qwen), and show that CAI significantly predicts failure.

However, all analyses are conducted on models in the 7–8B parameter range. The authors explicitly note that scaling behavior is unknown, leaving open whether the observed robustness patterns and mechanistic signatures persist, attenuate, or change qualitatively at larger model scales.

References

We evaluate only 7--8B parameter models; scaling behavior is unknown.

— Fragile Reasoning: A Mechanistic Analysis of LLM Sensitivity to Meaning-Preserving Perturbations  (2604.01639 - Han et al., 2 Apr 2026) in Limitations

Our analyses show that the exposed failures cannot be explained by simple surface-level artifacts alone, but what linguistic or representational factors fundamentally make models vulnerable under semantic-preserving variations remains an open question.

— Hallucination-R1: Robustness-Oriented Paraphrase Generation for Factual Consistency  (2609.21227 - Yu et al., 18 Sep 2026) in Limitations, third limitation

Future work should investigate whether similar structural sensitivity exists across a broader range of multilingual reasoning tasks.

— Lost in Reordering: Structural Sensitivity of Multilingual LLMs under Semantics-Preserving Perturbations  (2609.03511 - Nhayakkat et al., 3 Sep 2026) in Limitations

Encouragingly, some fault tolerance has been built into particular billion-weight models , but only at a single scale: we do not know whether that behavior is part of a persistent trend or sits at the edge of a cliff beyond which errors consume models.

— Fault-tolerant foundation models  (2610.10311 - McCourt et al., 7 Oct 2026) in Main text, paragraph beginning “The most pressing question concerning fault-tolerant neural network utility”

Reading the text does not reveal which rewrites these are, so a human monitor could not anticipate them, and this tail is the part of the convention question that remains open.

— On the Monitorability of Natural Language Autoencoders: Identification vs Discrimination  (2610.18727 - Denamganaï, 13 Oct 2026) in Section 6.2, paragraph “The damage is concentrated in a few rewrites”