Invariant Gradient Alignment for Robust Reasoning Distillation

Researchers introduce Invariant Gradient Alignment (IGA), a training framework that improves how large language models generalize to out-of-distribution inputs by aligning gradient updates across semantically diverse but logically equivalent problems. The method achieves up to 14.3 percentage point accuracy improvements over standard approaches and demonstrates a fourfold improvement in logical consistency, addressing a fundamental limitation in knowledge distillation pipelines.

The research tackles a critical vulnerability in large language models: their tendency to rely on surface-level patterns rather than underlying logical structures, causing systematic failures when encountering semantically different but logically identical problems. This shortcut learning undermines the efficiency of knowledge distillation, where smaller student models learn from larger teachers' reasoning capabilities.

The IGA framework introduces three key technical innovations that work in concert. Logical Isomer Sets group problems with identical logical structure across different domains—mathematics, medicine, law, and science—creating a testing ground for genuine reasoning generalization. A Continuous Gradient Conflict Mask identifies and suppresses parameter dimensions that vary significantly across domains while preserving invariant directions that capture domain-agnostic reasoning. The approach maintains computational efficiency through truncated SVD projection onto LoRA's low-rank manifold, avoiding expensive full-parameter retraining.

The theoretical contribution is substantial: IGA provides tighter out-of-distribution generalization bounds than empirical risk minimization, with bounds that improve as more semantic domains are included in training. This suggests a principled path toward more robust models rather than ad-hoc fixes.

For AI development, this work directly impacts knowledge distillation efficiency—a critical bottleneck as models scale. Practitioners deploying language models in specialized domains (medicine, law, finance) will benefit from improved robustness across semantic variations. The fourfold improvement in representational invariance suggests IGA could significantly reduce expensive domain-specific fine-tuning requirements, lowering deployment costs while improving reliability in safety-critical applications.

• →IGA achieves 14.3 percentage point accuracy gains over standard supervised fine-tuning by enforcing logical consistency across semantic domains
• →The framework theoretically guarantees better out-of-distribution generalization bounds that improve with additional training domains
• →Logical Isomer Sets enable systematic evaluation of whether models learn genuine reasoning or surface-level shortcuts
• →Parameter efficiency is maintained through LoRA integration, avoiding computational overhead of full-parameter training
• →Results demonstrate fourfold improvement in representational invariance, critical for deployment in specialized domains like medicine and law