Invariance Training via Adversarial Robustification
Use robustified mathematical problem reformulations as a data augmentation strategy during fine-tuning to prevent models from overfitting to specific phrasings.
Time
Compute level
Resources
Prerequisites
Scope
The Gap
Section 5.1 and Table 9 reveal that model performance drops significantly (up to 11.2%) on 'robustified' problems compared to originals, indicating that models are overfitting to specific phrasings or memorizing training data.
Limitation
Models lack robustness to surface-level changes (variable names, distractors, value shifting), suggesting they rely on pattern matching rather than deep reasoning.
Proposed Solution
Implement the robustification pipeline described in Appendix A.4 (reformulation, distractors, numerical shifts) as a training-time data augmentation strategy. Fine-tune a model on pairs of {Original, Robustified} problems from datasets like MATH or GSM8K, forcing the model to align its reasoning paths or hidden states across these semantically equivalent but syntactically distinct inputs.
Why This Matters
This directly addresses the 'memorization vs. reasoning' dilemma, ensuring that high benchmark scores reflect actual mathematical understanding rather than dataset contamination.
How to Get Started
How to test it: We hypothesize that fine-tuning on robustified dataset variants will improve out-of-distribution robustness. We evaluate using accuracy drop between original and robustified benchmarks (like MATH or GSM8K). If fine-tuning degrades performance on original unperturbed tasks by more than 5%, we will halt.
- Generate paired datasets of original and robustified mathematical problems using rules for numerical shifts and syntactic distractors.
- Set up a fine-tuning pipeline on public mathematics benchmarks like MATH and GSM8K using a large compute cluster.
- Train the model to align its hidden states across semantically equivalent original and robustified problem inputs.
- Evaluate robustness gains and generalization performance on perturbed evaluation splits.
Code linked to this paper on Hugging Face: github.com/google-deepmind/superhuman
Suggested by AI. Adapt it to your setup.
Has This Been Done?
The listed works such as 'Understanding adversarial training' and 'Adversarial Training is a Form of Data-dependent Operator Norm Regularization' discuss general adversarial training and stability for classification or regression. They do not address fine-tuning LLMs on pairs of original and robustified mathematical word problem datasets to align reasoning paths.
- Understanding adversarial training: Increasing local stability of supervised models through robust optimization · 2018 · 475 citations
- Adversarial Training is a Form of Data-dependent Operator Norm Regularization · 2019 · 51 citations
We searched OpenAlex for “Use robustified mathematical problem reformulations as a data augmentation strategy during fine-tuning to prevent models from overfitting to specific phrasings.” and “Invariance Training via Adversarial Robustification” and an AI compared the idea with the top 7 results. A small search can miss relevant work.
Source Paper
Towards Robust Mathematical ReasoningThang Luong, Dawsen Hwang, Hoang H. Nguyen +17 more · 2025
This paper introduces IMO-Bench, a challenging benchmark suite designed to evaluate mathematical reasoning at the International Mathematical Olympiad (IMO) level, comprising tasks for short-answer verification (IMO-AnswerBench), proof writing (IMO-ProofBench), and proof grading (IMO-GradingBench). The authors demonstrate that current frontier models, while improving, still struggle with robustification (reformulated problems) and subtle logical fallacies in proofs, proposing a new gold-standard model (Gemini Deep Think) that achieves significant gains through agentic search and verification.
Related Ideas
- Distilling Explicit Reasoning into Latent SpaceRobust-R1: Degradation-Aware Reasoning for Robust Visual Understanding🟡Fairly easy
- Interpretable Reasoning via Decodable Think TokensDeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing🟠Medium
- BitDistill for Reasoning-Intensive DomainsBitNet Distillation🟠Medium
- Multi-Corpus Pretraining and Post-Training Reasoning Evaluation of Loopback TransformersLoopback Transformer: Towards a Unified View of Residual Looped Transformers🟠Medium
- Adaptive Token-Level Gating for Heterogeneous Modality AlignmentInternVL-U: Democratizing Unified Multimodal Models for Understanding, Reasoning, Generation and Editing🟡Fairly easy