UNMASK: Discovering and Causally Verifying Spurious Shortcuts in Text Classifiers
Abstract
UNMASK automatically discovers and mitigates spurious correlations in text classifiers via causal verification and group-based reweighting without manual annotations.
Neural language models trained on large crowdsourced corpora frequently exploit spurious surface patterns tied to target labels without true linguistic or causal relevance, boosting benchmark performance while failing on adversarial or out-of-distribution inputs. Existing approaches either require manual specification of the feature vocabulary or automate discovery only partially, leaving the gap between dataset-level correlation and model-level exploitation unaddressed. We present U N M ASK, a fully automated pipeline that discovers, causally verifies, and mitigates spurious correlations in text classifiers without additional human annotation. Given unlabeled training examples, U N M ASK generates candidate surface patterns as executable boolean expressions, filters them through a statistical validation protocol with independent replication, and establishes causal model dependence via verified counterfactual interventions. Causally confirmed features then serve as annotation-free group definitions for Deep Feature Reweighting, eliminating the group labels that standard DFR requires. Applied to BERT and RoBERTa trained on MNLI, our pipeline independently rediscovers established lexical-overlap and negation biases, verifying 9 of 10 features on BERT and 6 on RoBERTa, and improving HANS accuracy by up to 12.58 pp. On CivilComments-WILDS, programmatic groups match the 70.1% worst- group accuracy of hand-labeled DFR (Kirichenko et al., 2023) without demographic annotation. We further demonstrate that the discovery and validation stages generalize to reward model preference data, surfacing interpretable spurious correlations in RewardBench2.
Community
Language Models are very good at exploiting shortcuts. Existing approaches either require manual specification of the spurious feature or automate discovery only partially. The gap between dataset-level correlation and model-level exploitation is unaddressed. UNMASK closes that. The pipeline generates candidate shortcuts as executable boolean predicates.
has_neg(h) AND overlap(p,h) >= 0.8
is a hypothesis you can run on the corpus - statistically test it, rank it against every other candidate, and reuse the exact same function downstream. Stage 2 repairs them: an evaluator LLM checks each predicate against its actual execution trace and broken ones are rewritten. Survivors run over every training example, and only features passing Fisher's exact test under BH control - replicated on a held-out split get through. But a correlation in the data isn't a shortcut the model uses. Correlation doesn't mean causation right? So we intervene. Generate a minimal edit that removes the surface pattern while preserving the label, then measure the paired shift in predicted probability as causal signal. On MNLI: BERT causally exploits 9 of 10 validated features. RoBERTa only 6. And the strongest correlation in the entire dataset - always/every in premise + never/no in hypothesis, OR 10.01 is used by neither. A high odds ratio is a hypothesis, not a finding. And because the predicates are executable, they hand you group labels for free: 71.84% worst-group accuracy on CivilComments-WILDS, matching hand-labeled DFR, with no additional demographic annotation. For NLI tasks, HANS accuracy improves by up to 12.58 pp. We tried the discovery half at RewardBench2. No debiasing, just which surface predicates separate chosen from rejected and Math's strongest signal is the phrase "let me help you solve this step by step." OR 22.7. This forces to learn string, not a reasoning style.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Discovering Latent Groups for Robust Classification (2026)
- Diagnosing Fine-Grained Inconsistency Classification in Financial Disclosure Text (2026)
- Do LLM Attribution Metrics Transfer? Auditing Retrieval-Augmented Generation Evaluation Across Datasets and Constructs (2026)
- SURGELLM: Rethinking Multi-Task Evaluation through Task-Aware Feature Gating with Class-Balanced Normalization (2026)
- SABRE: Scalable and Automated Benchmarking of VLMs under Stress (2026)
- Detect, Unlearn, Restore: Defending Text Summarization Models Against Data Poisoning (2026)
- Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Get this paper in your agent:
hf papers read 2608.09209 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper