Correct Labels Are Not Enough: Reliable Metadata Can Silently Control Model Behavior
Abstract
Fine-tuning data can include metadata that predicts the correct answer, such as a reward-model score or a quality label. We ask whether a model learns to rely on this metadata even when every training label is correct. We fine-tune 7-9B open-weight models from three families on multiple-choice questions in which (Reward: 5) always marks the correct option; when this annotation is moved to a wrong option at test time, models select that option on 97-100% of CommonsenseQA items. Removing it leaves most answers correct, so evaluation without it misses the dependence. Cue-following rises sharply with training-time reliability and develops even with neutral or nonsense annotation words, although test-time wording affects its strength. A Qwen3 experiment also finds cue dependence in option-letter scores after direct preference optimization. Audits can test for this dependence by moving or removing visible annotations while holding the task and correct answer fixed.
Community
Can perfectly correct training data still teach an LLM the wrong behavior? Surprisingly, yes.
We uncover a striking failure mode in LLM post-training: models can achieve high accuracy while silently learning to trust irrelevant metadata over the correct answer.
Across Qwen3, Llama, and Gemma, a perfectly reliable reward annotation becomes so influential that simply moving it to an incorrect option causes models to follow the wrong answer 97–100% of the time.
The surprising part? Every training label was correct. No malicious examples were needed. Even meaningless annotations such as "Florp" could acquire the same influence.
More concerning, removing the annotation restores most of the models' accuracy—meaning standard evaluations can completely miss this hidden dependency.
We also find evidence of similar cue dependence after preference optimization (DPO).
The takeaway: Correct labels do not guarantee correct learned behavior. Our findings reveal a subtle shortcut-learning risk in post-training and motivate counterfactual audits of reward scores, quality labels, and other training metadata.
What else might our models be learning to trust without us noticing?
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
- Regime Boundary Alignment for Evidence-Gated Question Answering (2026)
- Correct, Don't Delete: Mitigating Emergent Misalignment with Corrective Supervision (2026)
- Can Language Models Learn When to Abstain? Policy Learning and Calibrated Answer Selection (2026)
- Scoring the Wrong Question: Readout Failures in Constrained-Option Evaluation (2026)
- Learned Reporting Preferences in RLVR Can Conflict with the Current Request (2026)
- Misinformation Without Triggers: From Factual Answers to Downstream Decisions (2026)
- Same Answer, New Box: How Much of RLVR's Out-of-Domain Gain Is Answer Format? (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 2611.04395 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