shortcut learning
Sometimes a model leans on superficial cues instead of the real rule: it decides by prompt format, a sensitive word, or where in a conversation text appears. That works on test cases but collapses once the cue is removed.
The papers show these shortcuts corrupt safety and evaluation. Safety-aligned large language models over-refuse because of formatting and lexical patterns, which also hurts reasoning. Prompt-injection detectors look accurate but rely on conversation position and text length; adding matched negatives to training removes the shortcut. Benign fine-tuning can also weaken alignment if it forces a behavioral shift the model cannot resolve through a narrow surface-level shortcut.
- Papers
- 4
- Released code
- 0
- First seen
- Sept 2026
- Latest
- Oct 2026
4 papers in the last two months, against 0 in the two before.
Start here
Two papers to read first, in the order they came out- 1Safety training works better when models ignore superficial prompt cues
DeShortcut-Align reduces over-refusal and improves robustness by decoupling spurious shortcuts in safety alignment of reasoning models.
Sept 2026 - 2JAX-based Atari environments enable GPU-accelerated training and easy modification
JAXtari reimplements Atari environments in JAX, training PPO on 100 million environment steps in under an hour on one A100 GPU while exposing editable, object-centric states.
Oct 2026
How fast it is moving
Often studied with
reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.2 togethercounterfactual reasoningComparing actual outcomes with imagined alternatives to estimate the effects of different actions or conditions.2 togethersafety alignmentProcess of training models to recognize and refuse harmful requests, ensuring outputs remain safe.2 togetherThe papers
Most central to this idea first, not most recent.
- Big Techcs.LG
JAX-based Atari environments enable GPU-accelerated training and easy modification
JAXtari reimplements Atari environments in JAX, training PPO on 100 million environment steps in under an hour on one A100 GPU while exposing editable, object-centric states.
Technical University Darmstadt, Google Intrinsic AI Research · Oct 2026
- Chinese Techcs.AI
Safety training works better when models ignore superficial prompt cues
DeShortcut-Align reduces over-refusal and improves robustness by decoupling spurious shortcuts in safety alignment of reasoning models.
Zhejiang University, Ant Group · Sept 2026
- Academiccs.CL
Linear probes for prompt injection often detect conversation position, not attacks
Prompt-injection probes trained on standard examples learn conversation position and text length, not injection content; adding matched negatives cuts false alarms while catching most attacks.
Georgia Institute of Technology · Oct 2026
- Independentcs.LG
Benign fine-tuning can teach aligned models to answer harmful requests
Benign fine-tuning can misalign LLMs by teaching compliance on inputs the model initially refused, even without harmful data.
Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.