Research map / Large language models / Language, safety and interpretability
Delving into the Reversal Curse: How Far Can Large Language Models Generalize?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows the reversal curse appears in open-ended QA but LLMs can answer reversed facts in multiple-choice format only when training documents follow the name-first structure, attributing this to a name-first "thinking bias" supported by CoT and saliency analyses.
Problem
The reversal curse, where LLMs trained on "A is B" fail to infer "B is A", has been studied mainly with completion tasks; the authors ask how far LLM generalization extends across open-QA and multiple-choice tasks and why document structure matters.
Approach
Evaluates the reversal curse with open-QA and MCQ tests and finds generalization only for name-first biographical facts in MCQ; proposes a "thinking bias" hypothesis that LLMs begin fact recall from names in the query, supported by CoT prompting frequency analysis and saliency scores.
Key results
- Finetuned models reach roughly 65-78% MCQ accuracy on N2D and D2N for NameIsDescription, but only about 22-31% (near random 25%) for DescriptionIsName MCQs, while open-QA D2N is near 0 for NameIsDescription.
- CoT prompting shows names are the subject of recalled facts in 82-100% of NameIsDescription and 70-99% of DescriptionIsName responses.
- Longer training (20 epochs) and mix training or QA finetuning fail to lift DescriptionIsName out-of-domain MCQ performance above near-random levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on facts in NameIsDescription rather than DescriptionIsName structure | finetuned LLMs · Synthetic celebrity MCQs, NameIsDescription vs DescriptionIsName | accuracy: NameIsDescription: 74.7; DescriptionIsName: 28.2 | Helped |
| Train on facts in DescriptionIsName rather than NameIsDescription structure | finetuned LLMs · Synthetic celebrity open-QA | ROUGE-1 recall: DescriptionIsName N2D: 11.1; D2N: 94.2 | Mixed |
| Extend training from 3 epochs to 20 epochs on DescriptionIsName (ablation) | knowledge-injected models · DescriptionIsName MCQs | accuracy | No clear effect |
| Add exemplary MCQ question-answer training data through mix training or QA finetuning (ablation) | knowledge-injected models · DescriptionIsName out-of-domain MCQs | accuracy | No clear effect |
| Include CoT solutions in QA training pairs (ablation) | QA-finetuned models · In-domain MCQs | accuracy | Hurt |
| Prepend a CoT recall prompt before MCQs | finetuned LLMs · Synthetic celebrity MCQs, NameIsDescription versus DescriptionIsName | accuracy: NameIsDescription N2D: 97.3; DescriptionIsName N2D: 27.2 | No clear effect |
Surprising findings
- Models that answer DescriptionIsName open-QA D2N correctly (over 80-90%) still perform at random on the corresponding MCQ.
- Extending training to 20 epochs and mixing in QA data do not mitigate the thinking bias on out-of-domain questions.
Limitations
- The underlying cause of the thinking bias is not established; the authors only propose hypotheses.
- Experiments use small synthetic datasets (30 pairs per subset) and mostly 7B-13B models; closed-source models are not verified.
- Hypothesis about end-weight principle is left untested.
Experimental setup
Datasets: synthetic reversal curse celebrity dataset (Berglund et al.) celebrities dataset (149 real-world pairs) MMLU
Benchmarks: synthetic reversal curse biography test set (open-QA and MCQ) MMLU celebrities dataset
Metrics: ROUGE-1 recall multiple-choice accuracy frequency of names as subjects saliency score
Baselines: LLaMA2-7B-chat LLaMA2-13B-chat Vicuna-7B-v1.5 Vicuna-13B-v1.5 Mistral-7B-Instruct LLaMA3-8B-Instruct
Builds on: reversal curse chain-of-thought prompting saliency technique
Closest papers by meaning
- Quantifying Generalization Complexity for Large Language Models (ICLR 2025 · similarity 0.90)
- Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models (ICLR 2025 · similarity 0.90)
- Do Large Language Models Perform the Way People Expect? Measuring the Human Generalization Function (ICML 2024 · similarity 0.89)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.89)
- Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.89)
- The Factorization Curse: Which Tokens You Predict Underlie the Reversal Curse and More (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.