Pith. sign in

REVIEW 24 cited by

The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2309.12288 v4 pith:VQI3MMGE submitted 2023-09-21 cs.CL cs.AIcs.LG

Lukas Berglund , Meg Tong , Max Kaufmann , Mikita Balesni , Asa Cooper Stickland , Tomasz Korbak , Owain Evans This is my paper
classification cs.CLcs.AIcs.LG
keywords cursereversalmodelanswermodelsreversetrainedwill
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form "A is B", it will not automatically generalize to the reverse direction "B is A". This is the Reversal Curse. For instance, if a model is trained on "Valentina Tereshkova was the first woman to travel to space", it will not automatically be able to answer the question, "Who was the first woman to travel to space?". Moreover, the likelihood of the correct answer ("Valentina Tershkova") will not be higher than for a random name. Thus, models do not generalize a prevalent pattern in their training set: if "A is B" occurs, "B is A" is more likely to occur. It is worth noting, however, that if "A is B" appears in-context, models can deduce the reverse relationship. We provide evidence for the Reversal Curse by finetuning GPT-3 and Llama-1 on fictitious statements such as "Uriah Hawthorne is the composer of Abyssal Melodies" and showing that they fail to correctly answer "Who composed Abyssal Melodies?". The Reversal Curse is robust across model sizes and model families and is not alleviated by data augmentation. We also evaluate ChatGPT (GPT-3.5 and GPT-4) on questions about real-world celebrities, such as "Who is Tom Cruise's mother? [A: Mary Lee Pfeiffer]" and the reverse "Who is Mary Lee Pfeiffer's son?". GPT-4 correctly answers questions like the former 79% of the time, compared to 33% for the latter. Code available at: https://github.com/lukasberglund/reversal_curse.

Discussion (0). Sign in to comment.

Forward citations

Cited by 24 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Emergent Misalignment Recruits a Pre-existing Persona Subspace

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Fine-tuning on narrow bad data recruits a low-rank persona subspace already present in a frozen instruction-tuned model; holding that subspace out of activations prevents broad misalignment, and injecting it into the ...

  2. Can a Language Model Learn Facts Continually in Its Weights?

    cs.CL 2026-07 unverdicted novelty 7.0 of 10

    Weight writes create question-keyed knowledge that later sequential writes redirect, so context—not weights—is the reliable channel for facts that must survive or be composed.

  3. DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning

    cs.CL 2026-07 unverdicted novelty 7.0 of 10

    DiscoLoop adds a discrete embedding channel to looped transformers to fix representational misalignment in two-hop reasoning, yielding near-perfect accuracy on synthetic tasks and better pretraining loss on real data.

  4. Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

    cs.AI 2026-05 conditional novelty 7.0 of 10

    Masked diffusion language models, not larger autoregressive LLMs, are the better building block for text-based world models in agentic RL, improving rollout fidelity, diversity, and downstream task success.

  5. Can Gradient Descent Simulate Prompting?

    cs.CL 2025-06 conditional novelty 7.0 of 10

    A MAML-style meta-training objective makes a single gradient step on new text recover part of the performance that prompting achieves, on reversal-curse and passage-QA tasks.

  6. Quantifying Cross-Modality Memorization in Vision-Language Models

    cs.CV 2025-06 conditional novelty 7.0 of 10

    Fine-tuning VLMs on image-only or text-only personas yields a significant, asymmetric cross-modal memorization gap that persists with model scale, unlearning, and multi-hop reasoning.

  7. Hierarchical Domain Generalization

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Over infinite domains, hierarchy-uniform domain generalization is impossible for every nontrivial hypothesis class; a length-generalization bound is a property of the length hierarchy, not a hierarchy-free guarantee.

  8. DeFrame: Debiasing Large Language Models Against Framing Effects

    cs.CL 2026-02 conditional novelty 6.0 of 10

    LLM fairness scores shift substantially with positive vs negative framing of the same question, and DeFrame—a three-step self-revision prompt—reduces both average bias and this framing gap.

  9. Breaking the Reversal Curse in Autoregressive Language Models via Identity Bridge

    cs.AI 2026-02 conditional novelty 6.0 of 10

    Identity-bridge regularization, rephrased into an out-of-context reasoning form, yields ~40% reversal accuracy in a 1B LLM and provably fixes reversal in an idealized one-layer transformer.

  10. NarrativeTrack: Evaluating Entity-Centric Reasoning for Narrative Understanding

    cs.CV 2026-01 conditional novelty 6.0 of 10

    NarrativeTrack shows that video AI models, including GPT-4o, falter at tracking a specific person across scene changes, outfit changes, and similar-looking characters in long videos.

  11. Conflict-Aware Fusion: Mitigating Logic Inertia in Large Language Models via Structured Cognitive Priors

    cs.AI 2025-12 unverdicted novelty 6.0 of 10 partial

    Conflict-Aware Fusion mitigates Logic Inertia in LLMs through a four-stage pipeline of SFT, DPO, logical invariance regularization, and reinforcement learning from a symbolic oracle, saturating four stress tests on ru...

  12. The SuperActivator Mechanism: Transformers Concentrate Reliable Concept Signals in the Tail

    cs.LG 2025-12 conditional novelty 6.0 of 10

    Reliable concept presence in transformers is concentrated in the extreme high-activation tail of in-concept tokens; thresholding that tail improves concept detection and localization.

  13. Orchestrating Dual-Boundaries: An Arithmetic Intensity Inspired Acceleration Framework for Diffusion Language Models

    cs.CL 2025-11 conditional novelty 6.0 of 10

    ODB-dLLM accelerates diffusion LLM inference by adaptively shortening responses and using a jump-share speculative decoding strategy, achieving 2.6-6.3x over Fast-dLLM.

  14. Diffusion-Inspired Masked Fine-Tuning for Knowledge Injection in Autoregressive LLMs

    cs.CL 2025-10 unverdicted novelty 6.0 of 10

    Masked fine-tuning enables autoregressive LLMs to inject new factual knowledge without paraphrases and with reversal-curse resistance, matching diffusion LLM advantages on QA tasks.

  15. Energy-Based Transformers are Scalable Learners and Thinkers

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Energy-Based Transformers learn to predict by gradient-descent minimization of a learned energy function, and the paper reports faster pretraining scaling and inference-time thinking gains over Transformer++ and Diffu...

  16. Potemkin Understanding in Large Language Models

    cs.CL 2025-06 conditional novelty 6.0 of 10

    LLMs frequently pass definition questions yet fail to use the same concepts in classification, generation, and editing tasks, a gap the authors call potemkin understanding.

  17. SV-TrustEval-C: Evaluating Structure and Semantic Reasoning in Large Language Models for Source Code Vulnerability Analysis

    cs.SE 2025-05 conditional novelty 6.0 of 10

    Large language models perform poorly on a new C-code vulnerability benchmark, indicating they rely on pattern matching rather than genuine reasoning.

  18. Dynamics Reveals Structure: Challenging the Linear Propagation Assumption

    cs.LG 2026-01 conditional novelty 5.0 of 10

    Under the linear propagation assumption, first-order feature geometry cannot support both negation and composition: the only feature map satisfying both is the zero map.

  19. Reinforcement Learning Improves Traversal of Parametric Knowledge in LLMs

    cs.CL 2025-11 reject novelty 5.0 of 10

    Reasoning LLMs seem better at retrieving hierarchical facts not because they know more but because they navigate better; the key supporting RL experiment is missing from the paper.

  20. Noise or Nuance: An Investigation Into Useful Information and Filtering For LLM Driven AKBC

    cs.CL 2025-09 conditional novelty 5.0 of 10

    In a no-retrieval, no-fine-tuning setting, LLM-provided hints, self-judging filters, and regex-based parsing substantially change triple completion quality, with the judge filter roughly doubling F1.

  21. SymbolicThought: Integrating Language Models and Symbolic Reasoning for Consistent and Interpretable Human Relationship Understanding

    cs.CL 2025-07 reject novelty 5.0 of 10

    A human-in-the-loop system that adds logical rules and an interactive interface to LLM-based character relationship extraction, improving recall and cutting annotation time.

  22. PlanE: Meta Planning of Data, Tuning, and Inference for Extractive-based LLMs

    cs.AI 2026-05 conditional novelty 4.0 of 10

    A quadratic meta-planner trained on a few model-dataset runs selects the optimal data-tuning-inference configuration for extractive LLMs, matching grid search on three IE tasks.

  23. Adversarial Activation Patching: A Framework for Detecting and Mitigating Emergent Deception in Safety-Aligned Transformers

    cs.LG 2025-07 reject novelty 4.0 of 10

    A framework that borrows activation patching to adversarially induce and measure deception, supported only by an underspecified toy network simulation.

  24. Prompting as Scientific Inquiry

    cs.CL 2025-06 conditional novelty 4.0 of 10

    Position paper arguing that prompting LLMs is a form of behavioral science and should be recognized as a core scientific method alongside mechanistic interpretability.

Pith tools