Is the reversal curse in LLMs real?

Original Article ↗ Hacker News Discussion ↗

Claims that large language models suffer from a “reversal curse” — failing to infer “B is A” from “A is B” (e.g., mapping from “Tom Cruise is the son of Mary Lee Pfeiffer” back to “Mary Lee Pfeiffer is Tom Cruise’s mother”) — have prompted debate over what these systems actually learn. Commenters argue over whether this is a fundamental failure of logical deduction, an artifact of how training stores associations, or simply a reflection of natural language ambiguity where “A is B” often doesn’t imply a reversible relationship. The thread also explores parallels with human memory, the gap between in-context reasoning and baked-in training, and the limits of using black-box, constantly changing commercial models like GPT‑4 to draw general scientific conclusions about LLM capabilities.

Reality and Significance of the “Reversal Curse”

  • Many see the “curse” (failing to go from “A is B” to “B is A”) as unsurprising given how next-token models work and the asymmetry of language.
  • Others argue that in identity-like cases (“X was the ninth chancellor”) most users expect reversibility, so failures are a bug from a product perspective.
  • Some say the paper overclaims: the effect is real but more about recall/representation than inability to perform logical deduction.

Semantics of “is” and Natural Language

  • Several comments stress that “A is B” usually does not imply “B is A” in natural language (“a bird is an animal” vs “an animal is a bird”).
  • Distinctions raised: “is of identity” vs “is of predication”; “A is B” vs “A is a B”; definite vs indefinite articles.
  • Others respond that the paper focuses on clearly reversible forms like “X is the Y” or unique roles (mother/son, chancellor number).

Training vs In‑Context Behavior

  • Strong consensus that LLMs can reverse relations reliably when both sides appear in the prompt; the issue arises in training-time generalization.
  • Some frame this as “the model can infer B from A, but the training process doesn’t store facts in a symmetric way.”
  • Debate over whether training is just “dumb optimization” versus any kind of “smart” meta-learning.

Comparisons to Human Cognition

  • Multiple analogies: language learning flashcards, Anki directionality, tip-of-the-tongue, asymmetric memory of movie roles, “force = mass × acceleration” versus using the formula in reverse.
  • Some argue humans also don’t get reversals “for free,” only after explicit inference and practice.
  • Others counter that for many simple factual relations humans are far more robust and nearly order-invariant.

Prompting, Data, and Fine‑Tuning

  • Several commenters emphasize that underrepresented facts, ambiguous prompts, and small or poorly designed fine‑tuning sets magnify the effect.
  • Examples show that wording (“is” vs “was”, “eighth” vs “ninth”) and extra context can flip answers from wrong to right.
  • Some worry this makes fine‑tuned models brittle for factual retrieval, even if base models handle reversals in rich contexts.

Counting, Numbering, and the German Chancellor Case

  • A long subthread examines misanswers about “the eighth/ninth Federal Chancellor of Germany.”
  • Disagreements arise over whether an “acting” officeholder should count; some say the model is clearly wrong, others say the ambiguity explains its choice.
  • More broadly, LLMs’ difficulty with positional/ordinal questions (e.g., “fifth word in a list”) is noted as a related weakness.

Mechanistic and Architectural Explanations

  • One line of discussion cites work showing feed‑forward layers acting as key–value stores: subject tokens provide keys, values store facts, and attention plus relation phrases retrieve them.
  • On this view, reversals are not free: “B is A” requires its own separate key–value mapping, so asymmetric storage is expected.
  • Some speculate bidirectional models (e.g., BERT-style) might fare better, but this remains unclear in the thread.

Broader Implications and Skepticism

  • Skeptics say the reversal issue reinforces that LLMs don’t “understand” or do general logical deduction, just statistical pattern completion.
  • Others argue that LLMs do approximate reasoning in-context, but training doesn’t exploit that capability to enforce symmetric generalization.
  • There is frustration about drawing strong scientific conclusions from proprietary, evolving models; calls appear for more open, mechanistic studies.
  • Some conclude the effect is mostly a nuance to keep in mind rather than a devastating indictment; others see it as emblematic of deeper limits.