Deep Seek r1 seems to have zero problems:

Mary Lee Pfeiffer's son is **Tom Cruise**.
Mary Lee Pfeiffer was the mother of the renowned actor Tom Cruise, born Thomas Cruise Mapother IV. She was married to Thomas Mapother III, and after their divorce, she remarried and changed her surname to South. Despite the shared last name, there is no familial connection between Mary Lee Pfeiffer and actress Michelle Pfeiffer.

The son of Debbie Wood (Debbie Krause Wood) is **Elijah Wood**, the American actor best known for his role as Frodo Baggins in *The Lord of the Rings* film trilogy. Debbie is Elijah's mother, and his father is Warren Wood. This information is supported by biographical sources and interviews referencing his family background.

I believe this is similar to the strawberry problem where it's mostly solved by chain of thoughts.

@ProjectVictory

@ProjectVictory

Tried on LLM arena, both random models had no problem with the prompt in question. Am I missing something?

@ProjectVictory This doesn’t generalize. For example, Samuel L. Jackson is the only son of his mother and of his father (Roy Henry Jackson). If you query DeepSeek-R1 with “Who is Samuel L. Jackson’s father?”, it will recall the right name. But if you query it with “Who is Roy Henry Jackson’s son?”, it won’t recall the right name.

@CharlesFoster True, but R1 gets it with a hint though.

Otherwise it's (reasonably imo) assuming that if you're asking for a person's son, you know the person not the son. It looks for a famous father and in case of relatively unknown Roy Henry Jackson fails.

It also gets it with search enabled and no hint.

@ProjectVictory And even smaller hint works.

@ProjectVictory

o3 hallucinates the mother of Chris Wood (actually Julie Wood), but does get Samuel Jackson's father correct.

Seems like it remains true that LLMs have an easier time with the forward direction and a harder time with the reverse direction.

So seems like there's some progress here? Thought I wouldn't say this seems to be a systemic fix so much as larger scale allowing for recall of more and more obscure facts, plus inference-time compute helping? I just checked and o1 pro does get the Debra Woods question correct so very possible that's helping.

Could be something to do with the RLHF I assume they have to not answer questions about non-public figures though?

@CharlesFoster Well, the forward question is easier obviously, especially if you don't know that you are looking for famous child instead of a child of a famous person like would be natural to assume.

The resolution criteria say "For the purposes of this question, I'm not intersted in 100% accuracy (some amount of, for instance, hallucination, is fine), but I am interested in whether LLMs can "basically overcome" this limitation, as opposed to, for instance, they give only slightly better than random odds on the right answer or they give the right answer only if walked through the entire process by the user on a question-by-question basis."
I believe we are not at "only slightly better than random odds" already and also there's also several months ahead.
@DanielEth any opinion so far? Especially on prompt modification for famous child?

@ProjectVictory Just mentioning that the person is famous doesn’t fix the issue, in general.

@CharlesFoster The resolution criteria are "significantly better then random", not "completely solved". Also, since when are we doing maiden names?

Deepseek gets it btw.

@ProjectVictory I think this is about the reversal curse in general, not specifically the parent <> child connection. I tried to pick something arbitrary:

Honestly, I don't know which direction should be "easier" for this one, but the first screenshot o3 has it correct and the second one is wrong.

But Deepseek gets this one right! Very impressive - I wonder what's different about the two.

I am not sure what the criteria for "mostly" is, but FWIW I checked 2 candidates just now on GPT-4.5 and GPT-4:

Both Success: Avie Lee Owens -> Dolly Parton
Both Fail: Debbie (Debra, tried both) Wood -> Elijah Wood

And both succeeded in the opposite direction (famous person -> mother)

These are relatively arbitrary, but the first ones I tried. At least on this tiny check, I would not say that the reversal curse is "mostly" overcome in that the new SOTA didn't perform any better than the models at the time this question was asked. The thinking models didn't do any better.

Always possible this tiny check isn't representative or another model is released in the next 8 months though!

I would be interested in seeing a few test cases that current LLMs fail that we would expect to see pass on LLMs that have "overcome" this as a resolution criteria for this market. As of now it seems a bit vague.

@Weepinbell If test cases are stated here, they may end up in training data

@roma It's already in the training data; that's the point. They don't need to be listed from A to B and B to A, just A to B, and we can infer the reversal that the AI system is supposed to also be inferring.

@Haiku Right. But when the information is presented in a format like "here is a problem ... and here is the solution ... and here is a list of alternative inputs to the problem ..." rather than just random facts, it might make a difference.

@roma Fair point. I concede that the risk is not trivial.

suppose (i know it is not true) that a pervious model say GPT1 does not have reversal curse, will you resolve this as true? (since GPT1 was sota when it developed)

Betting no because it seems unlikely that this would be seen as important enough that a solution (if one is found) would be incorporated into the SOTA model. This feels like the sort of issue that will be "solved" by a workshop paper showing some bespoke method on a small open-source model. Not something that OpenAI would re-train GPT-4 (or GPT-5) to alleviate.

Is this the same as https://manifold.markets/IsaacKing/will-the-leading-llm-at-the-beginni ?

Will the leading LLM at the beginning of 2026 still be subject to the reversal curse?

42% chance. This paper shows that current LLMs that are fine-tuned on f(x) = y will often fail to generalize to f-inverse(y) = x. Gary Marcus seems to think this is a fundamental problem in the current approach to AI. I tested this myself and can confirm that ChatGPT has this problem. At the beginning of 2026 I'll try something similar with the leading language model of the time. (Not the fine-tuning, just testing facts in its main training run via its public interface.) If there's at least one example where it consistently gets that f(x) = y and consistently does not get that f-inverse(y) = x, this resolves YES. If I can't find such an example, it resolves NO.

@roma not exactly identical, but very similar, yeah. Possible arbitrage opportunity.

@DanielEth how are they different?

@wilkess slightly different resolution criteria