HackerPapers

arXiv:2410.05229cs.LG

GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models

Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, Mehrdad Farajtabar

Abstract

Recent advancements in Large Language Models (LLMs) have sparked interest in their formal reasoning capabilities, particularly in mathematics. The GSM8K benchmark is widely used to assess the mathematical reasoning of models on grade-school-level questions. While the performance of LLMs on GSM8K has significantly improved in recent years, it remains unclear whether their mathematical reasoning capabilities have genuinely advanced, raising questions about the reliability of the reported metrics. To address these concerns, we conduct a large-scale study on several SOTA open and closed models. To overcome the limitations of existing evaluations, we introduce GSM-Symbolic, an improved benchmark created from symbolic templates that allow for the generation of a diverse set of questions. GSM-Symbolic enables more controllable evaluations, providing key insights and more reliable metrics for measuring the reasoning capabilities of models.Our findings reveal that LLMs exhibit noticeable variance when responding to different instantiations of the same question. Specifically, the performance of all models declines when only the numerical values in the question are altered in the GSM-Symbolic benchmark. Furthermore, we investigate the fragility of mathematical reasoning in these models and show that their performance significantly deteriorates as the number of clauses in a question increases. We hypothesize that this decline is because current LLMs cannot perform genuine logical reasoning; they replicate reasoning steps from their training data. Adding a single clause that seems relevant to the question causes significant performance drops (up to 65%) across all state-of-the-art models, even though the clause doesn't contribute to the reasoning chain needed for the final answer. Overall, our work offers a more nuanced understanding of LLMs' capabilities and limitations in mathematical reasoning.

arXiv comments
ICLR camera ready + additional discussion in the appendix
Journal reference
DOI
Subjects
cs.LGMachine Learning cs.AIArtificial Intelligence

On Hacker News

Linked by 18 identifiable commenters in 19 comments across 17 threads, from Oct 2024 to Jul 2026. Found in 1 HN submission from Oct 2024 to Oct 2024. Open submission discussions.

Comments linking this paper, per year.

Discussed in comments17 threads

  1. What Emily Bender meant by "stochastic parrots" 1 comment

    • I am no linguist, but I believe this is referred to as surface structure and deep structure. What you describe is a line of text that is grammatically somewhat adequate to pass as readable and you treat all text the same. when we read the text, we decipher meaning out of…

      delis-thumbs-7eView on HNParent context

  2. What happens when people don't understand how AI works 1 comment

    • > The article skirts around a central question: what defines humans? Specifically, intelligence and emotions? > The entire article is saying "it looks kinds like a human in some ways, but people are being fooled!" > You can't really say that without at least attempting the…

      _petroniusView on HNParent context

  3. The Illusion of Thinking: Strengths and limitations of reasoning models [pdf] 1 comment

    • A slightly less cynical take is that they want to temper expectations for the capabilities of LLMs in people’s day-to-day lives, specifically in the context of Apple products. A “smarter Siri” is never going to be an autonomous personal assistant à la Jarvis from Iron Man, which…

      MontyCarloHallView on HNParent context

  4. Understanding Reasoning LLMs 1 comment

    • What everyone needs to understand about [reasoning] LLMs is that LLMs can’t reason. https://arxiv.org/pdf/2410.05229

      efitzView on HNParent context

  5. 30% drop in O1-preview accuracy when Putnam problems are slightly variated 3 comments

    • I think you are probably referring to the following paper: https://arxiv.org/abs/2410.05229

      intelkishanView on HNParent context

    • this work is similar to the GSM symbolic paper (applied to putnam) https://arxiv.org/html/2410.05229v1 going forward, llm performance must be reported on the confounded benchmark as well

      bwfan123View on HNParent context

    • >we don't know that they're capable of reasoning Apple AI researchers released a paper on it. They say no. https://arxiv.org/pdf/2410.05229

      ClubberView on HNParent context

  6. GPT-5 is behind schedule 1 comment

    • It's not like we discovered hot air ballons, and some people think we'll get to Moon and Mars with them... > Does person 1 have the burden of proof that rockets can be used to put things in orbit? Sure, but that doesn't make the reasoning used by person 2 valid to explain why…

      coldteaView on HNParent context

  7. OpenAI O3 breakthrough high score on ARC-AGI-PUB 1 comment

    • NovelQA is a great one! I also like GSM-Symbolic -- a benchmark based on making _symbolic templates_ of quite easy questions, and sampling them repeatedly, varying things like which proper nouns are used, what order relevant details appear, how many irrelevant details (GSM-NoOp)…

      loxiasView on HNParent context

  8. What happens if we remove 50 percent of Llama? 1 comment

    • https://arxiv.org/abs/2410.05229 An 18% drop in accuracy (figure 8) is not insignificant. Even 4o suffered 10% loss (figure 6), and 4o isn't a small llm. Competent performance should have near zero performance loss. The simplest benchmark merely changes things like "john had 4…

      zamalekView on HNParent context

  9. OK, I can partly explain the LLM chess weirdness now 1 comment

    • Mathematical reasoning is the most obvious area where it breaks down. This paper does an excellent job of proving this point with some elegant examples: https://arxiv.org/pdf/2410.05229

      griomnibView on HNParent context

  10. OpenAI, Google and Anthropic are struggling to build more advanced AI 1 comment
  11. Nvidia Rides AI Wave to Pass Apple as Largest Company 1 comment

    • Thanks, that is probably the one. Here is a link to the paper itself on arXiv: https://arxiv.org/abs/2410.05229

      ninjinView on HNParent context

  12. OpenAI builds first chip with Broadcom and TSMC, scales back foundry ambition 1 comment

    • did you not read https://arxiv.org/pdf/2410.05229 "We hypothesize that this decline is due to the fact that current LLMs are not capable of genuine logical reasoning; instead, they attempt to replicate the reasoning steps observed in their training data."

      turtlesdown11View on HNParent context

  13. Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku 1 comment

    • Indeed, test data like this constantly leaks into the training data, so these leaderboards are not necessarily representative for real-world problems. A better approach is to use variable evaluation like GSM-Symbolic (for evaluating mathematic reasoning)…

      carschnoView on HNParent context

  14. Sabotage evaluations for frontier models 1 comment

    • These are interesting tests but my question is, why are we doing them? LLM models are not “intelligent” by any meaningful measurement- they are not sapient/sentient/conscious/self-aware. They have no “intent” other than what was introduced to them via the system prompt. They…

      efitzView on HNParent context

  15. Language is not essential for the cognitive processes that underlie thought 1 comment

    • I feel you are putting too much emphasis on the importance and primacy of having a definition of words like 'reasoning'. As humanity has struggled to understand the world, it has frequently given names to concepts that seem to matter, well before it is capable of explaining with…

      mannykannotView on HNParent context

  16. ChatGPT's Name Bias and Apple's Findings on AI's Lack of Reasoning 1 comment

    • Apple's study suggested a 0.3% drop in benchmark results for GPT-4o from fuzzying the prompts, which to me really means no drop and actually shows the benchmark results are more robust than I'd previously expect. It also only looked at GPT-4o, ignoring the allegedly more…

      airstrikeView on HNParent context

  17. Cwal: Scripting Engine Without a Language 1 comment

Submitted to Hacker News

Connected through shared HN threads, not inferred similarity or a recommendation.