Large Language Model Reasoning Failures
Large Language Model Reasoning Failures
Citation
- Authors: Peiyang Song, Pengrui Han, and Noah Goodman
- Year: 2026
- Venue: Transactions on Machine Learning Research
- arXiv: arXiv:2602.06176
- URL: https://arxiv.org/abs/2602.06176
One-Line Takeaway
This survey organizes LLM reasoning failures into a structured taxonomy and is useful here mainly as a map of failure modes that RLVR papers often diagnose only in narrow slices.
Technical Core
- The paper proposes a two-axis taxonomy: reasoning types (informal, formal, embodied) crossed with failure types (fundamental failures, application-specific limitations, and robustness issues).
- It collects evidence that reasoning failures recur across working memory, compositional reasoning, arithmetic, social reasoning, physical reasoning, and robustness to superficial perturbations.
- A recurring theme is that many failures come from architecture, training objectives, and lack of grounding rather than from a single benchmark artifact.
Method
- This is a survey and synthesis paper rather than a new training method.
- Its contribution is the taxonomy, the root-cause analysis, and the survey of mitigation strategies already proposed in the literature.
Assumptions And Limits
Because it is a broad survey, it does not itself resolve the RLVR-specific question of whether reinforcement learning broadens support or merely sharpens it. Its value is mainly organizational and comparative.
Relevance To This Project
The paper is useful as background because the repo's motivation sits inside a broader class of reasoning failures: not just final-answer error, but failures of exploration, abstraction, transfer, and robustness. It helps keep the project's latent-strategy framing connected to the larger literature on why present LLM reasoning breaks.
Comparison Notes
Most papers in this llm-rlvr cluster study one bottleneck in detail, such as support shrinkage (Wu et al. (2025)) or hard-problem exploration (POPE). This survey is useful because it places those RLVR-specific issues inside a wider map of formal, informal, and embodied reasoning failure modes.
Retrieval: alphaXiv MCP get_paper_content on arXiv:2602.06176; no local cache or downloaded source artifacts were created.
Verified from source: verification in this ingestion pass was limited to the alphaXiv whole-paper report and arXiv metadata exposed through that report. I did not run a separate targeted PDF query for this note.
