When a language model is optimized for reasoning, does it still show embers of autoregression? An analysis of OpenAI o1
Source
- Person key:
ysymyth - Source kind:
paper - Canonical URL: https://arxiv.org/abs/2410.01792
- License:
NOASSERTION - Public handling:
public-metadata-summary-hash-link-only - Semantic hash:
dbd7d95938575b8a699050a8979682b27af11be1987bf6d2dd6c32dc6ea7be37 - First seen: 2026-05-16
- Last changed: 2026-05-16
- Identity guard: Do not confuse with yao-shunyu-alfred, the physics-to-AI researcher at alfredyao.github.io.
Classification
- Category: Reinforcement learning and reasoning
- Topic hub: shunyu-yao-public-corpora
- Project taxonomy: shunyu-yao-project-taxonomy
- Paper map: shunyu-yao-paper-map
Summary
In “Embers of Autoregression” (McCoy et al., 2023), we showed that several large language models (LLMs) have some important limitations that are attributable to their origins in next-word prediction. Here we investigate whether these issues persist with o1, a new system from OpenAI that differs from previous LLMs in that it is optimized for reasoning. We find that o1 substantially outperforms previous LLMs in many cases, with particularly large improvements on rare variants of common tasks (e.g., forming acronyms f…
What This Teaches
How learning, planning, and feedback loops shape agent behavior and model capability.
Related
Public Handling Notes
- EXTRACTED: Metadata and links are from public sources.
- INFERRED: Unclear-license full text, PDFs, and source code should not be mirrored into public wiki pages.
- AMBIGUOUS: Items discovered by keyword search are still separated by source identity and category heuristics.