--- title: "A new AI paper claims a 2.29x efficiency gain. The real number is closer to 1.21x." description: "A cited look at a new AI reasoning-compression paper: a real method, verified numbers, and one headline figure that bundles two separate effects into one." publishedAt: 2026-09-11 author: Ena Pragma url: https://enapragma.co/field-notes/astar-thought-v2-efficiency-claim tags: ["ai-research", "verification", "llm-reasoning"] --- A new AI research paper claims its method makes a model's reasoning 2.29 times more efficient. That number is real, it appears in the paper's own results table, and it is also not a clean measurement of what the paper actually invented. Figuring out why is a useful, transferable skill: almost every AI efficiency claim you will see this year has the same kind of hidden seam in it, and knowing where to look costs about five minutes. ## What the paper actually does The paper is A*-Thought-V2, published on arXiv on September 7, 2026, by 16 researchers from Beijing University of Posts and Telecommunications, The Hong Kong Polytechnic University, Tsinghua University, JIUTIAN Research, and OpenBMB. It tackles a real, unglamorous problem: when a large language model reasons through a hard question, it often writes out a long chain of thought, including dead ends, self-corrections, and false starts ("wait, that's not right, let me try again"). That reasoning is useful for getting the right answer, but it is expensive: every word the model writes costs time and computing power, whether or not that word ends up mattering. The paper's fix is genuinely clever. It takes each step in a model's reasoning trace and represents it as a point in a simplified three-dimensional space, then measures the angle between each small step and the overall direction from the original question to the final answer. Steps that stay closely aligned with that overall direction, the productive, on-track parts of the reasoning, are left as normal, readable text. Steps that swing away from it, the backtracking and self-correction, get compressed into a single internal marker, a vector the model can still use to inform its next steps, but which no longer prints out as words at all. The paper reports a real, checked pattern behind this: reasoning steps with a small angle from the main direction tend to contain words like "simplify" and "therefore," while steps with a large angle tend to contain words like "wait" and "maybe this is wrong." The compression targets exactly the reasoning that reads as a detour, not the reasoning that reads as progress. Every specific accuracy and length number describing this method, checked directly against the paper's own results table rather than taken from any secondhand summary, holds up exactly. On one tested model, accuracy moved from 89.0 to 91.5 percent while the average response shrank from roughly 16,150 tokens to roughly 13,460. On a larger model, accuracy moved from 92.9 to 93.9 percent with a similar drop in length, from about 14,000 tokens to about 11,800. Those are real, verified improvements: a model that reasons only slightly better while writing meaningfully less. ## The number that oversells itself, and how to catch it yourself The paper's own highlighted claim is a 2.29 times gain in what it calls "accuracy per computation unit," a single number meant to summarize the whole efficiency story. Checking that number against the paper's own table shows exactly how it is built, and exactly what it leaves out. The comparison behind 2.29x measures the new method against the completely untouched, out-of-the-box model. But the new method is not trained from that untouched starting point. It is trained on top of an intermediate version of the model that has already been fine-tuned on the same training data used to build the final method, and that intermediate version is already meaningfully more efficient than the untouched original, before the paper's actual invention is added at all. Comparing the finished method only against the very first starting point folds two separate improvements into one number: the general benefit of fine-tuning on that dataset, plus the specific benefit of the new compression technique. The paper's own table contains the fairer comparison, one step over: measured against that already-fine-tuned intermediate version rather than the raw original, the efficiency gain drops to roughly 1.21 times. Both numbers are real and both are in the same table. One of them is the headline. The other is the honest measurement of what the new idea specifically contributes, isolated from everything else that was already true before it was added. This is not a case of a research team fabricating a result. Comparing against the most recognizable baseline, the completely untouched model, is a common and defensible framing choice in this field. But it is also a framing choice that happens to make a smaller, real contribution look larger than it is, and the fix for a reader is always the same: before repeating an efficiency or performance multiplier, ask what it was actually measured against, and whether that starting point already included some of the improvement being credited to the new idea. ## Where the skepticism went slightly too far The research that surfaced this paper, a video from the YouTube channel Discover AI, deserves credit for catching the 2.29x-versus-1.21x gap correctly and explaining it to a general audience before this analysis existed independently. But checking every specific claim against the paper itself, rather than treating any single source as automatically right, also turned up a place where the video is a shade too skeptical. Discussing the method's limits, the video says there isn't "enough data" on whether the underlying idea holds up outside of math problems. The paper's own evaluation section contradicts that specific claim: it tests the method on two genuinely different kinds of questions, general science and commonsense reasoning, and graduate-level science, alongside math, and reports that the core pattern holds across both. What remains genuinely untested, and what the paper itself does not claim to have checked, is whether the method holds up on a different model family entirely or in a language other than the one it was trained and tested in. Those are the real open questions, not the cross-domain gap the video raised. ## Why this is worth five minutes even if you never read a research paper You do not need to understand three-dimensional geometry or model training to use the lesson here. Any time you see a company, a research team, or a headline claim a specific multiplier, faster, cheaper, more efficient, more accurate, the single most useful question is not whether the number is fake. Usually it is not. The useful question is what exact starting point the comparison was measured against, and whether that starting point was a fair, apples-to-apples version of the thing being improved, or a weaker stand-in that makes the improvement look bigger than the specific idea being sold actually earns. That question works on AI research papers, and it works just as well on a vendor's product announcement, a diet study, or a sale that claims fifty percent off. This is also a working example of the discipline worth applying to any single source, technical or not: verify the specific numbers, credit what the source got right, and name plainly what it got wrong in either direction, oversold and underclaimed alike, rather than treating one factual slip as license to distrust everything else a source says. ## What this does not establish A*-Thought-V2 is six days old as of this writing, has not been through peer review, and its public code repository is early-stage. Its accuracy and efficiency numbers come from one paper's own testing setup, on two specific model sizes and one specific training dataset, and have not yet been independently reproduced by anyone outside the team that built it. Whether the geometric pattern it relies on holds up on different model families or in other languages is, by the paper's own account, still an open question. ## Sources - [A*-Thought-V2 paper, arXiv](https://arxiv.org/abs/2609.07821) - [A*-Thought-V2, GitHub repository](https://github.com/AI9Stars/AStar-Thought) - [A*-Thought-V2 on Hugging Face Papers](https://huggingface.co/papers/2609.07821) - ["AI Just Learned to Think Between the Tokens," Discover AI](https://youtu.be/It_8QWMQvaA) - [Coconut: Training Large Language Models to Reason in a Continuous Latent Space (Meta)](https://arxiv.org/abs/2412.06769) - [SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs](https://arxiv.org/abs/2510.05069)