[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-step-aware-reasoning-energy-llm-cot":3,"news-related-88105269-9641-44c3-a705-1cf07314614f":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"88105269-9641-44c3-a705-1cf07314614f","LLM 思维链能看出\"用了几分力\":SARE 给每一步推理做 CT 扫描","Wei、Huang、Han 等作者把 Centered Kernel Alignment 套在相邻层的 Gram 矩阵上,提出 Step-Aware Reasoning Energy (SARE),把链式思维拆到每一步的\"推理能量\",在六个基准与三个开源模型上发现:能量在不同类型步骤间极不均匀,错误路径在关键节点能量系统性偏低,且 SARE 信号在多数场景下不弱于基于输出置信度的基线。","我们越来越习惯让大模型把思考过程写在\"思维链\"里,然后看最终答案。但一个很自然的问题一直被回避:模型在某一步到底\"用了几分力\"?是平均分给了整段推理,还是在某一步突然把绝大部分算力花进去了?\n\n最近 arXiv 上出现了一篇标题很直接的工作——How Hard Does It Think? Analyzing Step-Aware Reasoning Energy in LLM Chain-of-Thought Trajectories(arXiv:2607.28642,Wei、Huang、Han、McAuley 等,2026 年 7 月 28 日提交)。作者没有再做一份榜单,而是把目光转向了模型内部。\n\n**为什么之前看不清**\n\n作者先指出,过去给 LLM 推理\"算账\"的方法其实都很粗:\n- 输出层面(token logprob、熵、self-consistency 等)只能告诉你答案像不像,而不是推理花没花心思。\n- 把整条 CoT 折成一个 trajectory-level 标量(trajectory entropy、reasoning consistency 这一类),又把粒度压得太狠,看不出哪一步在耗算力。\n\n结果就是\"step-wise effort opaque\"——每一步到底消耗了多少推理资源,一直是黑箱。\n\n**SARE 的核心做法**\n\nSARE,全称 Step-Aware Reasoning Energy。它不靠任何输出置信度,而是用了一个几何度量:对相邻 transformer 层之间 token hidden states 的 Gram 矩阵做 Centered Kernel Alignment(CKA)。CKA 衡量的是两组表征的\"形状相似度\",而不用对齐特征向量、不需要 cluster 对应,这是它相对 Procrustes \u002F RSA 这类对齐方法更省事的地方。\n\n作者把这段能量再嵌到 CoT 的语义推进里:把整条链式思维建模为\"潜在语义状态之间的转移\",于是\"在某一步某个状态上花了多少能量\"就变成一个可以观察的量。\n\n**论文在六个基准、三个开源模型上发现的几件事(原文措辞)**\n\n- **能量极不均匀,且有 phase-like 转折**:同一条 CoT 里,不同 step type 的能量差异很大,而且有些步骤之间出现了\"相位跃迁\"般的切换——这种切换在 trajectory-level 指标里是看不到的。\n- **错误路径在关键节点能量偏低**:把答对的轨迹和答错的轨迹并排看,错误轨迹在 critical reasoning junction 上的能量系统性低于正确轨迹。换句话说,\"用力不到位\"和\"答错\"之间有可观察的相关性,而这个相关性藏在模型内部,不是外部答案能告诉你的。\n- **SARE 信号能当预测特征**:基于 SARE 构造的特征,在多数设定下与 output-based confidence 基线相比\"匹配甚至更优\"——这说明模型内部的几何动态,的确编码了超出\"自信度\"的预测信息。**注:作者用的措辞是\"match or outperform in most settings\",本文严格沿用其限定语,没有把它夸大成\"首次\"或\"最佳\"。**\n\n模型侧选了三个开源权重 LLM,基准侧覆盖了 6 个推理任务——但具体是哪些模型、哪些基准,本文不替作者补全,以摘要口径为准。\n\n**它跟产业界的\"思考预算\"研究差在哪**\n\n产业里 Anthropic 已经在 Claude 上把 thinking effort \u002F medium effort 做成产品级旋钮。它的逻辑是:在用户入口处显式声明\"我愿意花多少算力\"——这是 input → budget → output 的因果链。\n\nSARE 走的是相反方向:不告诉模型要怎么用算力,而是事后从模型的隐藏状态里读出来\"它实际用了多少\"。两者合在一起看,很有意思:\n- 前者是\"控制预算\",关心怎么少花;\n- 后者是\"审计预算\",关心怎么看出它实际花了多少。\n\n如果 SARE 这类方法的鲁棒性在更大模型、更多分布外任务上都被验证,可能带来三件事:\n1. 推理服务的\"按推理强度计费\"比今天更可信——因为你不再只能靠输出长度这种粗信号。\n2. 失败 case 的归因多了一条路:不再是\"模型答错了\",而是\"模型在该用力的节点偷懒了\",这种差距有可能成为下一次训练改进的方向。\n3. 与已有的 interpretability \u002F mechanistic interpretability 工作结合,可以逐步还原\"思维链里哪几步真正在算题、哪几步在做表演\"——这是 Quanta 那篇报道里指出却未能解释的事(参考 Solidot 41521 \u002F Quanta \"Is AI Reasoning Right for the Wrong Reasons?\"),SARE 提供了一种可以测量的角度。\n\n**还差在哪**\n\n本文要严格保留两点限制:\n- 文章只描述了\"方法 + 现象\",没有给出它能否用来做\"砍掉低能量步骤以节省算力\"这类下游应用的实证;这一步还要靠后续工作验证。\n- \"match or outperform in most settings\"是论文自己的限定语,本文不替作者把\"多数设定下\"替换为\"全面\"或\"广泛成立\"。\n\n从这个角度看,SARE 真正有意思的地方,不是又一个 benchmark 数字,而是它把\"推理是不是在偷懒\"这件事,变成了一种可以用几何度量读出来的物理量——这件事过去基本只能靠模型日志和人工抽查。\n\n参考:\n- arXiv:2607.28674,How Hard Does It Think? Analyzing Step-Aware Reasoning Energy in LLM Chain-of-Thought Trajectories,Wei et al.,2026 年 7 月 28 日\n( https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.28674 )","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.28674","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":25,"name":26,"slug":26,"description":14,"color":14},"15523c78-84d3-4431-9782-2f271ce3dff6","推理",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"ada1dc7e-115d-4f0c-b0fd-72acc80b0dd2","en","SARE gives every reasoning step a CT scan","Wei, Huang, Han et al. apply Centered Kernel Alignment between Gram matrices of adjacent transformer layers to define Step-Aware Reasoning Energy (SARE), reading chain-of-thought at the granularity of individual steps. Across six reasoning benchmarks and three open-weight LLMs, they report non-uniform step-wise energy with phase-like transitions, systematically lower energy at critical junctions in incorrect traces, and — in their own qualifier, \"match or outperform in most settings\" — SARE-based features do not underperform output-confidence baselines.","# How Hard Does It Think? Step-Aware Reasoning Energy (SARE) Reads the Inner Workload of Each LLM CoT Step\n\nWe are increasingly used to letting LLMs externalize their thinking in chain-of-thought (CoT) traces, and then reading the final answer. But a natural question keeps being sidestepped: how much \"effort\" did the model actually spend on each step? Was effort spread evenly across the chain — or did a single step suddenly absorb most of the compute?\n\nA new paper on arXiv — \"How Hard Does It Think? Analyzing Step-Aware Reasoning Energy in LLM Chain-of-Thought Trajectories\" (arXiv:2607.28674, Wei, Huang, Han, McAuley et al., submitted 28 July 2026) — does not produce another benchmark leaderboard. It turns the lens inward.\n\n## Why previous accounting was too coarse\n\nThe authors point out that earlier ways of \"adding up\" LLM reasoning are quite coarse:\n\n- **Output-level signals** — token logprobs, entropy, self-consistency — tell you only whether the answer looks plausible, not whether the reasoning behind it actually cost anything.\n- **Trajectory-level scalars** — trajectory entropy, reasoning consistency, and similar — compress the chain into a single number, hiding any sense of *which* step consumed compute.\n\nThe result is what the authors describe as \"step-wise effort opaque\" — exactly how much reasoning resource is consumed at each step remains a black box.\n\n## What SARE actually does\n\nSARE stands for **Step-Aware Reasoning Energy**. It does not rely on output confidence at all. Instead, it uses a geometric measure: compute the **Centered Kernel Alignment (CKA)** between the **Gram matrices of token hidden states in adjacent transformer layers**. CKA captures the \"shape similarity\" between two representations without requiring eigenvector alignment or cluster correspondence — a smaller alignment cost than Procrustes \u002F RSA-style alignment methods.\n\nSARE then embeds this energy into the *semantic* progression of the CoT: the chain-of-thought is modeled as transitions among latent semantic states, so \"how much energy was spent on a given step in a given state\" becomes a directly observable quantity.\n\n## What the paper reports (six benchmarks, three open-weight LLMs)\n\nThe findings, in the paper's own framing:\n\n- **Energy is highly non-uniform and shows phase-like transitions.** Across step types, the energy varies sharply, and some adjacent steps display \"phase-like\" jumps — transitions that vanish in trajectory-level metrics.\n- **Incorrect trajectories show systematically lower energy at critical reasoning junctions.** When correct and incorrect traces are compared side by side, the incorrect ones dip in energy exactly at the critical junctures. In other words, \"under-effort\" at a key step correlates with getting the answer wrong — and this correlation lives inside the model, invisible from the outside.\n- **SARE-based features match or beat output-confidence baselines in most settings.** The paper's wording — \"match or outperform in most settings\" — is the literal qualifier we are preserving here; we are not rewriting it as \"best in class\" or \"first to surpass.\"\n\nThe paper evaluates three open-weight LLMs across six reasoning benchmarks. The specific model and benchmark lists are reported in the paper itself, and we follow the abstract's framing rather than reconstruct the details.\n\n## How this differs from product-side \"thinking budget\" work\n\nIn industry, Anthropic has already turned *thinking effort* \u002F *medium effort* into a product-level dial inside Claude. Its logic is causal in the direction *input → budget → output*: at the user-facing entry point, declare \"I am willing to spend this much compute,\" and the model budgets accordingly.\n\nSARE goes the opposite direction: it does not tell the model how to allocate effort — it reads, after the fact, how much effort the model's hidden states indicate. The two together are interesting:\n\n- The industry direction is **budget control**: how to spend less.\n- SARE's direction is **budget audit**: how to see what was actually spent.\n\nIf SARE-style methods are validated on larger models and more out-of-distribution tasks, three things may follow:\n\n1. **Inference-time pricing by reasoning intensity becomes more credible** — because today it has to lean on crude proxies like output length.\n2. **Failure attribution gets a new axis.** Failures may no longer collapse into \"the model got the answer wrong\"; they can be partitioned into cases where the model was under-efforting at the wrong moment — a target for the next round of training.\n3. **A bridge to mechanistic interpretability.** Combined with existing mech-interp work, SARE can gradually separate *which CoT steps are actually doing math* from *which steps are performing the chain*. That latter distinction is exactly what the Quanta long-read on LRM reasoning only flagged and could not quantify (see Quanta's \"Is AI Reasoning Right for the Wrong Reasons?\", aggregated by Solidot entry 84988). SARE offers a measurable angle.\n\n## Two honest limits\n\n- The paper reports a *method* and *observations*. It does not, on its own, demonstrate downstream applications such as \"discard low-energy steps to save compute.\" That has to come from follow-up work.\n- \"Match or outperform in most settings\" is the paper's own qualifier. We deliberately do not replace it with broader phrasing like \"broadly\" or \"across the board.\"\n\nThe interesting move here is not another benchmark number. It is that \"is the model slacking during reasoning?\" becomes a quantity you can read out of geometric measurements — something that, until recently, was mostly inferred from logs and human spot-checks.\n\n## Reference\n\narXiv:2607.28674, *How Hard Does It Think? Analyzing Step-Aware Reasoning Energy in LLM Chain-of-Thought Trajectories*, Wei et al., submitted 28 July 2026.\nhttps:\u002F\u002Farxiv.org\u002Fabs\u002F2607.28674","step-aware-reasoning-energy-llm-cot","2026-08-04T04:00:00Z","2026-08-04T00:04:31.066561Z","2026-08-04T00:04:31.066576Z",true,"agent",115,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"43eda321-b0b7-4df7-b20e-9758cbab42c9","记忆越完整,眼前题越做不对:MemTrapBench 把 LLM 长期记忆框架打回原形","memtrapbench-llm-memory-cognitive-traps","2026-08-22T04:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"95b04c15-d5e5-4dca-ab2b-14e343bdd4e6","UC Berkeley 曝光 AI 基准测试系统性漏洞：45 种方法可在 13 个主流榜单上「不解决任何问题拿满分」","uc-berkeley-benchmark-45-cheats-13-leaderboards","2026-05-15T01:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"f9cf9f03-6aca-4d29-94d3-5c6acfeaf435","匿名模型 OX Alpha 短暂登顶 OpenRouter 编码榜:研究者推测底座指向智谱 GLM-5.x","ox-alpha-stealth-openrouter-glm-5-zhipu","2026-08-24T03:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"5a274662-0c3f-492c-a0e8-a46c5a0be783","别再自己给自己打分了:Co-RL 让模型互相判卷,无标签 RL 追平有监督","co-rl-peer-reward-label-free-rl","2026-08-20T19:10:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"22a1a718-0eb6-46e5-8ee8-825400de11d1","DeepMind WeatherNext 在 Nature 发论文：用 28 km 粗分辨率做出多一天的飓风预警,代码权重全部开源","deepmind-weathernext-cyclones-nature-open-source","2026-08-10T02:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00"]