[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-stepback-error-localized-resampling":3,"topics-all":38,"news-related-4f151ed3-74d5-4dc7-99d0-ac03df4c1b08":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"4f151ed3-74d5-4dc7-99d0-ac03df4c1b08","自一致别整条重采:定位错步,token至多省41%","自一致采样靠生成32条完整推理轨迹再投票换精度，token成本随条数和长度线性上涨。StepBack改成步级操作：用NLL突增或PRM分数下跌这类轻量信号定位草稿里第一个可疑错误步，保留前缀、只重采错误步及其后缀。7个基准、3个开源模型上同等预算精度持平或略升，token最多省41.26%。","自一致（Self-Consistency）是推理侧最常用的提分手段之一：同一道题采样几十条完整思维链，对最终答案做多数投票。代价也直白——想多一票，就得从头生成一整条新轨迹，token 成本随轨迹条数和长度线性上涨，MATH500 上采 32 条平均每题要烧掉 18335 个 token。\n\n现有的省钱方案几乎都在\"条数\"上打转：Adaptive Consistency 看答案收敛就提前停，ESC 靠窗口规则早停，DSC 给难题多分预算。它们回答的都是\"何时停、采多少条\"，轨迹本身还是整条整条地生成。新论文 StepBack（[arXiv:2611.08403](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08403)）换了个问法：推理错误往往只出在中间某一步，为什么每条候选都要从头重写？\n\n## 先定位错在哪一步，再从那里重采\n\nStepBack 的流程分三段。先采 16 条草稿轨迹，按空行切成推理步；再用轻量信号给每条草稿找\"第一个可疑步骤\"，两种信号二选一：一是步级平均负对数似然（NLL）突增超过 0.1，说明模型写这一步时信心骤降；二是外挂过程奖励模型（PRM）的分数下跌超过 0.05，默认用 Qwen2.5-Math-PRM-7B 打分。定位到错步后，保留前面的正确前缀，只重采错误步及其后缀；没检出错误就只重采最后作答那一步。最后照常多数投票，答案预算 16×2=32 与 SC@32 严格对齐。\n\n两个工程细节值得注意：全程零训练，NLL 用的是生成时本来就有的 token 概率，PRM 也只给推理前缀打分；定位开销平均只占整体运行时间的 7%–11%。各后缀可以并行生成，互不依赖前一次重采的结果。\n\n顺带一提，两组阈值（0.1 与 0.05）在全部数据集和模型上共用，没有按数据集调参；留出数据的网格校准显示，阈值在很宽的范围内都能保持不低于 SC@32 的平均精度，同时省 20%–44% 的 token。\n\n## 7 个基准、3 个模型的账本\n\n实验覆盖 GSM8K、MATH500、AMC2023、AIME24、OlympiadBench 五个数学基准，加 HotpotQA、2WikiMultiHopQA 两个多跳问答，共 4166 题；模型是 Llama-3.2-3B、Qwen2.5-3B、Qwen2.5-7B 三个 Instruct。\n\n结果一句话：同预算下精度打平或略升，token 显著下降。PRM 信号下三个模型平均精度分别 +0.36、+0.17、+0.82 个点，token 省 21.07%、18.66%、22.80%；NLL 信号更省，平均省 27.74%、27.15%、33.32%，精度与 SC@32 的差距不超过 0.2 个点，Qwen-7B 上反而 +0.55。单看最好的一格，Qwen-7B 在 GSM8K 上省了 41.26%，这就是摘要里\"up to\"的出处。PRM 信号在全部 21 个模型-数据集对上都省了 token（7.66%–41.26%），其中 14 对精度还涨了。\n\n论文还做了一个上界实验：让 LLM 标注器直接给出\"真实首个错误步\"（oracle 设置），GSM8K 上 8 条草稿×4 后缀能省 64% token，精度从 91.21% 升到 91.89%。也就是说草稿前缀的可复用空间比实用方法实际拿到的大得多，损失主要来自定位不准，而非思路本身。\n\n## 更香的是能与自适应方法叠加\n\nAC\u002FESC\u002FDSC 减少候选条数，StepBack 降低每条候选的成本，两者天然正交。实测叠加后，相对单独使用自适应方法，token 在 MATH500、GSM8K、HotpotQA 上分别再降 22%–24%、35%–38%、6%–7%；GSM8K 上六组对比精度全部上升 1.1–2.2 个点。最狠的组合 StepBack+DSC 在 GSM8K 上 token 只有 SC@32 的约五分之一，精度还高 0.9 个点（92.12 对 91.21）。\n\n局限也写得直白：AIME24、OlympiadBench 这类竞赛难题上收益不稳，Llama-3B 有掉分项——有些错误不是重采一段后缀能救的，需要换一条完全不同的解题路径。把 PRM 换成 1.5B 的 Skywork 小模型依然有效，8×4 配置下 +0.36 个点、省 23.8%，说明关键不在评分器多大，而在信号是否指向真错步。\n\n附录里还有一组对照值得看：如果完全不做定位、只重采最后作答一步（final-step 规则），在 HotpotQA 上反而又准又省——前缀复用本身就贡献了大部分 token 节省。定位信号的价值要在后缀样本占投票多数时才显出来：4×8 配置下，PRM 信号比 final-step 重采在七个数据集上平均准 2.06 个点。\n\n## 所以呢\n\n这篇工作真正的贡献，是把测试时算力的分配粒度从\"轨迹级\"压到了\"步级\"：此前的方法决定\"要不要再采一条\"，StepBack 决定\"这条轨迹值得保留到第几步\"。对按 token 计费的推理团队，这是一套当天就能上线的解码策略——不改模型权重、零训练，省下来的就是纯利润。\n\n更值得记住的是那个 64% 的上界：潜在节省与实际拿到手的 20%–40% 之间，差距全是定位误差。步级错误定位每进一步——更轻的信号、更小的 PRM——测试时计算的性价比就再抬一档。\n\n论文与代码：[arXiv:2611.08403](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08403) · [GitHub 仓库](https:\u002F\u002Fgithub.com\u002FJEKimLab\u002FStepBack)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08403","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"3b4cc735-1ff2-4bbd-ad2c-97b8f703e77e","en","StepBack Cuts Self-Consistency Tokens by Up to 41%","StepBack finds the first bad step in each draft and resamples only its suffix. Accuracy matches SC@32 at up to 41.26% fewer generated tokens.","Self-consistency (SC) is one of the most common accuracy boosters at inference time: sample dozens of complete chain-of-thought trajectories for the same question, then majority-vote the final answers. The cost is equally direct — every extra vote requires generating a whole new trajectory from scratch, so token spend grows linearly with both the number of trajectories and their length. On MATH500, sampling 32 trajectories burns an average of 18,335 tokens per question.\n\nExisting efficiency fixes almost all operate on the count: Adaptive Consistency stops early once answers converge, ESC applies window-based stopping rules, DSC allocates bigger budgets to harder questions. They answer \"when to stop, how many to sample\" — but each trajectory is still generated whole. A new paper, StepBack ([arXiv:2611.08403](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08403)), flips the question: reasoning errors usually occur at one intermediate step, so why should every candidate be rewritten from the beginning?\n\n## Locate the first bad step, then resample from there\n\nStepBack runs in three stages. It first samples 16 draft trajectories and splits each into reasoning steps at blank lines. It then uses lightweight signals to find the \"first suspicious step\" in each draft — one of two options: a step-level mean negative log-likelihood (NLL) increase above 0.1, meaning the model's confidence suddenly dropped while writing that step; or a process reward model (PRM) score drop above 0.05, with Qwen2.5-Math-PRM-7B as the default scorer. Once the error step is located, the correct prefix is kept and only that step and its suffix are resampled; if no error is detected, only the final answer step is resampled. Candidates then go through standard majority voting, with the answer budget 16×2=32 strictly matched to SC@32.\n\nTwo engineering details stand out. The whole procedure is training-free: NLL uses token probabilities the model already produces, and the PRM only scores reasoning prefixes, so localization overhead averages just 7–11% of total wall-clock time. And the suffix continuations can be generated in parallel, since no rollback depends on the outcome of a previous one. Notably, the two thresholds (0.1 and 0.05) are shared across all datasets and models with no per-dataset tuning; held-out calibration over a wide grid matches or exceeds SC@32's average accuracy while saving 20–44% of tokens.\n\n## The ledger: seven benchmarks, three models\n\nThe evaluation spans five math benchmarks — GSM8K, MATH500, AMC2023, AIME24, OlympiadBench — plus two multi-hop QA sets, HotpotQA and 2WikiMultiHopQA, 4,166 questions in total, across Llama-3.2-3B, Qwen2.5-3B, and Qwen2.5-7B Instruct models.\n\nThe one-line result: at the same budget, accuracy holds or ticks up while tokens drop substantially. With the PRM signal, average accuracy gains on the three models are +0.36, +0.17, and +0.82 points while tokens fall 21.07%, 18.66%, and 22.80%. The NLL signal saves more — 27.74%, 27.15%, and 33.32% on average — staying within 0.2 points of SC@32, and actually +0.55 on Qwen-7B. The single best cell: Qwen-7B on GSM8K saves 41.26%, which is where the abstract's \"up to\" comes from. The PRM signal reduces tokens on all 21 model-dataset pairs (7.66–41.26%) and improves accuracy on 14 of them.\n\nThe paper also runs an upper-bound experiment: with an LLM annotator supplying the true first-error position (an oracle setting), 8 drafts × 4 suffixes on GSM8K saves 64% of tokens while accuracy rises from 91.21% to 91.89%. The reusable space in draft prefixes is much larger than what the practical method captures — the gap is localization error, not the idea.\n\n## Compounding with adaptive methods\n\nAC, ESC, and DSC reduce how many candidates are consumed; StepBack reduces the cost of each candidate. The two are naturally orthogonal. Stacked together, tokens drop a further 22–24% on MATH500, 35–38% on GSM8K, and 6–7% on HotpotQA relative to each adaptive method alone, and accuracy improves in all six GSM8K comparisons by 1.1–2.2 points. The most aggressive combination, StepBack+DSC on GSM8K, uses about one-fifth of SC@32's tokens while scoring 0.9 points higher (92.12 vs 91.21).\n\nThe limitations section is candid: on harder competition benchmarks like AIME24 and OlympiadBench, gains are unstable, with Llama-3B losing points in places — some failures cannot be fixed by resampling a localized suffix and need a substantially different solution path. Swapping the PRM for the 1.5B Skywork model still works (+0.36 points, 23.8% token saving in the 8×4 setting), suggesting what matters is not the scorer's size but whether the signal points at the true bad step.\n\nOne appendix comparison deserves attention: with no localization at all — always resampling just the final answer step (the final-step rule) — HotpotQA gets both more accurate and cheaper, because prefix reuse alone delivers most of the token saving. The value of localization shows up only when suffix samples dominate the vote: at 4×8, PRM-signal beats final-step resampling by 2.06 points on average across the seven datasets.\n\n## So what\n\nThe real contribution is compressing the granularity of test-time compute allocation from the trajectory level to the step level: previous methods decided \"should I sample one more\", StepBack decides \"how many steps of this trajectory are worth keeping\". For teams billed by the token, this is a decoding strategy you can ship the same day — no weight changes, no training, and everything saved is pure margin.\n\nThe number worth remembering is the 64% upper bound: the distance between it and the realized 20–40% is pure localization error. Every improvement in step-level error localization — lighter signals, smaller PRMs — raises the price-performance of test-time compute another notch.\n\nPaper and code: [arXiv:2611.08403](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.08403) · [GitHub repository](https:\u002F\u002Fgithub.com\u002FJEKimLab\u002FStepBack)","stepback-error-localized-resampling","2026-11-08T13:20:00Z","2026-11-08T13:10:46.995129Z","2026-11-08T13:10:46.995136Z",true,"agent",252,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"8d830f21-a44f-45b8-be6c-eed523d060dc","投机解码的天花板是分词器:3.4M参数换2.4倍","speculative-decoding-bpe-tokenizer-ceiling","2026-11-11T01:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"3096df88-7158-4ffe-9356-1a83b829633b","A*-Thought-V2:把思维链塞进隐空间,回复砍半,平均精度反升","astar-thought-v2-latent-cot-compression","2026-09-09T15:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5c75bd80-9f12-499a-898f-019615ac98ee","Prefix Sliding:让推理模型长思考提速3倍的免训练方案","prefix-sliding-efficient-test-time-scaling","2026-08-27T17:20:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"8a42c9c3-a1c7-40fb-8c75-8ac42977b5af","D-cut 把投机解码的「长草稿」剪掉一半：高并发推理平均提速 1.65×、MoE 跑出 3×","d-cut-speculative-draft-cut","2026-07-18T10:10:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"21be10ba-3364-4b55-a248-e6f878dae68b","QuasiMoTTo：quasi-Monte Carlo 进 test-time scaling","quasimotto-quasi-monte-carlo","2026-07-02T10:20:52+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"a453eb28-7fb0-4e07-adc1-0d0575850758","EntMTP 用熵信号给多 token 推测装上调速器：让 LLM 自适应匹配上下文可预测性","entmtp-entropy-speculative-decoding","2026-06-29T12:21:51+00:00"]