← 返回 2026-10-19

LEV:用循环决策模型实现快而深的思考 LEV: Thinking Fast and Deep with Looped Decision Models

Dingjie Song, Zhiling Yan, Yuxuan Zhang, Lichao Sun 📅 2026-10-14 👍 71 2026-11-11 18:30
System 1.5 决策模型 循环Transformer 推理效率 早退机制 自适应计算

在原生循环骨干上每步读出选项并联训所有深度,让决策模型按题加深循环步数,兼顾延迟与精度

前置知识

决策模型(decision model)

对固定选项集合打分的判别式模型:把状态和带分隔符选项的问题打包进一条序列,一次前向传播后在 $\langle \mathrm{decide}\rangle$ 位置与各选项收尾符的隐状态之间做指针头内积,经 softmax 输出每个选项一个概率,不生成任何文本。代表如 Jev/Kev,单题延迟数十毫秒,概率可用单一温度缩放校准。

LEV 就是这类模型。理解其“一次前向、深度在训练时固定”的本质,才能理解本文要解决的算力分配问题:简单题和难题花同样的计算。

循环 Transformer(looped transformer)

把同一组 $L$ 层解码器 $F_\theta$ 重复跑 $R$ 次:$h^{(t)}=\mathrm{Norm}(F_\theta(h^{(t-1)}))$,以 $L\cdot R$ 层的等效深度换取仅 $L$ 层的参数量。“原生循环”指 Ouro 等模型在预训练阶段(每步都带语言建模损失,共 7.7T token)就以循环形式训练,而非事后把现成模型的若干层圈起来重跑。

LEV 的全部“思考深度”来自重跑原生循环骨干。论文用控制实验证明后补循环基本无效,只有预训练即循环的骨干才“从出生就开着”。

早退与自适应计算

让不同输入消耗不同算力的技术总称:早退(early exit)在某个中间层置信度足够时直接输出;margin/entropy 阈值、patience 规则同理;级联(cascade)则把不确定样本转给更大的第二个模型。已知缺陷是浅层出口未被训练、在新数据上过度自信(confidently wrong)。

LEV 的按需深度与停止器正是对这套思路的改造。文中图 8、表 16 与这些基线规则逐一对比,不懂它们的形态就无法理解对比结论。

指针头(pointer head)

源自 Pointer Network 的打分方式:对问题表示 $h_{dec}$ 与每个选项的收尾符表示 $h_k$ 各做线性投影后计算 $z_k=(W_K h_k)^\top(W_Q h_{dec})/\sqrt{d_p}$,softmax 后即各选项概率。它把“从词表选词”换成“直接在给定选项上形成分布”,避免生成式模型对选项顺序和标签的敏感性。

LEV 在每个循环步后都接同一个共享指针头读出,这是“每步可读、一套权重服务所有深度”的关键部件。

一致性目标与无标注阈值拟合(CATs)

CATs 的思想:不预测“当前答案对不对”,而预测“提前停下会不会与跑完的完整模型结论一致”,即 $y_k=\mathbb{1}[\hat{o}^{(k)}=\hat{o}^{(R)}]$。因为目标由模型自身最深答案定义,拟合分类器与调阈值都不需要任何人工标注;阈值取满足“≥98% 最终答案与最深一致”的最小值。

这是本文停止规则的灵魂,也是“部署方可在自家无标注流量上零成本适配”这一卖点的根据,与在标注数据上调阈值的传统早退形成对照。

双过程理论(System 1 / System 2)

Kahneman 提出的认知框架:System 1 快而省、凭直觉即刻作答;System 2 慢而深、付出串行推理的代价。本文把单次前向的决策模型类比为 System 1、生成式思维链类比为 System 2,并把 LEV 定位为 System 1.5——按题在循环步数上连续调节深浅。

论文标题 “Thinking Fast and Deep” 与全部叙事都建立在这对概念上,理解它才能把握文章的定位语言。

研究动机

决策模型(如 Jev/Kev)一次前向就给出每个选项的概率,速度快但计算深度在训练时固定:简单题和难题消耗同样的算力。作者在 JevBench 的 231 道公开题上逐题对比同源训练的 Kev-0.8B 与 Kev-4B:后者多答对 28 题,且增益全部来自 standard 与 hard 档(easy 档两者均已全对),而两模型结论不同的只有 42 题——大模型的额外参数被浪费在大多数题上。现有加算力的方案各有硬伤:级联要重读输入并需要第二个模型(Kev-0.8B 中位延迟 79ms、Kev-4B 108ms,被升级的请求比大模型单独跑还慢);生成式思考在 hard 档提升 9–13 分,但 hard 题中位延迟慢 3–4 倍,且简单题也要为推理文本付费(djev 思考模式把 48 道 easy 题做错 2 道,534 个决策里 72 个超 8,192 token 预算无有效答案);中间层早退的浅层读出没被训练过,在新数据上自信地出错;把循环“改装”到普通预训练模型上,ReZero 门控训练后平均幅度仅约 0.006,循环基本不开启,去掉门控反而让新数据掉 2.2 分。

本文的目标是本文要造一个“按需深度”的决策模型:同一个 checkpoint 在每道题上可选择循环步数 $k\in\{1,\dots,R\}$,简单题一次前向即答,难题自动多跑几步;加深的过程不生成 token、不引入第二个模型、不重读输入(从上一步隐状态继续)。同时,任一深度的读出都必须是训练过的合格决策模型(而非顺便看一眼的中间层),使一个模型覆盖延迟-精度曲线上的多个工作点;决定“何时再跑一步”的停止器,其阈值必须能在完全无标注的真实流量上拟合,让部署方零成本适配自己的数据分布。

与已有工作不同的是,独特切入有三点。其一,不在预训练模型上后补循环(作者实测门控趋零、循环关死),而是直接构建在预训练阶段就以循环形式训练的原生循环语言模型(Ouro、Nanbeige4.2)之上,决策微调只需“激活”而非“塑造”循环。其二,在每一步循环后都用同一个共享指针头读一次选项,并把 $R$ 步的 log loss 联合平均训练($\mathcal{L}=\frac{1}{R}\sum_t \ell(p^{(t)},y)$),使每个深度都是合格的决策模型。其三,停止器不预测“当前答案是否正确”(浅层在新数据上过度自信、正确性阈值迁移差),而预测“当前答案是否与最深答案一致”(CATs 一致性目标);该目标源自模型自身,因此阈值可以在无标注请求上以“≥98% 与最深答案一致”的准则拟合,这是此前早退文献没有在决策模型上做过的组合。

核心方法

直觉:把“再想一会儿”从生成文本改成重跑同一组层。LEV 把决策模型的骨干换成原生循环语言模型,让同一解码栈 $F_\theta$ 跑 $R$ 次,每步之后都用一个共享指针头读一次选项,得到该深度的分布 $p^{(t)}=\mathrm{softmax}(z^{(t)}/T_t)$。请求打包成一条序列:状态 $s$ 在前,各问题连同选项以分隔符 token 包裹并以 $\langle\mathrm{decide}\rangle$ 结尾跟随;块因果掩码让每个问题只看到状态和自己的 token,位置索引在状态后重排,相当于各问题被独立作答而状态只编码一次。训练时对所有步的损失取平均 $\mathcal{L}=\frac{1}{R}\sum_{t=1}^{R}\ell(p^{(t)},y)$,指针头与 LoRA(rank 16)跨步共享;推理时先只跑 1 步,随后一个轻量逻辑回归停止器在每步后判断哪些问题还需加深,加深从上一步隐状态继续而不重读输入;每个深度各拟合一个温度 $T_k$ 做校准。请求的计算代价定义为“状态实际跑到的深度 / $R$”,按 token 数加权平均。

核心创新有三。其一:每步可读、逐步联合训练。以往循环要么后补(门控趋零、形同虚设),要么只有最后一步被监督(实验中只训最后一步的模型 depth-1 只有 36.4%,浅层读出不可用);LEV 让每个循环深度都经过决策训练,因此“一个 checkpoint 服务所有深度”,加深只是继续往前跑。其二:停止目标从“正确性”换成“与最深答案的一致性” $y_k=\mathbb{1}[\hat{o}^{(k)}=\hat{o}^{(R)}]$。正确性目标在标注校准集上调阈值后迁移性差(两个 Ouro 模型都落到固定深度前沿之下),而一致性目标无需标注、可在部署方自己的流量上拟合,阈值 $\tau$ 取满足“≥98% 最终答案与最深一致”的最小值(即最便宜的停止)。其三:加深以整个循环步为单位,不增加序列位置、不重读输入;同一请求内状态与问题一起加深,状态深度为 $d(s)=\max_j d(q_j)$。

方法步骤详情

流程五步。(1) 构造请求:状态 $s$ 置于序列开头,各问题的选项用分隔符 token 包裹并以 $\langle\mathrm{decide}\rangle$ 结尾,块因果掩码保证问题间互不可见。(2) 循环前向:$h^{(t)}=\mathrm{Norm}(F_\theta(h^{(t-1)}))$,每步用共享指针头 $z_k=(W_K h_k)^\top(W_Q h_{dec})/\sqrt{d_p}$ 打分,得各深度读出 $p^{(t)}$ 与答案 $\hat{o}^{(t)}$。(3) 训练:LoRA rank 16 加共享指针头,损失为步平均 $\mathcal{L}=\frac{1}{R}\sum_t\ell(p^{(t)},y)$,1 epoch、31,329 条记录、峰值学习率 $10^{-4}$、至多 4 张 H200。(4) 校准:在校准分区为每个深度拟合温度 $T_k$,不接触任何基准集。(5) 停止:每深度一个逻辑回归 $\hat{y}_k=\sigma(w_k^\top\phi_k+b_k)$,特征 $\phi_k$ 来自前后两步读出:最高概率、选项间边际、归一化熵、与上步答案是否一致、$\mathrm{KL}(p^{(k-1)}\|p^{(k)})$、概率变化量、$\log K$、$k$;阈值在无标注请求上于 $[0.5,1]$ 搜索,取满足“≥98% 与最深答案一致”的最小值;停止深度 $d_\tau(q)=\min\{k<R:\hat{y}_k\ge\tau\}\cup\{R\}$,状态跑到 $d(s)=\max_j d(q_j)$。

技术新颖性

与已有工作的本质区别:相对早退(CALM、BERxiT 等),LEV 的浅层读出是训练过的决策头而非未监督的中间层,且停在整步、不增加位置,阈值无标注可得;相对级联,不需要第二个模型、不重读输入;相对生成式思考(djev/OpenJev),加深成本是整次前向的 $k/R$ 而非逐 token 解码,简单题零额外开销;相对把循环后补到密集模型(作者实测收益不足 0.5 分、门控均值约 0.006),LEV 依赖预训练即循环的骨干(Ouro 以 7.7T token 预训练、每步带语言建模损失),使循环“开箱即用”。此外,把“何时停”形式化为 CATs 式一致性预测、并系统证明无标注阈值拟合足矣(每折 500 条无标注题即达全程拟合 0.2 分以内),以及“冻结配方后再读 held-out 面板”的实验纪律,都是决策模型文献中少见的严谨做法。

LEV with R loop steps, drawn for R = 4. (a) A request is packed into one sequence, and the same decoder stack runs R times. One shared pointer head reads the options after every step. (b) After step k, a stopper predicts whether the answer already agrees with the deepest one. (c) Each question exits at its own depth, and the state runs as deep as its deepest question.
Figure 3: LEV with R loop steps, drawn for R = 4. (a) A request is packed into one sequence, and the same decoder stack runs R times. One shared pointer head reads the options after every step. (b) After step k, a stopper predicts whether the answer already agrees with the deepest one. (c) Each question exits at its own depth, and the state runs as deep as its deepest question.

实验结果

四条结论。(1) 加深提精度、边际递减:LEV-2.6B 在 JevBench 从 63.2%(深度 1)到 76.2%(深度 4,+13.0 分,McNemar $p<0.001$;hard 档 +17.1),LEV-1.4B 58.0→68.0%,两步的 LEV-3B 69.3→72.7%;第一个额外步贡献最大(held-out 均值 51.3→62.0,深度 4 也只 65.7),且离训练数据越远收益越大。(2) 一个 checkpoint 多工作点:LEV-3B 深度 1 在 23ms 比 22ms 的 Qwen3-1.7B 高 4.0 分,LEV-2.6B 深度 2 在 101ms 比 105ms 的 Qwen3.5-2B 高 3.1 分,深度 4 与 Kev-4B 持平(76.2 vs 75.8)。(3) 按需深度省 15%–24% 计算、精度损失不足 1 分(锁定测试 vs 最深 −0.7~−1.0、vs 固定深度前沿 +0.1~+0.9),停止器自动把步数花在难题上:JevBench 三档平均 1.2/2.4/3.0 步。(4) 低延迟占优但优势在延迟而非算力:<50ms 时 LEV-1.4B/3B 比同延迟 dense 与级联高 5.4~8.8 分,dense 选项在 22–79ms 存在空档,按 FLOPs 计 LEV 不占优;闭源 Jev 仍以 86.6% 领先。改判分析显示深度主要在修正而非乱改(hard 档 29 题改对、10 题改错),但 LevBench 上 12.6% 的题浅层对而最深错,是一致性停止器的固有盲区。

Accuracy (%) and median latency per request of every system on the six panels of Table 5. Avg. is the unweighted mean of Kev-Breadth and LevBench.
Table 1: Accuracy (%) and median latency per request of every system on the six panels of Table 5. Avg. is the unweighted mean of Kev-Breadth and LevBench.
On-demand depth on the two held-out panels, pooled over their questions. Cost is the token-weighted share of the deepest pass.
Table 2: On-demand depth on the two held-out panels, pooled over their questions. Cost is the token-weighted share of the deepest pass.
Retrofitted loops on Qwen3.5-0.8B-Base. The first row is the accuracy (%) of the unlooped baseline, and every other cell is the paired accuracy difference to it in points, with a 95% CI, on development partitions.
Table 3: Retrofitted loops on Qwen3.5-0.8B-Base. The first row is the accuracy (%) of the unlooped baseline, and every other cell is the paired accuracy difference to it in points, with a 95% CI, on development partitions.
Training settings of LEV and the dense controls.
Table 4: Training settings of LEV and the dense controls.
Evaluation panels and the suites they come from. Suite names are those of the Kev repository.
Table 5: Evaluation panels and the suites they come from. Suite names are those of the Kev repository.
Our in-process evaluator against the official JevBench harness on the released Kev checkpoints.
Table 7: Our in-process evaluator against the official JevBench harness on the released Kev checkpoints.
Accuracy and latency on the public JevBench items per tier (48 easy, 72 standard, 111 hard), with the JevBench Intelligence score.
Table 8: Accuracy and latency on the public JevBench items per tier (48 easy, 72 standard, 111 hard), with the JevBench Intelligence score.
Top-label expected calibration error on the 111 hard JevBench items, with the LEV readouts and the control at their fitted temperatures.
Table 9: Top-label expected calibration error on the 111 hard JevBench items, with the LEV readouts and the control at their fitted temperatures.
Stopping rules and stopper variants on the pooled held-out panels, with the accuracy difference in points to the fixed-depth frontier at the same cost, and its 95% CI.
Table 16: Stopping rules and stopper variants on the pooled held-out panels, with the accuracy difference in points to the fixed-depth frontier at the same cost, and its 95% CI.
Accuracy of on-demand LEV minus a cascade or the best dense option at the same mean latency, in points with 95% CIs.
Table 17: Accuracy of on-demand LEV minus a cascade or the best dense option at the same mean latency, in points with 95% CIs.
Training data and full-weight fine-tuning, on the pooled held-out panels. Cost and the difference to the frontier are for on-demand depth.
Table 19: Training data and full-weight fine-tuning, on the pooled held-out panels. Cost and the difference to the frontier are for on-demand depth.
Accuracy of the jointly trained LEV-2.6B and of a model trained for each depth alone (Single), with the paired difference in points.
Table 21: Accuracy of the jointly trained LEV-2.6B and of a model trained for each depth alone (Single), with the paired difference in points.
(a) Loop steps that on-demand LEV-2.6B spends per JevBench tier. (b) LevBench test accuracy against latency, with the Pareto frontiers of LEV over all depths (solid) and of dense models and cascades (dashed). (c) JevBench accuracy of LEV-2.6B on the standard and hard tiers at depth one, on demand, and at depth four, against the dense Qwen3.5-4B control.
Figure 1: (a) Loop steps that on-demand LEV-2.6B spends per JevBench tier. (b) LevBench test accuracy against latency, with the Pareto frontiers of LEV over all depths (solid) and of dense models and cascades (dashed). (c) JevBench accuracy of LEV-2.6B on the standard and hard tiers at depth one, on demand, and at depth four, against the dense Qwen3.5-4B control.
Accuracy and median latency on the 231 public JevBench items. Each LEV checkpoint gives one operating point per depth, labelled by depth. Squares are separate dense checkpoints.
Figure 4: Accuracy and median latency on the 231 public JevBench items. Each LEV checkpoint gives one operating point per depth, labelled by depth. Squares are separate dense checkpoints.
Where depth helps. (a) JevBench accuracy on the standard and hard tiers. (b) LevBench questions that LEV-2.6B revises between consecutive depths. (c) Gain of LEV-2.6B from depth one to four per Kev-Breadth area. (d) LEV-2.6B on the LevBench tiers, with the cost of on-demand depth in parentheses.
Figure 5: Where depth helps. (a) JevBench accuracy on the standard and hard tiers. (b) LevBench questions that LEV-2.6B revises between consecutive depths. (c) Gain of LEV-2.6B from depth one to four per Kev-Breadth area. (d) LEV-2.6B on the LevBench tiers, with the cost of on-demand depth in parentheses.
Accuracy against cost. Gray points and line are the fixed depths and their frontier, and the blue line moves the threshold of the stopper to the chosen dot. The dashed line and the star repeat both on the locked test partitions.
Figure 6: Accuracy against cost. Gray points and line are the fixed depths and their frontier, and the blue line moves the threshold of the stopper to the chosen dot. The dashed line and the star repeat both on the locked test partitions.
LEV-2.6B on LevBench by difficulty k, such as the number of reasoning steps, on the generated reasoning families and the public datasets that define one, with the number of questions n and the cost of on-demand depth below each k.
Figure 7: LEV-2.6B on LevBench by difficulty k, such as the number of reasoning steps, on the generated reasoning families and the public datasets that define one, with the number of questions n and the cost of on-demand depth below each k.
Accuracy of each stopping rule minus the fixed-depth frontier at the same cost on the pooled held-out panels, with 95% CIs. Gray rows are early-exit rules.
Figure 8: Accuracy of each stopping rule minus the fixed-depth frontier at the same cost on the pooled held-out panels, with 95% CIs. Gray rows are early-exit rules.
The stopper of LEV-2.6B with few unlabelled requests. (a, b) Fitted with n unlabelled questions per fold. (c) Fitted on the unlabelled requests of one panel and applied to another, as the difference to the frontier.
Figure 9: The stopper of LEV-2.6B with few unlabelled requests. (a, b) Fitted with n unlabelled questions per fold. (c) Fitted on the unlabelled requests of one panel and applied to another, as the difference to the frontier.
Accuracy against mean latency on the LevBench test partition. Lines join the fixed depths of one LEV checkpoint, and stars mark its on-demand depth. The dashed line is the best dense model or cascade, and each number is the gap of on-demand LEV to it.
Figure 10: Accuracy against mean latency on the LevBench test partition. Lines join the fixed depths of one LEV checkpoint, and stars mark its on-demand depth. The dashed line is the best dense model or cascade, and each number is the gap of on-demand LEV to it.
Training choices. (a) LEV-2.6B on LevBench, trained jointly, at depth four only, or for depth one or two alone. (b) Pooled held-out accuracy by training set.
Figure 11: Training choices. (a) LEV-2.6B on LevBench, trained jointly, at depth four only, or for depth one or two alone. (b) Pooled held-out accuracy by training set.
Paired item-level differences in accuracy on the public JevBench items (231 in all, 111 hard), in points with 95% CIs. Open markers are the hard items.
Figure 13: Paired item-level differences in accuracy on the public JevBench items (231 in all, 111 hard), in points with 95% CIs. Open markers are the hard items.
查看结构化数据
任务指标本文基线提升
JevBench 多选决策(231 道公开题:48 easy / 72 standard / 111 hard) 总体准确率 LEV-2.6B:63.2%(深度 1)→ 76.2%(深度 4) Kev-4B 75.8%;同数据同配方的 dense 控制 Qwen3.5-4B 73.6% 较自身深度 1 +13.0 分(hard 档 +17.1);与 Kev-4B 持平(+0.4),超 Qwen3.5-4B +2.6 分
LevBench(自建冻结 held-out 基准,dev 4,370 题,id/near/deep/far 四档) 准确率 LEV-2.6B:46.9%(深度 1)→ 63.6%(深度 4) Qwen3.5-2B 53.1%;Qwen3.5-4B 62.1%(同配方 dense 控制) 深度 4 超 Qwen3.5-4B +1.5 分;首个额外步即修对 20.4% 的题
按需深度(两个 held-out 面板 pooled,锁定测试) 准确率 / 计算成本(相对最深前向) 成本 0.78–0.87,精度 59.3–63.7(三模型) 最深读出(成本 1.0)60.1–64.6 省 13%–22% 计算,仅损 0.7–1.0 分;对同成本固定深度前沿 +0.1~+0.9 分
低延迟区间(LevBench test,按需平均延迟 39–46ms) 对同延迟最佳 dense/级联的准确率差 LEV-1.4B / LEV-3B 按需:+5.4 ~ +8.8 分 Qwen3-1.7B→Qwen3.5-4B 级联及所有 dense 模型的最佳选项 LEV-3B 对最佳 dense +7.5 [6.5, 8.4],对级联 +8.8 [7.7, 9.9]
Kev-Breadth(14 个从未训练的数据集、5 个应用领域,test 3,031 题) 准确率 LEV-2.6B 67.7%(深度 4) 同配方 Qwen3.5-4B 67.4%;Qwen3.5-2B 64.7% +0.3 分;分领域看语言理解 +19.0、知识 +11.8、工具 +11.6,品味类任务约 0

局限与改进

作者承认:骨干仅 1.4B–3B(公开可得者止于此规模),深度曲线能否随规模延伸未知;JevBench 仅 231 道公开题;按需深度按循环步计费,尚无跨深度批处理的服务路径把节省变成墙钟时间;延迟在单请求、无融合核/图捕获/状态缓存下测得,各系统同等欠优化。我的补充:(1) 计算量并不减少——每步都跑完整骨干,FLOPs 口径下 LEV 落后 dense 最多 6.0 分,低延迟优势部分来自 22–79ms 的产品空档;(2) 一致性目标无法挽回“浅层对、深层错”的 12.6% 题,深度 4 还会把 10 道 hard 题改错;(3) 收益集中在前一个额外步,第四步至多 1 分;(4) 温度缩放后 hard 档 ECE 仍 0.18–0.26,远差于 Kev-4B 的 0.112 与 Jev 的 0.086;(5) 与闭源 Jev 绝对差距大(76.2 vs 86.6);(6) 加深让模型对不可知题更自信(top prob ≥0.9 从 0.0% 升到 6.3%),与弃答能力冲突。

独立分析的弱点

(1) 一致性目标的上限是最深读出,浅层对而深层错的 12.6% 题必被牺牲;可加入“加深是否改变答案”的双向预测,对会变的题比较浅/深置信度或做双读集成,并对少量流量始终跑满深度做对冲。(2) 加深非单调改进(hard 档 29 改对/10 改错),只取最深答案浪费中间读出;可用多深度读出加权集成或投票。(3) FLOPs 无优势、延迟优势依赖 dense 空档,一旦 dense 厂商补齐 20–80ms 档位即可能消失;需真正的步间增量计算,如后继步只跑部分层或复用 KV 缓存。(4) 校准差(hard ECE 0.18–0.26)且浅层在新分布过度自信(被停题在校准集 94.8% 正确 vs Kev-Transfer 78.9%),可加跨分布温度再校准或共形预测。(5) 状态须与最深问题同深($d(s)=\max_j d(q_j)$),一题难拖累全请求;可研究问题级状态快照或分组调度。(6) 品味类任务深度零增益,应对其自动退化为深度 1 以省算力。

未来方向

作者明确提出两个方向:设计能挽回“浅层对、深层错”的停止目标,使按需精度超越最深读出;把配方推广到更大的循环骨干,检验难题上的收益是否随规模增长。基于本文成果还可延伸:(1) 服务系统工程——对不同深度请求做连续批处理、融合核、图捕获与隐状态缓存,把步数节省兑现为墙钟时延,这是把论文主张产品化的关键缺口;(2) 把按需深度与“不可知题”检测联合学习“何时加深、何时弃答”(图 14 显示加深反而让模型对不可知题更自信);(3) 停止器的无标注拟合天然适配在线部署,可做持续更新与分布漂移监控(作者已证明跨面板迁移损失 ≤0.7 分,但训练源分布拟合会失效);(4) 训练侧,Train-120k 把 LEV-2.6B 深度 1 从 50.5% 提到 62.9%、按需成本降到 0.63,说明浅层读出可被数据显著强化,值得系统研究数据配比对深度曲线形状的影响;(5) 与生成式推理混合:对停止器判“加深后仍不一致”的少数难题升级为 token 级思考,构造真正的 System 1.5 流水线,兼取两种范式的长处。

复现评估

复现条件较好。代码开源于 GitHub(OpenLAIR/lev),三个模型权重发布在 HuggingFace(songdj/lev 集合),另有项目主页。训练成本不高:LoRA rank 16、仅 1 个 epoch、31,329 条记录,峰值学习率 $10^{-4}$,至多 4 张 H200;骨干 Ouro-1.4B/2.6B 与 Nanbeige4.2-3B 均公开。评测协议严谨透明:LevBench 在任何模型读它之前冻结,构建与查重规则写在附录 B.4;自建评测器在官方 Kev checkpoint 上与 JevBench harness 对齐(Kev-4B 完全一致,Kev-0.8B hard 档差不足 1 分);停止器拟合只需每折约 500 条无标注请求。剩余难点:闭源 Jev 无法复现只能引用;延迟数字依赖单张 H200 bf16、无优化 kernel 的特定环境,难以严格对齐;复现锁定测试需遵守其面板划分与“读一次”纪律。总体属中等偏易的复现难度,开放程度在同方向工作中属上乘。

相似论文推荐