每日科技文摘Daily Tech Digest

本期:一个模型被反复做了 600 次实验之后This issue: what 600 experiments taught one person about a world model

发布于 2026-11-15Published 2026-11-15

今日深读Today's Deep Read

600 次实验之后:JEPA 世界模型的瓶颈不在「预测不准」,而在「没人教它分段想」After 600 Experiments: The Bottleneck in a JEPA World Model Wasn't Bad Predictions — Nobody Was Making It Think in Chunks

今年,Yann LeCun 的团队发布了一个叫 LeWM 的 JEPA 世界模型,基准任务很朴素:机械推一个 T 形块,让它从桌上任意位置移动到目标图像里标出的位置。一个刚毕业的工程师 Haddy Alchaer 没有训练任何模型,只是把这几个月里能想到的「用法」都试了一遍——前后 600 组实验——然后写了一篇异常诚实的复盘。他的第一个发现就把流行的解释打掉了:长程任务失败,通常被归咎于「预测误差随步数累积」,按这个诊断,办法应该是换个更准的模型。他把 LeWM 的预测器整个换成模拟器的真实物理(相当于给模型一台永远不出错的预言机),在 75 步长的任务上,成功率从 78% 变成 77%——几乎没动。真实物理只比学出来的预测好 20 个百分点(42% 对 22%,在「一次性规划全程」的设定下),而把任务切成三段 25 步、每段单独规划,成功率从 22% 跳到 80%。同样是改「怎么用」,收益是换「物理引擎」的三倍。This year Yann LeCun's team released LeWM, a JEPA world model evaluated on a humble task: a robot pusher slides a T-shaped block across a table until it lands where a goal image says it should. Haddy Alchaer didn't train anything. He just spent a few months trying every way he could think of to *use* the published model — 600 experiments — and wrote an unusually honest postmortem. His first finding killed the prevailing diagnosis: long-horizon failures are usually blamed on prediction errors compounding with length, which implies you need a better model. So he swapped LeWM's learned predictor for the simulator's exact physics — an oracle that never errs — and on 75-step tasks success went from 78% to 77%. Nothing. Exact physics added 20 points (42% vs 22%) when the planner tries to cover all 75 steps at once; cutting the task into three 25-step segments added 58. Changing how you plan paid out three times what changing the physics paid.

为什么会这样?模型在 25 步内预测得近乎完美(89.7%,和真物理只差 0.8 个百分点),只是它的「可靠视野」就这么多。与其修模型,不如顺着视野规划:高级别挑子目标,低级别只走 25 步。这就是分层规划,被公认为世界模型的主要开放问题之一,而在 Alchaer 的实验里它值 60 个百分点。他还顺手做了一个反直觉的检验:多花四倍的搜索预算(跑四次规划取最优)只换来不到 1 个百分点——因为四次运行几乎同生共死,算力替代不了结构。Why? Because the model predicts near-perfectly within 25 steps (89.7%, within a point of exact physics) and that's its reliable horizon. Rather than fixing the model, plan around its sightline: a high level picks subgoals, a low level only ever walks 25 steps. That's hierarchical planning — widely called one of the main open problems for world models — and here it was worth sixty points. A bonus control was counterintuitive: running the planner four times and keeping the best plan cost 4x compute and bought less than one point, because the runs succeed and fail together. Compute can't substitute for structure.

第二个发现更有意思:怎么在没有专家演示的情况下自动挑出这些子目标?他把 LeWM 对训练演示的每一帧嵌入(约 46 万个状态)连成一张图,演示内相邻帧连边,不同演示里嵌入接近的状态也连边。关键性质是:嵌入距离恰好对应物理可达性——「25 步内可达」的判断 87% 正确,也就是说这个向量空间自带一张「够得着」的地图。规划时在这张图上从当前状态搜到目标状态,沿途状态就是子目标。第一版有个陷阱:跨演示的跳转和真实的一步同等计价,搜索就拼出一条物理上根本走不通的捷径(1015 条计划里 318 条用了六次以上跳转);给每次跨演示跳转加个代价,成功率从 62% 到 85%,追平拿着专家子目标的天花板。The second finding is more original: how to generate those subgoals without expert demonstrations? Alchaer embedded every frame of the training demonstrations with LeWM's own encoder (~460,000 states) and wired them into a graph — consecutive frames within a demonstration, plus states close in embedding space across demonstrations. The key property: embedding distance tracks physical reachability (87% accurate at ranking a reachable 25-step goal as closer), so the vector space doubles as a map of what the robot can reach. Search that graph from current state to goal, and the states along the path become subgoals. The first version had a trap — free jumps between demonstrations let the search chain shortcuts no physical robot could follow (318 of 1,015 plans used six-plus jumps). Putting a cost on every cross-demonstration jump took success from 62% to 85%, matching the planner handed ideal expert subgoals.

剩下的实验在拆「测试本身」。标准 Push-T 测试的目标位置就摆在演示路径上——几乎所有演示都以同一个终点收场,84% 的测试目标在记忆图里已有一模一样的状态。等于考卷答案写在草稿纸上。Alchaer 于是自建了一个更难的测试:起点取一条演示、目标取另一条演示,逼机器走「没有任何一条演示连接过的」状态,还把成功定义收紧(最后 10 步都必须停在目标上)。在这种测试下,纯记忆图只有 45%,而第三个技巧——利用推动物块的物理对平移、旋转、镜像不变——把别处的演示片段变换一下填进记忆的空白,成功率回到 85%,还能让 1% 的演示记忆顶到 10% 用。The remaining experiments went after the benchmark itself. The standard Push-T test picks goals that sit on the demonstration paths — every demonstration ends near the same place, so 84% of test goals already existed in the memory graph. The exam answers were printed on the scratch paper. So Alchaer built a harder test: start in one demonstration, goal in another, forcing travel through states no single demonstration connects, with success tightened to 'the block holds on target for the last 10 steps.' Plain memory drops to 45% there. A third trick — pushing physics is invariant under shifting, rotating, and mirroring the whole table — lets you paste transformed demonstration segments into holes in memory, restoring 85% and stretching 1% of the demonstrations to perform like 10%.

这篇复盘最值得带走的不是某个数字,而是方法论本身:当所有人都在造更大的预测器时,先做一个便宜的判决实验(把预测器换成真物理),问清楚瓶颈到底在模型的哪一块。他测出这个模型的有用视野是 25 步,然后把所有设计都围绕「不要越出视野」来做。诚实的部分同样值得抄:小样本能骗人(官方 50 个 episode 报 68%,同分布抽 250 个只剩 44%),换个随机种子就能翻转 38 个 episode 的成败;基准的出题方式本身就该被审计。What to take away isn't any single number but the method: while everyone else builds bigger predictors, run one cheap decisive experiment (swap the predictor for perfect physics) to find out which part is actually the bottleneck. Once he measured the model's useful horizon as 25 steps, every design decision bent around staying inside that horizon. The honesty is worth copying too: the official 50-episode sample flatters the model (68%; on 250 fresh same-distribution episodes it's 44%), and one planner seed flip changes 38 of 250 outcomes. Benchmarks deserve audits of how their questions are drawn, not just of their scores.

AI 思考AI Thoughts

数学之美Mathematical Beauty

程序员随想Programmer Reflections

哲学与生活Philosophy and Life

工具角Tools Corner