MMLAv4
用一次尝试的反馈，调整后续尝试。
Φ 改变生成方式，M 保存可供读取的记录。
Feedback from one attempt can inform the next.
Φ adapts generation. M retains records for later reasoning.
同一道题仍在进行时，反馈可调整 Φ，参与下一次生成。While a problem is still open, feedback can adjust Φ for the next generation.
片段关闭后逐事件处理，完整记忆行提交后供后续读取。After segment closure, events are processed in order; committed rows become available to later reads.
01 / 运行机制Mechanisms
Φ 和 M 都能影响下一次尝试，但保存的内容、更新的时机和写入方式各不相同。Both Φ and M can affect the next attempt. They store different things and have different update rules and timing.
POLICY STATE
AUTHORITATIVE MEMORY
更新 Φ，或不更新Update Φ, or leave it unchanged后续尝试读取当前的 ΦLater attempts read the resulting Φ逐事件生成候选，写入或 NULLFor each event: propose, then write or take NULL全部处理完，后续推理读取 MAfter all events are processed, later reasoning reads M“尝试”是一次候选生成；“片段”是系统划定的一段记录，结束后才能回看整合。两条路径可以分别发生，无需每次同时更新。An attempt generates a candidate. A segment is a delimited record that becomes available for consolidation after it closes. Either update path can run independently.
贯穿示例 · 非实验结果Worked example · Not an experimental result
题目先给出 x + y = 10。第一次尝试提交 (4, 6) 后，校验器才返回一条额外约束：x > y。The problem initially gives x + y = 10. After the first attempt submits (4, 6), the checker returns an additional constraint: x > y.
一次尝试开始时At the start of an attempt
模型把题目、当前 Φ 和 M 中可读取的信息用于生成。一次尝试固定读取这组版本，途中不会悄悄换成刚更新的状态。The model generates using the problem, current Φ, and information read from M. The attempt uses a fixed pair of versions throughout.
此时只知道 x + y = 10，可以提出 (4, 6)。反馈 x > y 尚未到达，不能提前使用。Only x + y = 10 is known, so (4, 6) is a possible candidate. The feedback x > y has not arrived yet.
同一问题仍在求解中While the same problem is being solved
更新规则把已收到的反馈转为对 Φ 的调整，例如更新小型适配器的参数。更新后的数值参与下一次计算；更新可以被拒绝，此时 Φ 保持原样。An update rule uses received feedback to adjust Φ, for example the parameters of a small adapter. Those values enter later computation. An update may also be rejected, leaving Φ unchanged.
反馈 x > y 现在可以用于更新。Φ 保存的是调整后的数值，不是这句约束的文字；更新也不保证下一次答对。The feedback x > y can now inform an update. Φ holds adjusted numbers, not the text of the constraint. Correctness on the next attempt is not guaranteed.
片段完成之后After a segment ends
片段结束后，系统逐事件处理。神经模型为各可行槽位提出内容；可信模块补齐身份、版本、来源和权限等字段，再校验完整行。Once a segment closes, events are processed in order. A neural model proposes content for each feasible slot. A trusted module supplies identity, version, provenance, and permissions, then validates each complete row.
可以把 x > y 提议为这道题的约束记录。候选包括可检查的内容及模型读取用的表示，此时还未写入 M。x > y can be proposed as a constraint for this problem. The candidate has inspectable content and a representation for model reads. It has not yet entered M.
每个事件的写入决策The decision for each event
准入控制器比较各槽位写入与 NULL 的预期后果，每个事件最多替换一整行。后一个事件接着读取处理后的 M；全部事件处理完，新版本才供后续推理使用。Admission compares the expected consequences of each slot write and NULL. Each event replaces at most one complete row. The next event uses that M; later reasoning can read it after all events are processed.
假定准入选择写入槽位 02，x > y 就成为可读取的记忆。若选择 NULL，候选被放弃，M 的内容和版本都不变。Suppose admission selects slot 02. Then x > y becomes readable memory. With NULL, the candidate is discarded and M keeps its content and version.
x + y = 10(4, 6)x > y 尚未返回，不能作为本次输入。x > y has not been returned and cannot be an input yet.
x > y
反馈 → 更新规则 → 新数值状态FEEDBACK → UPDATE RULE → NEW VALUES
Φi+1下次尝试读取新的策略版本The next attempt can read the new policy version
片段已完成Segment closedconstraint / x_gt_yx > y提出内容 → 组装并校验 → 等待准入Propose content → Assemble and validate → Await admission
x > y连同来源、版本等字段With provenance, version, and other fields原记忆行Original memory row↳=仅替换槽位 02 · 其余行保持原样Only slot 02 is replaced. All other rows stay unchanged.
NULL 不删除记忆，也不改变记忆版本。NULL neither deletes memory nor changes its version.
预测式记忆准入Predictive memory admission
“准入”就是决定一条候选能否进入 M，以及替换哪个槽位。容量固定，保存新信息可能挤掉仍有用的旧信息；有价值的候选也未必值得写入。Admission decides whether a candidate enters M and which slot it replaces. With fixed capacity, new information can displace useful old information, so even a useful candidate may not be worth writing.
在例子中，保存 x > y 可能方便以后读取；但若它已被可靠保留，或覆盖会丢掉更重要的约束，NULL 可能更合适。Retaining x > y may help later reads. If it is already reliably retained, or a write would remove a more important constraint, NULL may be preferable.
比较的是多种可能后续的平均表现，不能看过某个实际未来，再为它挑选动作。The target averages over possible continuations. An action cannot be chosen after seeing which future actually occurred.
估计各动作风险槽位 / NULLSlot / NULL“预测式”指用已有信息估计未来后果。部署时没有未来分支、标准答案或事后评估结果作为输入。“Predictive” means estimating future consequences from available information. Future branches, answer keys, and later evaluations are absent from deployment inputs.
机制与训练目标依据 v4 第 2–3 节。各模块的定义、实现条件和评估方法见下方论文详解。Mechanisms and training objectives follow Sections 2–3 of v4. The paper guide develops each module's definitions, implementation conditions, and evaluation methods.
论文机制The papers
完整机制详解Full mechanism guide五类状态、检索回退、关系绑定、接口与潜在读出。Five states, retrieval fallback, relational binding, interfaces, and latent readout.
02 / RTT怎样算同一道题内的学习What counts as learning within a problem反馈、策略更新、后续复用，以及因果验证和重置。Feedback, policy updates, reuse, causal tests, and reset.
03 / ATOMIC MEMORY ROWS一条记忆怎样获得写入资格How a memory row becomes authoritative行结构、可信组装、语义一致性、生命周期与恢复。Row schema, trusted assembly, semantic consistency, lifecycle, and recovery.
04 / PREDICTIVE ADMISSION怎样预测一次写入的后果How to predict the effects of a write分组未来、动作重放、多事件决策、不确定性与 NULL。Grouped futures, action replay, sequential decisions, uncertainty, and NULL.
05 / DUAL STATE策略和记忆怎样独立协作How policy and memory work together版本对、独立操作、交换与回滚、四格对照和成本。Version pairs, separate operations, swaps, rollback, factorial comparisons, and cost.
06 / SEGMENT CONSOLIDATION何时可以回看，何时允许读取When to review, and when to expose memory五种时钟、片段边界、重叠去重、异步提交与信息隔离。Five clocks, segment boundaries, overlap, deduplication, asynchronous commits, and isolation.
02 / 系统架构Architecture
策略和记忆各自维护版本、写入权限与回滚记录。一次推理读取兼容的版本组合，更新后的状态用于之后的推理。Policy and memory have separate versions, write permissions, and rollback records. Each attempt reads a compatible pair; updates become available to later attempts.
当前生成过程的工作状态，随 token 和尝试变化。它与跨片段保留的记忆 M 有不同的生命周期。The working state of generation changes across tokens and attempts. Its lifetime differs from memory M, which persists across segments.
汇总片段内已经发生的信息，容量有界。片段结束后，系统据此构造记忆候选。A bounded summary of information observed within a segment. After the segment ends, it supports the construction of memory candidates.
在预训练或离线训练中更新，与推理时可变的 Φ、M 分开管理。Updated during pretraining or offline training, and managed separately from Φ and M, which can change during reasoning.
03 / 实验Experiments
选自 v4 第 5–8 节的记忆管理、类型化传输与扩展上下文问答实验。Selected memory management, typed transport, and extended-context QA experiments from Sections 5–8 of v4.
三个随机种子下，各 300 条留出记录均正确执行了生命周期操作。任务使用合成数据。All 300 held-out records passed the lifecycle checks for each of three seeds in synthetic tasks.
表 7 · 第 20 页Table 7 · p. 20受控 anchor-filler 任务中，三个随机种子下各 240 条留出记录均完整匹配。All 240 held-out records matched exactly for each of three seeds in the controlled anchor-filler task.
表 9 · 第 25 页Table 9 · p. 25Qwen2.5-14B 在扩展上下文实验中的结果范围，覆盖三个选择器种子。The range across three selector seeds for Qwen2.5-14B in the extended-context experiment.
表 8 · 第 22 页Table 8 · p. 22多跳问答 · 论文表 8Multi-hop QA · Table 8
Qwen2.5-14B · 扩展上下文（约 8,200 词）
| 方法Method | F1 | 平均输入 tokensAvg. prompt tokens |
|---|---|---|
| 完整上下文Full context | 0.715 | 12,646 |
| 稠密检索Dense retrieval | 0.753 | 1,338 |
| BM25 | 0.771 | 1,330 |
| 驻留缓存 + 归档回退Resident cache + archive fallback | 0.813–0.830 | 1,368 |
展示表 8 的扩展上下文设置，含 300 道留出问题。F1 区间覆盖三个选择器种子；输入长度为完整提示词的平均 token 数。此处比较问答检索组件，归档、索引、检索延迟和能耗另行计量；完整设置见原表。The extended-context setting from Table 8, with 300 held-out questions. F1 ranges cover three selector seeds; token counts are averages for the complete prompt. This comparison covers the QA retrieval component. Archive storage, indexes, retrieval latency, and energy are separate costs; see the source table for full settings.
04 / 论文脉络The research program
五篇理论论文分别处理学习发生的时机、记忆的保存形式、写入的选择，以及两种状态如何协作。v4 把这些规则放进同一套架构。The five theory papers address when learning happens, how memory is represented, which writes to admit, and how the two states work together. v4 brings their rules into one architecture.
05 / 论文Publications
v4 汇总整体架构、理论分析和实验结果。五篇配套论文分别讨论其中的具体问题。v4 brings together the architecture, theoretical analysis, and experimental results. Five companion papers develop individual parts of the theory.
222 页。介绍策略状态与记忆状态的更新规则，分析推理时训练、记忆准入和片段整合，并汇总组件实验与后训练研究的结果。222 pages covering update rules for policy and memory states, reasoning-time training, memory admission, and segment consolidation, with results from component experiments and post-training studies.
五篇 R02 理论稿分别展开推理时训练、记忆表示、准入和整合。修订后的定义与结论以 v4 为准。Five R02 theory manuscripts develop reasoning-time training, memory representation, admission, and consolidation. Revised definitions and conclusions in v4 take precedence.
Reasoning-Time Training: Learning Before a Single Problem Ends
Atomic Memory Rows: A Bounded, Verifiable Substrate for Editable Reasoning
Learning What to Remember: Predictive Admission for Bounded Reasoning-Time Memory
MMLA-RTT: Dual-State Learning at Reasoning Time
Causal Generation, Retrospective Consolidation: Completed-Segment Bidirectional Memory Without Temporal Leakage
06 / 引用Citation
PNG
SVG
@article{zou2026mmla,
title = {{MMLA}: Memory-Mediated Learning Architecture for Predictive Dual-State Adaptation},
author = {Zou, Junyi and Donz, Avrova},
year = {2026},
eprint = {2606.28876},
archivePrefix = {arXiv},
primaryClass = {cs.cl},
note = {Version 4},
url = {https://arxiv.org/abs/2606.28876v4}
}