ARA canonical source:
ara/PAPER.md
SameTime 中文镜像源:ara/PAPER.zh.md
ARA 论文式总览:Semantic Prefix Routing 与 TreeHeap
状态:持续更新的研究工件
创建时间:2026-06-22
负责人:nio / Houming818
Review Engineer:Codex
仓库:SameTime
ARA 参考:https://arxiv.org/abs/2604.24658v3
这个文件是 SameTime 公开研究记录的根 ARA manifest。
它不是传统论文。它是一份紧凑的研究状态图,目的是让人类或 AI reviewer 能够重建下面这条链:
claim -> predict -> experiment -> evidence -> decision -> next step
主张 -> 预测 -> 实验 -> 证据 -> 判断 -> 下一步
面向人的叙事版本发布在 SPR blog 中。这个文件是给 reviewer 和后续 AI agent 使用的地图。
最新的细节以各层 logic/claims.md、logic/experiments.md 和 evidence/ 为准;本文负责给出全局结构。
ARA 目录结构
SameTime 使用四层 ARA 结构:
| 层 | 路径 | 用途 |
|---|---|---|
| Logic | ara/*/logic/ |
claims、predicts、problem、experiment design |
| Source | ara/*/src/ |
可复现实验脚本和环境说明 |
| Trace | ara/*/trace/ |
路线转向、被拒绝路径、决策 DAG |
| Evidence | ara/*/evidence/ |
summary、metrics、raw log 指针 |
主要课题:
| Topic | Path | Role |
|---|---|---|
| S1 Echo | ara/s1-echo/ |
容量、路径 hash、上下文路由 |
| S2 Translation | ara/s2-translation/ |
Fold stack、graph builder、probability container |
| M0 TreeHeap Math | ara/m0-treeheap-math/ |
代数、plus、kernel search、可训练性 |
| S3 Generation | ara/s3-generation/ |
后续 generation / decoder work 的占位 |
状态词汇
| Status | 含义 |
|---|---|
verified |
已有证据,并包含 baseline 或 falsification check |
supported |
有正向 pilot evidence,但 baseline 或 scale 不完整 |
open |
主张合理,但尚未测试 |
design |
架构或数学设计,还没有证据支撑 |
rejected |
已测试并失败 |
downgraded |
早期较强主张被后续证据收窄 |
全局研究主张
当前全局 claim 必须刻意收窄:
SPR / TreeHeap 不能靠断言替代 Transformer。
它是一族 addressable、path-aware、tree-structured operators。
这些算子可能拥有 MLP / CNN / Transformer 类似的机器学习能力,
同时在 address、substructure migration、prefix reuse、
probability container、delayed collapse 等问题上提供更强归纳偏置。
在提出 WMT 规模主张之前,研究必须先通过小任务证明:
M0 math toolbox
-> S1 capacity and context routing
-> TreeHeap trainability
-> structural existence proofs
-> S2 translation / fold-stack tasks
Claim Tree
C0:ARA 流程主张
| ID | Source | Claim | Status | Evidence / Pointer | Falsification |
|---|---|---|---|---|---|
| C0-001 | SPR-001, SPR-008 | 每个强 claim 都必须有 evidence 和 falsification criteria。 | verified | ara/README.md, ara/*/logic/claims.md, ara/*/evidence/README.md |
一个 claim 在没有 evidence pointer 或 failure condition 时被升级。 |
| C0-002 | SPR-008 | 正确工作循环是 predict -> claim -> experiment -> evidence -> trace。 |
supported | Blog SPR-008 和当前 ARA layout | 研究决策继续只依赖聊天记忆。 |
| C0-003 | SPR-019 | Blog 叙事不够;claims 必须保存在 registry artifacts 中。 | open | PAPER.md / PAPER.zh.md 是第一层 root manifest |
后续 SPR claims 与 registry 分离且不 reconciliation。 |
C1:S1 容量与顺序主张
| ID | Source | Claim | Status | Evidence / Pointer | Falsification |
|---|---|---|---|---|---|
| C1-001 | SPR-001, SPR-002 | S1 token path hash 对 WMT14 word echo 有足够容量。 | verified | ara/s1-echo/logic/claims.md S1-C01, solo rate 99.7% |
同 slice / seed 下 solo rate 低于 95%。 |
| C1-002 | SPR-002 | 纯 cyclic shift 会发生 order collision;sign alternation 能打破这个对称。 | verified | ara/s1-echo/logic/claims.md S1-C02 |
非退化 A,B 与 B,A 在 sign alternation 后仍表示相同。 |
| C1-003 | SPR-002 | Echo reconstruction 可以在没有 learned Transformer attention 的情况下达到近似完美 BLEU。 | supported | ara/s1-echo/logic/claims.md S1-C03, BLEU-4 99.99 |
打乱 leaf labels 后 BLEU 仍保持,说明只证明 lookup。 |
| C1-004 | SPR-003 | Token-only routing 不编码上下文语义。 | rejected old claim / verified rejection | ara/s1-echo/logic/claims.md S1-C11 |
只有当 token-only route 在 polysemy 上超过 context 和 random baselines 才能重开。 |
| C1-005 | SPR-007 | Context-conditioned routing 能在 controlled proof 中编码语义区别。 | supported | ara/s1-echo/logic/claims.md S1-C10, S1-C13 |
真实 corpus 或 BoW/random baselines 追平或超过 context route。 |
| C1-006 | SPR-006, SPR-007 | S1b 的最小接口是 route(token, context),不是 route(token)。 |
supported | ara/s1-echo/logic/claims.md S1-C13 |
在 controlled 和 real tests 中 context 没有提升。 |
| C1-007 | SPR-004 | S1 output 可以作为 downstream S2 的输入契约。 | open | ara/s1-echo/logic/claims.md S1-C21 |
用 matched random vectors 替换 S1 vectors 后 S2 不变。 |
C2:S2 Fold Stack 与 Graph Builder 主张
| ID | Source | Claim | Status | Evidence / Pointer | Falsification |
|---|---|---|---|---|---|
| C2-001 | SPR-005 | Semantic vectors 包含 fold-action 信息。 | verified | ara/s2-translation/logic/claims.md C-001 |
clean split 下 fold action prediction 降到 chance。 |
| C2-002 | SPR-005 | 32D semantic space 足够完成当前 fold-action prediction task。 | verified | ara/s2-translation/logic/claims.md C-002 |
重新计算时 32D 明显弱于 128D。 |
| C2-003 | SPR-005 | EN->ZH 和 ZH->EN 两个方向都能做 cross-lingual fold structure prediction。 | verified | ara/s2-translation/logic/claims.md C-003, C-004 |
cross-lingual AUC 对 frequency / BoW baselines 降到 chance。 |
| C2-004 | SPR-005 | Fold action types 可以被 small MLPs 预测。 | supported | ara/s2-translation/logic/claims.md C-005 |
leakage control 后 baselines 追平。 |
| C2-005 | SPR-005, SPR-008 | PP/VP/NP pattern 显示不同 collapse regimes。 | verified | ara/s2-translation/logic/claims.md C-006..C-010 |
recomputed grammar atlas 消除这些 collapse pattern。 |
| C2-006 | SPR-005 | Head、span、action detection 能用 supervised probes 完成。 | verified | ara/s2-translation/logic/claims.md C-011..C-013 |
clean split 下 probe 失败。 |
| C2-007 | SPR-005 | Fold representation 可以用 oracle 或 template edges 重建 sentence。 | verified / supported | ara/s2-translation/logic/claims.md C-014, C-015 |
oracle-free setting 下 reconstruction collapse。 |
| C2-008 | SPR-008 | 当前 graph assembly 瓶颈是 child/parent allocation,不是 fold action representation。 | verified | ara/s2-translation/logic/claims.md C-016..C-020 |
oracle child/parent ablations 无法缩小 UAS gap。 |
| C2-009 | SPR-008 | Parent candidates 应作为 probability containers 保留,而不是 early argmax。 | verified for graph-builder stage | ara/s2-translation/logic/claims.md C-028 |
top-k parent coverage 低,或后续模块无法使用 distributions。 |
| C2-010 | SPR-008 | 当前 3-epoch TreeHeap vectors 过度坍缩,不能支撑 syntax-energy claim。 | verified | ara/s2-translation/logic/claims.md C-027, strategy audit |
重新训练后 vectors 在 baselines 下显示可分离 syntax energy。 |
| C2-011 | SPR-008, SPR-009, SPR-010 | 历史 checkpoints 不能承载更强的 “TreeHeap syntax energy solved” claim。 | downgraded | ara/s2-translation/evidence/strategy_audit/, tmerge_diagnostic/ |
新独立训练 checkpoint 通过 energy 和 syntax baselines。 |
| C2-012 | SPR-010 | 当前 world-model / background-field training evidence 只是诊断,不是正向 translation claim。 | downgraded | ara/s2-translation/evidence/world_model_long_20260617_180554/ |
新 loss 在 controls 下显示下游 translation 或 structure gains。 |
C3:M0 TreeHeap 代数主张
| ID | Source | Claim | Status | Evidence / Pointer | Falsification |
|---|---|---|---|---|---|
| C3-001 | SPR-011 | TreeHeap 必须先成为数学工具箱,再提出语言 claim。 | design | ara/m0-treeheap-math/logic/problem.md, logic/solution/algebra.md |
在没有 closure / inverse / projection tests 的情况下推进语言层工作。 |
| C3-002 | SPR-011, SPR-013 | Minimal TreeHeap algebra 应支持 closure、non-commutativity、inverse-like operations、projection、subheap matching、probability normalization。 | supported pilot | ara/m0-treeheap-math/logic/predicts.md P-MATH01; evidence/treeheap_math_probe/ |
synthetic exact mode 下 closure 或 order distinction 失败。 |
| C3-003 | SPR-012 | Subheap kernel search 是 TreeHeap 对 local structure 做 convolution 的类比。 | design / partly supported by toy | ara/m0-treeheap-math/logic/experiments.md existence suite B |
kernel relocation 无法迁移到 seen positions 之外,或退化成 root matching。 |
| C3-004 | SPR-013 | Pure-math probe 可以建立非语言数学 pilot。 | supported | evidence/treeheap_math_probe/README.md |
probe 无法复现 closure / inverse / subheap matching。 |
| C3-005 | SPR-014 | TreeHeap order 应由 primitive + plus 生成,而不是只由外部 index 指定。 | design | ara/m0-treeheap-math/logic/predicts.md P-MATH02 |
没有 plus candidate 能产生 successor、cycle 或 order margin。 |
| C3-006 | SPR-015 | Addressable TreeHeap 可以在 toy 中把 plus 用作 successor、information gain、mod-base fold、cyclic-kernel addressing。 | supported toy | evidence/primitive_plus_probe/ |
toy trace 无法复现 address successor 或 mod fold。 |
| C3-007 | SPR-016 | 在 TreeHeap encoder/decoder 之前,本地 learning stack 必须通过 linear、nonlinear、modular toy tasks。 | verified for local stack | evidence/trainability_quiz/ |
linear regression、XOR 或 modular addition deterministic rerun 失败。 |
| C3-008 | SPR-016 | 通过 trainability quiz 只能证明本地 ML stack capacity,不证明 TreeHeap language intelligence。 | verified boundary | ara/m0-treeheap-math/logic/experiments.md trainability interpretation |
blog 或 registry 把 toy tasks 升级为 language evidence。 |
C4:TreeHeap 存在性偏置主张
| ID | Source | Claim | Status | Evidence / Pointer | Falsification |
|---|---|---|---|---|---|
| C4-001 | SPR-017 | TreeHeap 的存在性主张不是 “它也能学”,而是 structural inductive-bias claim。 | design | Blog SPR-017 | TreeHeap 只匹配 generic MLP/Transformer,没有结构优势。 |
| C4-002 | SPR-017 | Addressable closure / mod-fold tasks 应测试 TreeHeap 是否在 length extrapolation 下保留 address 和 overwrite semantics。 | open | ara/m0-treeheap-math/logic/predicts.md P-EXIST01 |
flatten baselines 在 extrapolated address tasks 上追平或超过 TreeHeap。 |
| C4-003 | SPR-017 | Subheap kernel relocation 应显示 local pattern 跨 unseen tree addresses 迁移。 | open | ara/m0-treeheap-math/logic/predicts.md P-EXIST02 |
kernel 只在训练位置工作,或 false positive 很高。 |
| C4-004 | SPR-017 | Prefix compression 和 delayed collapse 应显示 shared prefixes reuse 与 calibrated candidates。 | open | ara/m0-treeheap-math/logic/predicts.md P-EXIST03 |
prefix tree 没有 compression gain 或 probability container calibration。 |
| C4-005 | SPR-018 | 单纯 pattern matching 是错误的 B 实验;真正 TreeHeap proof 需要 learned encoder 加 conjugate query/decoder kernel。 | design | Blog SPR-018 | hand-designed pattern matching 仍是唯一可工作机制。 |
| C4-006 | SPR-018 | Learned ordered TreeHeap search 应把 key/value data 编码进 searchable tree,并用 query kernel 解码。 | open | Blog SPR-018 planned experiment | learned encoder 无法形成超过 trivial memorization 的 searchable structure。 |
| C4-007 | SPR-018 | Learned weighted prefix TreeHeap 应类似 Huffman-like compression,适合 skewed symbol distributions。 | open | Blog SPR-018 planned experiment | expected path length 无法击败 fixed-length baseline。 |
C5:Soft TreeHeap / Gradient 主张
| ID | Source | Claim | Status | Evidence / Pointer | Falsification |
|---|---|---|---|---|---|
| C5-001 | SPR-019 | TreeHeap 需要 differentiable operator lifting,才能像 MLP / Transformer 一样训练。 | design | Blog SPR-019 | 只用 hard non-differentiable operators 就成功训练 useful TreeHeap。 |
| C5-002 | SPR-019 | Soft TreeHeap 应是 Hard TreeHeap operators 的 probability lifting:SoftO(H)=sum_a p(a) O_a(H)。 |
mathematical design | Blog SPR-019 | one-hot soft operation 无法恢复 hard operation。 |
| C5-003 | SPR-019 | Naive soft memory write 不足以作为 TreeHeap claim,因为它更新 array slots,而不是 plus algebra。 | design / critique | Blog SPR-019 | slot interpolation 被形式化证明等价于 TreeHeap plus。 |
| C5-004 | SPR-019 | Soft Plus 应 lift TreeHeap plus:`H_next=sum_a p(a | H,x) (H plus_a x)`。 | supported pilot | ara/m0-treeheap-math/evidence/soft_plus_probe/ |
| C5-005 | SPR-019, SPR-020 | Kernel-guided Soft Plus 应使用 TreeHeap convolution kernel 决定 write/merge route。 | open / scoped pilot evidence | ara/m0-treeheap-math/evidence/soft_plus_probe/;GLM audit 显示当前 collapse 依赖 engineered alignment features |
在 clean-feature ablation 中 kernel-guided plus 弱于 naive memory write 或 encoder soft plus。 |
| C5-006 | SPR-019 | Multi-kernel / staged training 优于一个 “big pot” loss。 | open | Blog SPR-019 Experiment 3 loss ablation | matched budget 下 single total loss 更稳定且泛化更好。 |
| C5-007 | SPR-019 | Soft collapse 应恢复 legal Hard TreeHeap structure。 | supported pilot | ara/m0-treeheap-math/evidence/soft_plus_probe/: collapse_accuracy_tau_0.05=1.0 |
更大或 noisy tests 中 collapse legality、route interpretability、hard-soft gap 失败。 |
| C5-008 | SPR-021 | C05 必须暴露 TreeHeap structure:path、subheap、recursive route/plus;否则退化成 flat soft memory。 | supported pilot | ara/m0-treeheap-math/evidence/structural_c05_probe/: flat/path-only test acc 0.0; subheap/path+subheap test acc 1.0 |
unseen-depth relocation 中 flat address 或 path-only baselines 追平 subheap/path-subheap kernels。 |
SPR Blog Source Map
| SPR | File | Main role in ARA |
|---|---|---|
| 001 | blogs/.../spr/001-problem.md |
问题定义与 ARA 规则:每个 claim 需要 evidence / falsification |
| 002 | blogs/.../spr/002-s1-evidence.md |
S1 echo capacity 与 order evidence |
| 003 | blogs/.../spr/003-s1-falsification.md |
token-only semantic route 被拒绝 |
| 004 | blogs/.../spr/004-architecture-decision.md |
SPR 层级拆分与 S1/S2 架构 framing |
| 005 | blogs/.../spr/005-s2-fold-stack.md |
Fold stack 与 S2 方向 |
| 006 | blogs/.../spr/006-next-experiments.md |
Baseline battle 与 claim decision workflow |
| 007 | blogs/.../spr/007-context-proof.md |
controlled context routing proof |
| 008 | blogs/.../spr/008-s2-strategy-audit.md |
strategy audit、probability container、历史 syntax claims 降级 |
| 009 | blogs/.../spr/009-world-model-frames.md |
术语:world model / background field |
| 010 | blogs/.../spr/010-world-model-night-run.md |
night-run diagnostic;不是 positive WMT claim |
| 011 | blogs/.../spr/011-treeheap-algebra.md |
语言 claim 前的 TreeHeap algebra design |
| 012 | blogs/.../spr/012-subheap-kernel-search.md |
subheap kernel / convolution-like reasoning proposal |
| 013 | blogs/.../spr/013-treeheap-math-probe.md |
M0 pure math probe |
| 014 | blogs/.../spr/014-primitive-plus-order.md |
primitive + plus as order source |
| 015 | blogs/.../spr/015-primitive-plus-probe.md |
addressable plus / mod fold toy proof |
| 016 | blogs/.../spr/016-trainability-quiz.md |
ML entrance quiz:linear、XOR、modular addition |
| 017 | blogs/.../spr/017-treeheap-existence-proofs.md |
existence proof claims A/B/C |
| 018 | blogs/.../spr/018-treeheap-structural-inductive-bias.md |
learned encoding 与 conjugate kernel refinement |
| 019 | blogs/.../spr/019-soft-treeheap-gradient.md |
soft algebra、kernel-guided soft plus、multi-kernel training |
| 020 | blogs/.../spr/020-soft-treeheap-audit.md |
GLM audit、ARA scope repair、clean-kernel next proof |
| 021 | blogs/.../spr/021-c05-structural-proof.md |
C05 structural proof:path、subheap、recursive plus |
| 022-033 | blogs/.../spr/022-*.md 到 033-*.md |
后续数学底座、kernel、S1、decoder 进展;以各层 registry 为准 |
Canonical local blog source:
../../blogs/www.grepcode.cn/src/spr/
../../blogs/www.lostmap.cn/src/spr/
Public blog URLs:
https://www.grepcode.cn/spr/
https://www.lostmap.cn/spr/
Evidence Registry Pointers
| Evidence | Supports |
|---|---|
ara/s1-echo/evidence/README.md |
S1 capacity、token-only falsification、controlled context routing |
ara/s2-translation/evidence/strategy_audit/ |
S2 graph-builder bottleneck、vector collapse、probability containers |
ara/s2-translation/evidence/tmerge_diagnostic/ |
t_merge / background-field diagnostics |
ara/s2-translation/evidence/world_model_long_20260617_180554/ |
world-model night-run diagnostic,不是 positive WMT claim |
ara/m0-treeheap-math/evidence/treeheap_math_probe/ |
minimal algebra probe |
ara/m0-treeheap-math/evidence/primitive_plus_probe/ |
addressable plus 和 mod-fold toy |
ara/m0-treeheap-math/evidence/trainability_quiz/ |
linear regression、XOR、modular addition learning checks |
ara/m0-treeheap-math/evidence/soft_plus_probe/ |
Soft Plus gradient path、toy collapse、GLM feature-ablation audit |
ara/m0-treeheap-math/evidence/structural_c05_probe/ |
C05 structural ablation:flat/path-only vs subheap/path-subheap relocation |
ara/m0-treeheap-math/evidence/treeheap_diff_algebra_probe/ |
TreeHeap diff/distance/finite-difference learning signal |
ara/m0-treeheap-math/evidence/algebraic_decoder_probe/ |
finite-field TreeHeap algebraic decoders |
ara/s1-echo/evidence/s1_wmt_echo_kernel_probe/ |
WMT short BPE echo write/read kernel |
ara/s1-echo/evidence/s1_probabilistic_read_kernel_probe/ |
probabilistic stop/left/right read collapse |
当前 Open Proof Queue
最高优先级开放实验:
Write Mechanism Ablation- Claims:C5-003..C5-005
- 比较:
A: naive soft memory write
B: encoder soft plus
C: kernel-guided soft plus
-
Hard/Soft Consistency- Claims:C5-002, C5-007
- 测量 hard-soft output gap、collapse legality、route interpretability。
-
Existence Proof Suite- Claims:C4-002..C4-004
- 测量 address extrapolation、subheap relocation、prefix compression。
-
Real Context S1b- Claims:C1-005, C1-006
- controlled proof 必须面对 BoW、keyword、random-hash、real-corpus baselines。
-
Internal Decoder Bridge- Claims:M0-DEC-C01 与 S1-READ-C01 后续
- 目标:把 algebraic decoder 接到 learned semantic decoder,验证 internal node state 的可读性。
当前降级或拒绝的主张
| Claim | Decision |
|---|---|
| Token-only routing encodes contextual semantics | rejected |
| Current 3-epoch TreeHeap vectors solve syntax energy | downgraded |
| Historical checkpoint proves TreeHeap syntax | downgraded |
| Naive soft memory write proves TreeHeap algebra is trainable | rejected as too weak |
| WMT / Transformer replacement claim | not allowed yet |
Reviewer Notes
当前研究不允许声称:
TreeHeap beats Transformer on WMT.
TreeHeap has learned syntax from the current historical checkpoint.
Soft TreeHeap training is already proved.
Kernel-guided soft plus has learned clean routing from raw TreeHeap geometry.
当前研究允许声称:
S1 has strong path-hash capacity.
Token-only semantic routing failed.
Context-conditioned routing works in a controlled proof.
S2 fold/action signals are probe-predictable.
Current graph builder needs probability containers and better allocation.
TreeHeap math/toy probes support a narrow algebraic toolbox direction.
Soft Plus has a working gradient-path toy proof.
Kernel-guided clean route learning remains open.
Subheap structure carries unseen-depth relocation in the C05 structural toy.
Finite-field/path-addressed TreeHeap has algebraic decoders in M0 toy.
维护规则
当新的 SPR blog 改变 claim 时,必须在同一个 commit 中更新本文,或在长实验前补 follow-up commit。
每个新 claim 应该有:
ID
source
claim sentence
status
evidence pointer
falsification condition
next experiment