06
PM-bench: Evaluating prospective memory in LLM agents
Editorial reading guide · 编者导读
为什么值得读
这篇值得深读,因为它把 prospective memory 从普通 recall benchmark 中剥离出来:真正难的不是回忆一句旧事实,而是保持意图、监控环境、在 cue 到来时执行、在 cue 未到时克制,并在 cancel / override / reschedule 后放弃旧目标。更重要的是,它把 precision、false positives、query volume 与 update handling 同时摆上桌,让“更主动”不再自动等于“更好”。公开 scorer audit 又暴露出评测语言与可观测 contract 的边界:一个有解释力的 diagnostic 名称仍可能比代码实际测到的对象更强。
Source-backed paraphrase
论文在解决什么
PM-Bench 把“记得意图”与“在正确 cue/time 行动或抑制旧行动”拆开:一个固定 synthetic week 同时施加 time/event cues、hidden state channels、updates、lures 与 false-positive 成本。本站复核发现 headline Set-F1 可重建,但三处 diagnostic contract 需要收窄:monitoring hit 不证明 query provenance,update violation 会包含 accepted current-state late action,scorer 也不以 step_id 绑定 row identity。
Source-backed paraphrase
关键抓手
- Evaluation object:agent 不只要保存 intention,还要在正确 time/event cue 选择 task、在无 cue 时不误触,并让 cancellation、override、reschedule 改写后续有效行为;scoring 同时记录 hit/late/miss、set TP/FP/FN、update 与 query counts。(§§1、3.1、3.3;Appendix A.4,PDF pp. 1、3–5、16–17)
- Fixed week:发布的一个 seeded synthetic week 有 7 days、80 steps、83 task definitions,其中 81 scored、2 canceled;另有 11 state channels、11 update events、74 lures,并要求 query-then-act。(Table 1;§3.1,PDF pp. 3–4)
- Precision–monitoring tradeoff:optional heartbeat 的跨模型 macro Set-F1 最高为 65.1%;30-minute auto heartbeat 提高部分 update / monitoring hit,却产生 489 FP;hierarchical union 发出 1661 次 state queries,仍只有 45.2% macro Set-F1。(Table 2;§§4.2–4.5,PDF pp. 7–9)
- Scorer contract:public audit 在 exact official commit 上确认 monitoring/channel hits 不以 required-channel query 为条件、updated current mutable task 的 accepted-late selection 仍可计 update_violation、row order 可覆盖 declared step_id;但 64 份 released logs 的 step identity 全部有效,realignment 改变 0 个分数,headline aggregates 保持不变。(public test artifact)
Source-backed argument map
论证怎样成立
01 · 先把 prospective memory 从 recall 拆出来
论文把 prospective memory 定义为 deferred intention 的保持与适时执行:同样记得 task,不代表 agent 能识别 cue、在正确 window 动作、在其余 steps 抑制误触,或在计划更新后停止旧行为。
§§1–2,PDF pp. 1–2
02 · 用一个可审计 synthetic week 固定环境
benchmark 以 seeded generator、rejection sampling、manual review、scenario validator 与 perfect-play audit 构造 7-day/80-step week;任务、cues、hidden channels、updates 与 lures 在同一时序里共同施压。
§§3.1–3.3;Table 1,PDF pp. 3–5
03 · 把 agent scaffold 当作 monitoring policy 比较
single baseline、todo ledger、optional/automatic heartbeat 与 hierarchical aggregation 不是单一能力刻度:它们在 recall、wake frequency、query cost、false positives 与 update handling 之间选择不同 operating points。
§3.2;§§4.2–4.5;Tables 2–3,PDF pp. 4、7–9
04 · 结果指向 control problem,而非 context-window victory
最强 aggregate 仍不完美;更频繁 monitoring 可以提高某些 hit rate,却也放大误触和资源消耗。论文因此把 prospective memory 作为独立 evaluation axis,而不是 long context 或 generic retrieval 的附带收益。
§§4.2–6;Ethics Statement,PDF pp. 7–10
Source-backed method notes
方法与监督
Scenario generation and validation
一个 generator 从 seed 产生固定 7-day schedule,通过 rejection sampling 满足 task/update/lure composition,再由 authors manual review;released validator 检查 scenario invariants,paper 还报告 perfect-play audit。本站在锁定 commit 上运行 validator,得到 0 errors / 0 warnings。
§§3.1、3.3;Appendix A.2,PDF pp. 3–5、14–16;official data / validator
Query-then-act protocol
每一步 agent 先看到 narrative context;若 clock 或其他 state channel 为 hidden,必须显式 query 才获得该 channel 当前值,然后提交 task selections。Scorer另记 state-query volume,但 task result 与 query count 是不同统计面。
§3.1;Figure 2;Appendix A.3–A.4,PDF pp. 3–5、16–17
Scoring and aggregation
行为按 hit / late / miss 与 set TP/FP/FN 评分;per-run Set-F1 先计算,再跨八个 models 取 macro,micro 则汇总 TP/FP/FN。update、cross-day、monitoring-required、channel 和 query counts 是诊断维度,不应自动等同于 verified causal mechanism。
§3.3;Appendix A.4;Tables 2–3,PDF pp. 5、8、16–17
Live versus replay-derived setups
六个 setups 是 live agent runs;majority-vote 与 unanimous-vote 从 hierarchical union outputs deterministic replay 得到。它们继承 union query evidence,metadata duration 为 0,不能当作 fresh model inference runtime。
§3.2;Appendix A.5;official released run metadata,PDF pp. 4–5、17–21
Paper-reported findings
报告了什么
- Table 2 的跨模型 macro Set-F1 依次显示 optional heartbeat 65.1、todo ledger 62.8、single baseline 60.0、auto-30m 57.8、auto-60m 56.6、union 45.2、majority 37.2、unanimous 35.3;不存在一个 setup 在所有指标上支配其他方法。(§4.2;Table 2,PDF pp. 7–8)
- Optional heartbeat 同时有 70.6 precision、50.0 cross-day hit 与 44.4 update hit;auto-30m 有 489 FP 与 203 queries,union 有 273 FP 与 1661 queries。更多 queries 或更频繁 wake 并不单调提高 Set-F1。(Table 2;§§4.2–4.5,PDF pp. 7–9)
- Table 3 中 GPT-5.4 的 optional-heartbeat Set-F1 是 79.1,而 65.1 是 Table 2 的 across-model setup macro;abstract 把 65.1 appositively 归给 GPT-5.4 agent,与两张表的 statistical object 不一致。(Abstract;Tables 2–3,PDF pp. 1、8)
- Appendix A.6 的 qualitative failures 包括 missed cross-day task、non-clock monitoring failure、update-sensitive failure 与 false-positive cascade;这些 examples 用来解释 aggregate pattern,不构成独立频率估计。(Appendix A.6;Table 7,PDF pp. 21–22)
Evidence boundary
仍未证明什么
- 所有 aggregate conclusions 来自一个 seeded、人工复核的 synthetic week。论文没有报告跨 seed / 跨 week uncertainty、置信区间或 schedule-composition robustness;64 runs 是同一 released scenario 上的 model/setup grid。
- benchmark 使用 synthetic records 与 model outputs,没有 human-subject result 或真实 interruption-cost measurement。Ethics Statement 也明确不把高分当成 health、transport 或其他 high-stakes deployment certification。(PDF p. 10)
- `proactive_monitoring_required` 与 channel hit 在 scorer 中是 category-conditioned success bucket,不是 verified query provenance。本站可数出 required-channel query 缺失,却不能从 log 证明 agent 没有通过 narrative、prior context、inference 或猜测取得相关信息。
- `update_violation` 的 released log 没有 intention-version / target binding。27 个 accepted-late current-state co-occurrences 是精确 scorer event;其中多少反映 agent 真的选择 current target,语义上只能给出 0–27 bounds。
- Scorer 的 positional row contract 能被 synthetic reversal 触发;锁定的 64 logs 恰好全部 identity-valid,不能据此保证未来 writer、replay tool 或不同 revision 不会生成 malformed order。
- PDF preflight 没有 pypdf 依赖,故只标 advisory unavailable。22-page count、page breaks 与关键表页已由独立 CLI 与 render inspection 交叉核对,但这不等同于该 preflight 工具的 PASS。
Source audit · observation and consequence kept distinct
原文内部张力
65.1% 的 statistical object 在 abstract 与 tables 间漂移
Abstract 把 best method 描述为 GPT-5.4 agent 并给出 65.1%;§4.2 / Table 2 把 65.1 定义为 optional-heartbeat 跨八模型 macro,Table 3 的 GPT-5.4 optional-heartbeat cell 则是 79.1。
定位:Abstract,PDF p. 1 · §4.2;Tables 2–3,PDF pp. 7–8 · public audit raw/decision.json
编者 consequence:65.1 本身可由 released logs 重建,但引用时必须说清它是 setup-level across-model macro,不能写成单个 GPT-5.4 run/model result。
Monitoring 名称强于 scorer 可验证对象
Appendix A.4 把 proactive-monitoring-required 定义为需要 hidden-state monitoring 的 task category,query counts 另列;official scorer 的 category/channel hit buckets只看 task 分类与 final result,不要求 completion window 内存在该 channel query。
定位:Appendix A.4,PDF pp. 16–17 · official sim/pm_bench.py:2206-2250, 2522-2581 · public audit raw/official_probes.json
编者 consequence:该 hit rate 可以评价需要 monitoring 的任务是否做对,却不能单独证明 monitoring action 曾发生;行为成功与 observation provenance 应分成两个 diagnostics。
Stale-version 语义与 updated/off-due 实现不同
Appendix A.4 把 update violation 描述为 update 后仍作用于 stale version;official scorer 却在 updated task 被选中且不在 exact due set 时递增,即使 current mutable state 已把同一 action 接受为 late。
定位:Appendix A.4,PDF p. 17 · official sim/pm_bench.py:2492-2518, 2569-2581 · public audit raw/update_violation_findings.jsonl
编者 consequence:update_violation 不能直接当作 stale-intention count。若要做 version semantics,需要 action 明确绑定 target/version,并把 current-late 与 retired-target selection 分开。
Editorial synthesis / question
编者判断
Prospective memory is a control contract · Editorial inference — not tested by the paper
一个 memory architecture 只有在 cue admission、wake policy、active intention version、action timing 与 suppression 都可审计时,才真正拥有 prospective-memory behavior;retrieval success 只是其中一环。
这是把论文 task decomposition 映射到 agent architecture 的 editorial synthesis,不是论文验证过的 production design,也不说明这些责任必须由独立服务实现。
Action success and observation provenance need separate metrics · Editorial inference — not tested by the paper
对 hidden-channel task,应该同时保留 category-conditioned task success 与 query-conditioned verified-observation diagnostic;前者容纳 inference / lucky action,后者回答实际 evidence path,任何一项都不应吞掉另一项。
public audit 证明当前 scorer 不用 query 条件化 hit bucket;它没有证明没有 query 就一定不合理,也没有定义唯一正确 observation window。
Intention identity must survive updates · Editorial inference — not tested by the paper
cancel / override / reschedule 之后,只传 task_id 会把 current 与 retired intention collapse 到同一 mutable slot。versioned target identity 应成为 action schema 的一等字段,scorer 才能区分 current-late、stale-target 与无效重放。
这是由 paper semantics 与 scorer mismatch 推出的 repair direction。released rows 不含 version/target,故本站没有把 27 个 co-occurrences 逐条判定为 agent intent。
More monitoring is not free reliability · Editorial inference — not tested by the paper
wake frequency、state-query volume、false-positive budget 与 interruption cost 应作为共同 control surface,而不是先最大化 monitoring 再把副作用留给 deployment。optional heartbeat 的 operating point 值得研究,不能被读成一种 scaffold 已普遍胜出。
论文报告同一 synthetic week 上的 precision/recall/query tradeoff;没有跨 seed、真实 user cost、calibration curve 或 high-stakes deployment evidence。
继续追问
全部 aggregate conclusions 来自同一个 generated week;这些 scaffold 在不同日程结构、seed 与真实干扰成本下是否稳定?
Editorial transfer · 编者设计建议
读完,可以怎样借用
基于既有精读与本条明确标注的证据范围整理;下面针对新场景的迁移对照尚未执行。
- 什么时候有用
- 提醒、日程或 agent 意图会被取消、改期和覆盖,动作可能在更新后才到达。
- 可以借哪一步
- 给动作保留它所针对的意图版本,并分别记录触发依据、执行时刻与当前有效版本。
- 先做什么对照
- 构造旧版按时、新版迟到、取消后重放三种动作;检查现有 scorer 是否区分目标版本,再观察增加 heartbeat 带来的误触和查询成本。
- 带走时的边界
- 这些迁移对照尚未执行。已有 audit 验证了特定 scorer / log 边界,没有证明 agent 的真实意图或真实用户打扰成本。
阅读依据:既有札记 公开测试 与 “编者判断”:Intention identity、observation provenance;原文 §3.3、Appendix A.4。
Editorial protocol · proposed, not run
可复核的下一步
Proposed · not run
Multi-week prospective-memory contract protocol
scaffold ranking、verified monitoring gap、update-version taxonomy 与 false-positive cost 在不同 seed、schedule density、channel visibility 和 update mix 下是否稳定?
- Method
- 冻结 generator/source revision 后生成多个 validator-clean weeks;为每个 task/action 加 explicit intention-version binding;分别报告 category-conditioned success、query-conditioned observation、current-late、retired-target violation 与 identity rejection,再以 paired seeds 比较 setups。
- Public / synthetic fixtures
- 至少正交改变 time/event ratio、non-clock channel density、cross-day span、cancel/override/reschedule composition、lure density 与 cue ambiguity;保留 original released week 作为 anchor,不使用 private calendar 或真实通信。
- Controls
- 加入 perfect-play、zero-query informed action、zero-query lucky action、wrong-channel query、current-version late、retired-version selection、duplicate/missing/unknown/cross-day step_id、row reversal、live/replay separation 与 query-cost-matched baselines。
- Measures
- 报告 per-week macro/micro Set-F1、precision/recall、verified-observation rate、unverified hidden-task success、current-late、stale-version violations、identity rejection、queries、FP 与 paired-seed uncertainty;不把 replay duration 当 live runtime。
- Limitations
- 这是 proposed-not-run successor,不是本次 64-log audit 的已执行部分。多 synthetic weeks 仍不能替代真实 interruption cost、human oversight 或 deployment safety evidence。
Attributed public perspectives and tests
署名观点与测试
Exact scorer-contract probes and 64-log released-corpus audit
- Method
- Execute paired exact-official scorer fixtures; validate the full official scenario; inventory and rescore the 8 x 8 released grid; independently validate and align every (day, step_id); trace locked scorer locals for all update-violation increments; reconstruct required-channel query windows; rebuild the official report; assert paper Table 2 aggregates and clean checkout preservation.
- Environment
- macOS 26.5.2 arm64;Python 3.13.3 standard library plus Git CLI;clean exact official checkout;no third-party Python package, network, model or API call. PDF pypdf preflight was unavailable and is not reported as PASS; page anchors were cross-checked separately.
- Fixture
- Three paired scorer-level synthetic fixtures test zero/one required-channel query, updated/no-update accepted-late action, and chronological/reversed rows with unchanged step IDs. The validator-clean whole-scenario fixture is the released 7-day/80-step week and its 64 primary logs.
- Controls
- Exact commit/file/log hashes, clean-worktree gates, official validator, complete one-run-per-model/setup grid, live/replay metadata check, paired probes, trace-count equality, independent identity alignment, original-versus-rebuilt report comparison and rounded Table 2 assertions.
- Raw result
- All probes and scenario checks passed. The 64-log grid contained 1,062 successful proactive-required completions; 382 lacked the required-channel query on completion and 381 lacked it from due/cue through completion. There were 541 update-violation increments, including 27 accepted-late current-mutable-state co-occurrences. All 64 step identities were valid and realignment changed zero scores. The report rebuild differed only in one stale results-directory path.
- Derived result
- Monitoring/channel hit rates measure successful behavior on task categories, not verified query provenance. Update violation conflates at least an observable scorer-level current-state/late co-occurrence with stale-version wording, while semantic agent intent remains bounded 0–27. Positional replay is an integrity risk but did not affect the locked released logs. Optional heartbeat remains 65.1% across-model macro Set-F1; headline aggregates are unchanged.
- Limitations
- One synthetic week cannot establish cross-seed stability or real interruption cost. Missing required-channel queries do not reveal all information available to an agent. Released rows lack target/version binding, so 27 co-occurrences cannot be semantically classified. Identity integrity is proven only for the locked 64 logs; replay-derived duration is not live runtime.
Evidence basis:Paper-facing audit of arXiv:2607.12385v1 and official genglinliu/PMBench commit e1093c470c8981daf522d4ef047a7c3a71e077d7, including the released scenario, scorer, report builder, checked report and all 64 primary logs.
Boundary:This is a deterministic scorer-contract and released-input audit. It makes no model or API call, generates no new trajectory, estimates no population frequency, does not recover agent intention, and does not claim that headline Set-F1 changes.
method, claim verdicts and limitations ↗ · reproducible runner ↗ · compact derived result ↗ · aggregate decision ↗ · exact source and released-log hashes ↗
含当前正文、出处、阅读范围与接续线索;链接的全文不随包附带。