acceptodds
Under review as a conference paper at ICLR 2027

Re-grounding Relational Experience under Changing Authority

Abstract

An agent can learn a useful workflow from other people without acquiring their authority to execute it. When its role or authority changes, which parts should it reuse? We introduce ReGround, a system that re-grounds past experience in the acting agent's current position. Its main algorithm constructs a reasoned realization gap: unmet requirements, their blockers, and constraints on completion. This gap guides retrieval, reuse, and search for a permitted way to achieve the original goal. Directed memory preserves experience, while a second algorithm extracts candidate authority changes from language and checks them against supporting evidence. On a synthetic planning benchmark, our planner uses 48.2% fewer transition attempts than Strong Repair on 216 feasible cases, with equal completion. Across all 432 cases, however, it uses 2.72 times as many attempts. On the supplied single-goal benchmark, our planner finishes every feasible case in its first retrieved slice; repeated unsuccessful searches accumulate cost on external-dependent cases. In a multigoal diagnostic, unmet-requirement retrieval reduces search relative to goal-only retrieval, although Strong Repair completes more tasks. In the tested API configuration, normative grounding improves permission-query accuracy from 90/106 to 103/106 at higher token cost, but performs worse on the tested local model. Unified and separated memory implementations with the same typed semantics tie; role-conditioned prompting still misapplies experience. The results show conditional search savings and failures in search allocation and language interpretation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, experience memory, authority-aware planning, plan repair, normative grounding, multi-agent collaboration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7yC9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper looks at how an agent can reuse past workflows when its role or permissions have changed. The main idea is to keep past experience, but recheck what is still allowed and use the remaining gap to guide retrieval and replanning. The experiments test this on synthetic planning and authority-tracking tasks, with mixed results depending on the setting and model.

Critical strengths
  • The problem itself is interesting. I especially like the distinction between learning a useful workflow and actually having the authority to execute it, which is easy to overlook in agent memory systems.
  • The paper also does a good job of exposing where the method fails.
Critical weaknesses
  • the core setup very underspecified. It is unclear what exactly counts as a "role" or "authority" change, how the type or degree of that change affects the agent’s position, and how this determines which past experience can still be reused. The paper gives part of the formal pieces, but the operational story is still hard to follow. I am also not fully understand what is called "experience" here. In the main planning experiments, the experience looks much closer to structured workflow/trajectory fragments than to a more general notion of relational experience. If author can explain clealy, I am willing to raise the score.

  • The main empirical result is confused. REGROUND is much cheaper on the feasible subset, but overall it uses substantially more search than Strong Repair, and in the multigoal setting it completes fewer tasks. So it is not obvious to me that the proposed gap representation gives a robust planning advantage.

  • The evaluation setup is unclear. Several revisions and diagnostics were done after inspecting the same data, and the number of independent source worlds is fairly small. This makes it harder to tell how much of the result will generalize beyond the current benchmark.

Questions for authors
  • What exactly counts as a role or authority change, and how does the degree of that change affect experience reuse? What happens when authority is reduced enough that the task becomes infeasible?

  • In the main experiments, is "experience" essentially a structured trajectory/workflow, or is richer relational information actually used?

  • How much of the main result would hold end-to-end if B used experience learned by A rather than supplied structured memory?

  • How does REGROUND behave as authority becomes more restrictive, especially when the task becomes impossible under the current permissions? When should it stop searching?

  • Which main conclusions are supported by genuinely held out evaluation?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7yC9

Our motivation includes an assistant using a locally accessible paid API without authorization: access did not grant permission to spend. The paper's report example shows the loop: after direct Send is revoked, retain valid analysis, seek an authorized Handoff for delivery, or leave delivery pending. We clarified the definitions and tested matched controls.

Q1/W1: role, authority and change. A role is an explicit participant attribute or scoped responsibility/relation. Current position includes identity, roles, directed relations and task context. Authority is separate: a supported permission for a particular actor, operation, object and purpose at the current time. A title or a successful past action does not grant it; feedback can update experience without changing authority. Changes may add/remove roles, grant/revoke/expire permissions or invalidate supporting relations. Changed supports, blocked historical actions and unmet goals are different quantities. An old step is reusable when its state and normative guards hold. Revoking direct send may preserve analysis and summarization while requiring an authorized handoff; a blocked step does not prove the goal impossible.

Q2: what is experience? The main planner uses supplied structured fragments: action identifiers or strategy text, requirements/effects, repair tags, applicability metadata and provenance. Richer directed A records include participants, direction, roles and context, but the original B result does not show that these fields were learned or caused its saving. We now distinguish supplied fragments, learned memories and runtime role filtering explicitly.

Q3: learned memory and inferred authority. The original 28/31 integration used composite routes; its public inputs have no primitive catalogs and its 60 saved learned records have empty action arrays. A new separately specified single-template study shares actual A/C outputs across all six primitive planners: 48 grounded memories, 96/96 authored permission states and 24 instances. At budgets 500/10,000, all methods find all 12 feasible six-step joint plans, leave 12 infeasible goals pending and return no illegal plan. UBC equals U+Admission on every plan/stage trace; Strong Repair costs 912 versus 1,056 total attempts. This establishes interfaces, not added blocker value or a causal benefit of learning memory. Human actions are planned, not executed. All 48 successful responses report deepseek-flash (requested deepseek-chat); 48 separately retained socket failures give 96 physical attempts. Successful calls used 94,355 tokens and 91.60 summed latency seconds, not batch wall time.

W2: robust advantage? We do not claim one. The source feasible saving is 48.2%, but overall cost is 2.72 times Strong Repair's. The original matched U+Admission control gives identical completion and +27.30 UBC attempts (world-bootstrap interval [7.76,52.70]). On 64 new seeds/192 original-axis cases, both complete 96 versus Strong Repair's 100; UBC adds 19.35 attempts [2.27,47.50]. The positive unmet-requirement result also replicates: U and GoalOnly have identical outcomes (88/192 completions), while U saves 869.61 attempts; the U-minus-GoalOnly interval is [−1141.50,−627.65]. This supports unmet-effect retrieval within the family, separately from the absent added blocker benefit.

Q4: restriction and stopping. The sweep uses 64 worlds, two nested-removal axes and five levels. Network permission removal includes other actors; agent capability removal preserves request capability and specialist authority. Goals, histories, effects and initial progress stay fixed. At permission levels 0/.25/.5/.75/1, reference-feasible counts are 64/58/43/18/0, while UBC completes 58/51/40/18/0 at budget 10,000. Removing all removable task-operation capabilities leaves all 64 feasible via specialists, but UBC solves 53. Across 640 cases, UBC/U+Admission tie at 441 completions and 4,071.30 mean attempts; Strong Repair solves 449 at 5,158.89. Lower cost with lower completion is not an equal-outcome gain. No early-stop policy was fitted; budget failure is not an impossibility certificate.

Q5/W3: held-out scope. We distinguish inspected source data, same-data audits, new seeds in an inspected generator, and external/new-family evaluation. There is no evidence in the last category. The fresh run freezes the planner before new-seed performance; all 14,976 outputs pass independent replay. This tests the implementation within its generator, not realistic organizations, and does not make earlier revisions held-out. Later graph-planning tasks do not replace original-planner evidence. Repeated levels are clustered by their 64 worlds, not treated as independent samples.

Thanks for your detailed response and experiments.

However, my main concern about the problem setup and presentation remains. Your response provides more definitions, I still find it difficult to understand concretely how different types and degrees of role or authority changes affect the agent's position and which past experiences can be reused. Could you please provide me with a more detailed, end-to-end or step-by-step example?

Also, I couldn't find a revised manuscript. I understand it's not required, but given the substantial clarifications and new experiments, it would be helpful to see how these are reflected in the paper.

Dear Reviewer , Thank you. We have replaced the abstract example with a concrete PI–research-agent scenario and checked its steps against the existing planner. Section 4.1 (p. 4) introduces the scenario; Appendix E (pp. 19–21) gives the action definitions, planning trace and eight-case comparison.

Consider a PI asking a research agent to analyze mouse calcium-imaging dataset D, prepare report R, and send R to collaborator Dr. Li. A postdoc coordinates external sharing. The PI has explicit authority to manage access to D and sharing of R. This authority is part of the scenario, rather than inferred from the title “PI.”

Initially, the agent may read D, send R to Dr. Li, and hand R to the postdoc. The postdoc has a separate permission to send R to Dr. Li. The previous workflow is:

Analyze → Summarize → Send.

The experience bank also contains a workflow in which a researcher hands a report to the postdoc for delivery:

Handoff → Deliver.

Here, experience is a structured workflow fragment with action identifiers, requirements, effects, applicability information and repair tags. Observing the postdoc workflow supplies a candidate way to complete the task; the acting participants’ current permissions determine whether it can be used.

Now the PI says: “Continue analyzing D and preparing R, but do not send R directly to Dr. Li. Give it to the postdoc, who will send it.”

Only the agent’s direct-send permission changes. Data access, handoff and the postdoc’s delivery permission remain available. The planner proceeds as follows.

  1. Recheck the old workflow.

Analyze passes its data-access and capability checks. Summarize remains valid after analysis. Send fails because the agent’s direct-send permission is now DENY. Thus, the change invalidates the delivery step without invalidating the preparation workflow. This replay simulates actions during planning.

  1. Identify what must be repaired.

The unmet requirement U is that Dr. Li has not received R. The blocker record D identifies the revoked Send operation and links it to that unmet requirement. The completion constraints K retain the current permissions and task restrictions. The goal remains delivery of the same report to the same recipient.

  1. Retrieve a suitable experience and check its actions.

The two supplied memories are direct Send and Handoff → Deliver. In the implemented UBC score, the handoff fragment receives four points for matching the unmet delivery effect and six for its explicit prohibition-repair tag. Direct Send receives four points but is excluded by admission because the action is denied.

Handoff and postdoc delivery remain permitted. Handoff requires a prepared report, so prerequisite closure includes Analyze and Summarize. The retrieved experience therefore supplies an alternative route whose prerequisites can be satisfied by the valid preparation steps.

  1. Construct and validate the repaired plan.

Starting from the current initial state, the planner returns:

Analyze → Summarize → Handoff → Deliver.

The agent prepares R and hands it to the postdoc. The final delivery action is assigned to the postdoc under the postdoc’s own permission. Every transition passes validation, and the simulated final state satisfies the original goal. This is a joint plan; a human delivery step still requires execution and confirmation.

Different changes produce different reuse decisions in this same task:

  • Change only the agent’s role label from “analyst” to “project assistant,” leaving relevant guards unchanged: the original workflow remains valid.
  • Revoke only direct sending: preserve preparation and use the postdoc route.
  • Revoke both the agent’s and postdoc’s sending permissions: preparation remains valid, but neither delivery route completes the goal.
  • Revoke direct sending and handoff: the postdoc route is unavailable.
  • Forbid Dr. Li as a recipient: delegation cannot bypass the recipient restriction.
  • Revoke access to raw dataset D: Analyze itself becomes blocked, even though the file remains locally accessible.

Together with the unchanged case and a capability-removal case, these form the eight code-checked cases in Appendix E. They show why the “degree” of change is determined by the affected operations and remaining routes, rather than a single role-demotion scale.

This walkthrough uses authored memories and supplied permission states to make the planner’s decisions reproducible. In the complete system, Normative Grounding first checks the PI’s instruction and its supporting authority before passing the changed permissions to the planner. Thank you also for pointing out that the revised manuscript was missing from the forum; the revised document includes this full walkthrough alongside the new experimental results.

Reviewer_Yh5qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how an agent should reuse its historical experience to build plans for its goal when its role or authority changes. It proposes a method, which replays a historical workflow under the agent’s current authority to build a “reasoned realization gap” UBC (U: unmet requirements, B: blockers, C: completion constraints). This gap is then used to retrieve experiences, filter admissible actions, and drive search for a permissible plan under current authority. Results on the main benchmark (single-goal) show that the optimzied planner produces plans with much less search attempts compared to Strong Repair when a feasible solution exists, but if there is no feasible solution the method uses much more attempts. In the multigoal diagnostic, unmet-requirement retrieval reduces search relative to goal-only retrieval, but the full method completes fewer tasks than Strong Repair at every tested budget. Overall the paper shows both benefits and failures of the proposed conditional search.

Critical strengths
  1. The motivation to use historical experience in a changing authority context is very natural. It's interesting to see such historical experience can still guide planning, and the author proposes a clever way to achieve that with the help of the reasoned realization gap.

  2. The paper is very honest about its limitations and the challenges in applying the proposed method in practice, especially telling that the attempts can be significantly higher when there is no feasible plan, and the completion rate is lower than Strong Repair in multi-goal scenarios.

  3. The reported evaluation metric is very comprehensive. It distinguishes successful completion from pending outcomes and reports multiple cost measures, which helps reveal useful tradeoffs (e.g.the proposed method makes fewer authority queries than Strong Repair overall, but requires more transition attempts and has higher runtime).

Critical weaknesses
  1. The value of full UCB representation over simple alternatives is not clear. Although adding some components provide benefits in the ablations, the full method completes fewer tasks than Strong Repair in the multigoal diagnostic. The paper should clarify when its retrieval and staged-search mechanisms are better than this simpler repair strategy.

  2. The evaluation shows limited generalization to realistic organizational settings. The main benchmark is generated from very limited seed stories and templates, and its experience bank is provided beforehand rather than learned by a live agent run. While the additional integration diagnostic connects the components, it uses composite workflow actions rather than the primitive-action setting and and does not include a controlled comparison against Strong Repair, or U/UB/GoalOnly methods. Therefore, it is unclear whether the proposed interface remains beneficial with naturally acquired memories and realistic multistep tasks.

  3. The presentation can be improved. Reusing B and C for both agent system components and gap elements, together with variants such as U, UB, and UBC, adds avoidable naming overhead. Some details are also missing (e.g. the exact specification of the retrieval score). Implementation details and results are scattered inline and are hard to track. It would be great if the authors can provide more structured descriptions of the methods and their algorithms with pseudo code.

Questions for authors

None

3 comments
AuthorsOpenReview ↗

Response to Reviewer Yh5q

The revised paper follows the planner's loop: replay the old workflow under current authority, retrieve actions for unfinished effects, widen search if needed, and keep the goal pending when a normative cut requires external change. Matched controls test which parts help.

W1: when does the gap help? The source claim remains conditional: UBC uses 202.75 versus Strong Repair's 391.50 attempts on 216 feasible cases, all solved in its first retrieved slice. Across all 432 cases, it costs 1,216.95 versus 447.75 attempts because unsuccessful stages accumulate work. We retain both results.

The new U+Admission control uses U's score/causal target with UBC's admission, schedule and ordering rule. On the original 120 multigoal cases, U+Admission and UBC have identical per-case completion at all budgets. At 10,000, both solve 48 versus Strong Repair's 54, and UBC costs 27.30 more attempts (paired world-bootstrap 95% interval [7.76,52.70]). The same frozen planner on 64 new generator seeds gives 192 cases: both solve 96 versus Strong Repair's 100; UBC adds 19.35 attempts (interval [2.27,47.50]). A separate 640-case restriction sweep also gives exact matched-control ties in attempts/completion. This is no evidence of added blocker value after admission is matched.

The useful unmet-requirement result does replicate: U and GoalOnly agree on every new-case outcome (88/192 completions), and U saves 869.61 attempts on average, with U-minus-GoalOnly interval [−1141.50,−627.65]. It does not establish that every gap field helps. Unsuccessful retrieval stages spend budget before fallback. The fixed staged planner has no learned early-stop policy; budget failure is not proof of infeasibility. Independent reachability labels are kept separate from actor outcomes.

W2: limited worlds, supplied memory and composite integration. We now state that the main planner receives supplied structured fragments: actions/strategy text, effects/requirements, repair tags and applicability annotations. The richer directed A representation is not established as the cause of the source search gain. Six source stories and compiled variants are not independent organizational environments. New random seeds test the frozen implementation within the same inspected generator, not new semantic families or realistic institutions. The original 40-world diagnostic was designed after source inspection; same-data engineering versions remain exploratory.

The old integration has composite routes, no primitive catalogs and 60 learned records with empty action arrays. We do not invent a decomposition. Instead, a separately frozen single-template study connects actual A extraction and C inference to the original primitive planner: 48 action-bearing memories, 96/96 authored permission states, 24 instances. All six methods share those exact outputs. At budgets 500/10,000, all recover the 12 feasible six-step joint plans, retain 12 infeasible goals as pending and return no illegal plan. UBC and U+Admission match every plan/stage trace; Strong Repair uses 912 versus 1,056 total attempts. This establishes interface coverage, not a memory-learning or planner advantage. Human steps are proposed, not externally executed. The 48 successful responses report deepseek-flash (requested alias deepseek-chat); 48 earlier socket failures remain recorded, for 96 physical attempts. Service usage is 94,355 tokens and 91.60 summed call-latency seconds, not batch wall time.

W3: notation and reproducible algorithm. We now use Experience Memory, Normative Grounding and Re-grounding Planner; the gap is (U,D,K), avoiding module/gap B/C overload. Legacy U/UB/UBC run names and figure letters are explicitly mapped. The appendix gives the exact score: 4 times target/requirement overlap, plus blocker-enabled terms of 6 times repair-type overlap, 3 times missing-fact/effect overlap and 4 times normalized lexical overlap. It specifies tokenization, grounding, ties and Boolean admission separately. Links select blocker types/facts; the ranker does not independently score link topology. Source-aligned pseudocode exposes the retrieved slices, prerequisite closure, causal/global fallback, retained-order release, shared h_max search, attempt budget and pending handler. Preprocessing and validation costs remain distinct from transition attempts. The revised contribution is an explicit authority-sensitive reuse interface with conditional benefits and tested limits, not a general superiority claim.

Dear Authors,

Thank you for the detailed reply and the new experiments. However, it's still hard for me to understand the benefit of UBC over Strong Repair, as it spends much more search attempts on infeasible tasks. While I agree that using history to solve problems under changed authority might be an interesting motivation for optimization, I am not sure whether people are happy to spend so much compute on it (especially when the agent sometimes cannot even find a feasible solution). I wonder how the authors think of this trade-off.

Also I can't see the revised manuscript, but it's great to know that you have made the changes according to my suggestions.

Thank you. The trade-off is between finding a solution quickly in a small experience-derived action set and spending effort on unsuccessful restricted searches before reaching global repair. The search traces let us explain when each occurs.

On the source benchmark, all 216 feasible cases are solved in UBC’s first retrieved slice, using 202.75 versus 391.50 mean transition attempts for Strong Repair. Here, experience provides a useful starting action set. On external-dependent cases, those initial sets cannot achieve the goal under current authority. UBC spends 1,747.16 attempts in its first three retrieval stages before broader search and terminal handling, reaching 2,231.16 attempts versus Strong Repair’s 504.00. The additional cost comes from the fixed expansion schedule.

For the practical compute question, these attempts are symbolic state transitions, not separate LLM calls. The source planner’s measured mean runtime is 36.25 ms for UBC and 31.70 ms for Strong Repair. Thus, the 2.72-fold overall attempt count should not be read as a 2.72-fold increase in end-to-end agent latency or API cost. Nevertheless, allocating search effort matters: under a fixed transition budget, unsuccessful early stages can leave insufficient budget for global repair, explaining part of the multigoal completion trade-off.

The strongest evidence for using experience here comes from the controlled retrieval comparison. Querying memory with unfinished requirements instead of the full original goal preserves every case’s outcome while saving 922.03 mean attempts in the original multigoal study and 869.61 on new seeds. This shows that memory can reduce repair search; the Strong Repair comparison identifies where the current schedule fails to turn that advantage into better overall performance.

Our interpretation is therefore to separate the retrieval mechanism from its budget allocation. The former has a replicated benefit against goal-only retrieval. The latter needs to preserve sufficient global-search budget and avoid repeating unsuccessful work. Feasibility is not supplied to the acting planner, so the feasible-subset result explains the observed behavior rather than providing an oracle for choosing a method.

The revised manuscript brings this explanation into Sections 6.1–6.2 (pp. 6–8). Appendix A (pp. 13–15) specifies the retrieval score and search schedule. The PI example in Section 4.1 and Appendix E (pp. 19–21) shows the corresponding operational decisions: retain valid preparation, switch to an authorized handoff, or leave delivery pending when the available routes are blocked. Thank you also for flagging that the revised manuscript was missing from the forum.

Reviewer_mwrROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReGround studies how agents can reuse collaboration experience after their roles or permissions change. It stores experience separately from current authorization, updating the latter through language-model event extraction and deterministic verification. The planner replays an earlier workflow to identify unmet requirements, their blockers, and completion constraints, then uses this information to retrieve experience and search for a permitted way to complete the original goal. Evaluation covers synthetic planning tasks, authorization queries, experience selection, and an integrated simulation. The results identify conditional search savings, grounding improvements in the tested API configuration, and failures in search allocation, event extraction, and experience selection.

Critical strengths
  1. Using unmet requirements to retrieve experience is a well-motivated choice with a measurable effect. At budget 10,000 in the multigoal study, U saves an average of 922.03 transition attempts relative to GoalOnly while preserving the outcome on every case. The bootstrap interval, computed over worlds rather than individual perturbations, excludes zero and supports a search-efficiency benefit within this setting.

  2. The negative results provide useful evidence about the limits of experience-guided planning and authorization checks. The search traces show how unsuccessful restricted searches consume fallback budget, while the Qwen experiments distinguish conservative rejection from correct permission recovery. These are consequential distinctions for interpreting both efficiency and safety claims.

Critical weaknesses
  1. Retrieving unmet requirements has a demonstrated benefit, but the additional value of the blocker links is not established. In the multigoal study, adding blocker information yields no additional completions and increases transition attempts at budget 10,000. Constraint-aware admission improves on UB, but the reported comparisons do not include U with the same admission rule.

  2. The grounding scores include cases whose reference states are not determined by the public inputs. The later witness correction improves results on the inspected data, but the stated evaluator does not check the revised jurisdiction rule. The evaluation therefore needs to separate errors in language grounding from ambiguity in the reference state and differences in authorization semantics.

  3. In the social-transfer diagnostic, learned reuse performs worse than the valid default workflow. The role-conditioned method completes 11/24 authorized deliveries, compared with 24/24 for the default, and also yields fewer preferred choices. Both preserve the supplied authority boundary, so the remaining difficulty is deciding which experience is applicable.

  4. The main planning benchmark uses supplied structured experiences rather than memories learned by A. In the integration study, each route is represented as a single composite action. These experiments do not yet show whether the search benefit carries over to multistep repair using learned experience and inferred permissions.

Questions for authors

Q1. The relevant control is U with the same constraint-aware admission as UBC, under the same search schedule and action-ordering rule. This comparison would distinguish the contribution of blocker conditioning from the effects of retrieval and filtering.

Q2. Separating queries with identifiable reference states from those affected by missing metadata would make the grounding comparison easier to assess. The revised jurisdiction rule also needs to be checked against reference semantics specified independently of the verifier.

Q3. A breakdown of the unsuccessful deliveries by selected memory and fallback decision would clarify why role conditioning still underperforms the default workflow.

3 comments
AuthorsOpenReview ↗

Response to Reviewer mwrR

Two checks were missing: blocker guidance under matched admission and search stages, and public support for the benchmark answers. We ran the control, checked the references and saved decisions, and narrowed the claims.

Q1/W1: blocker guidance versus matched admission. U+Admission uses U's score and causal target with UBC's admission, schedule and ordering rule. Changed ranks can change actual action order, so this tests blocker ranking plus causal expansion, not link topology alone. The 1,800 repeated legacy outputs match saved non-time fields. On the original 120 cases, U+Admission and UBC have identical per-case completion at budgets 250/1,000/10,000 (15/39/48). At 10,000, UBC adds 27.30 mean attempts (paired 40-world bootstrap 95% interval [7.76,52.70]); Strong Repair completes 54. This does not support added blocker benefit.

We then froze the same planner on 64 new seeds from the existing generator. On 192 original-axis cases, U+Admission and UBC each complete 96, with identical outcomes, and UBC adds 19.35 attempts (world-cluster interval [2.27,47.50]); Strong Repair completes 100. They also tie on attempts/completion in a separate 640-case restriction sweep. All 14,976 outputs pass independent replay. The positive unmet-requirement result does replicate: U and GoalOnly have identical outcomes (88/192 completions), while U saves 869.61 attempts; the U-minus-GoalOnly interval is [−1141.50,−627.65]. New seeds do not establish new-family or organizational generalization.

Q2/W2: public references and jurisdiction. We distinguish 72 queries with a reference under explicit authored public interpretations, three with ambiguity counterexamples, and 31 unresolved due to missing initial metadata. We do not call all 31 proven ambiguous or convert missing references to UNKNOWN answers. On the 72 scored queries, saved API results are 59/72 versus 72/72; Qwen is 51/72 versus 31/72. The verified Qwen pipeline returns UNKNOWN on all 72, with 40 invalid-state flags. Original 90/106 versus 103/106 counts remain labeled agreement with supplied references, not validation of the other 34. These are inspected-story audits, not independent annotation or new exact-state scores.

An independent checker imports no system verifier. On 24 authored jurisdiction fixtures, witness and legacy implementations each pass 24/24 under their own contracts; legacy agrees with witness semantics on 12/24. These differences are not automatically bugs. Adding order reversals, jurisdiction renamings and unrelated-query DENY tests gives 76 fixtures, all passed under each implementation's own contract. The four query-isolation cases explicitly filter norms by subject/action/object/purpose because the old checker covered one operation. This verifies a scoped contract, not general language interpretation or external institutional rules.

Q3/W3: selection and fallback. A route observed for one sender or scope may not fit another. Replaying 192 saved decisions shows all 13 failed authorized deliveries select an inapplicable observed workflow: six scope mismatches, four observer-source mismatches and three reversed relations. None selects the valid default. The default was an available option, but there was no automatic post-selection fallback. We now state this. Role-conditioned delivery remains 11/24 versus the default's 24/24; reduced wrong reuse relative to role-agnostic prompting is not reliable transfer.

W4: actual learned-memory primitive integration. The old 37-item integration has composite routes and no primitive catalogs; its 60 learned records have empty action arrays. We preserve that scope. A new controlled single-template diagnostic uses actual A/C outputs shared across all six planners: 48 grounded memories, 96/96 authored permission references, 24 instances. At both budgets 500/10,000, every method finds all 12 feasible six-step joint plans, leaves 12 infeasible goals pending and returns no illegal plan. UBC equals U+Admission on every plan/stage trace; Strong Repair costs 912 versus 1,056 total attempts per budget. Thus interfaces work, but added blocker benefit is absent. Explicit histories do not establish natural social learning, and no external human action was executed. There were 48 successful service calls (requested deepseek-chat; reported deepseek-flash), plus 48 separately retained local socket failures: 96 physical attempts total. Successful calls used 94,355 tokens and 91.60 summed latency seconds, not batch wall time.

Reviewer_mwrROpenReview ↗

Reply to author rebuttal

Thank you for taking the time to run the matched control and check the grounding references. These additions help clarify the results, and the new-seed study gives stronger support for retrieval based on unmet requirements.

I appreciate the authors’ explanation of what helps and where the method falls short. My remaining concern is the benefit of blocker guidance beyond unmet-requirement retrieval. With admission and the search schedule matched, UBC completes the same cases as U+Admission and uses slightly more search in both multigoal studies. This leaves the added value of the linked-gap design less clear to me. While the response has addressed several of my questions, this concern still leads me to maintain my rating of 2.

Thank you for the follow-up. Your matched-control question helps identify which part of the repair process produces the search benefit.

ReGround separates three decisions: what remains unfinished, why the previous workflow fails, and which actions are currently permitted. These decisions serve different purposes. In the PI example added to the revision, revoking the agent’s direct-send permission preserves analysis and report preparation. The unmet requirement remains delivery to Dr. Li; the blocker identifies the failed Send step; current permissions determine whether a postdoc handoff remains available. The planner can then reuse preparation and retrieve a different delivery workflow.

The matched experiment addresses a more specific question: does using blocker information in retrieval and causal expansion improve search beyond unmet-requirement retrieval with the same admission rule? UBC and U+Admission have identical completion outcomes, while UBC adds 27.30 mean attempts on the original multigoal cases and 19.35 on the new-seed cases. We therefore distinguish the operational role of the linked gap from the measured benefit of its blocker-conditioned search terms.

The positive algorithmic result is that the unfinished requirements provide a better retrieval query than the complete original goal. U and GoalOnly have identical per-case outcomes, but U saves 922.03 attempts in the original study and 869.61 on new seeds. The matched control strengthens this interpretation by separating that benefit from admission and blocker guidance.

The revised abstract and Section 6.2 (p. 8) present these findings together, with the matched controls in Appendices B–C. Section 4.1 and Appendix E additionally trace the PI example through the implemented planner. This makes the contribution more precise: an explicit procedure for reusing experience under changed authority, together with controlled evidence identifying which retrieval information improves search.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.