acceptodds
Under review as a conference paper at ICLR 2027

IEL: Intent-Level Experience Learning for LLM-based Customer Service Agents

Abstract

Large language model agents increasingly operate in interactive service environments where a single dialogue may contain multiple user intents with different outcomes. This creates a granularity mismatch for experience learning: episode-level success or failure cannot determine which intent-specific behaviors should be retained, discarded, or reused. We introduce Intent-level Experience Learning (IEL), which treats a cross-turn intent instance as the unit of experience attribution and procedural learning. IEL first reconstructs intent instances from multi-turn interactions, attributes dialogue and tool evidence to individual instances, and verifies their outcomes using intent-specific goals and completion criteria. It then induces reusable procedures from verified successful instances, while using verified failures, when available, to identify recurring error patterns. Taxonomy-absent intents are accumulated as candidates and may subsequently be merged with existing intents or promoted to new canonical intents as evidence grows. This design implements the principle of credit assignment before procedural abstraction, where credit assignment refers to associating interaction evidence and outcomes with individual intent instances rather than with the entire episode. We evaluate IEL on ECom-Bench and a cross-category extension covering previously unseen product categories, using three executor LLMs. IEL obtains the highest mean score in all evaluated executor–metric combinations, with improvements of up to 7.58 percentage points in pass1 and 13.18 points in pass3 in the in-domain setting. These results suggest that intent-level experience attribution provides a useful basis for learning reusable procedures from mixed-outcome service interactions.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
LLM agents, experience learning, agent memory, customer service agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Mj9YOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes IEL, an experience-learning framework for large language model-based customer service agents. Specifically, it aims to address the issue in interactive service environments where a single conversation may involve multiple different user intents, leading to a mismatch in the granularity of experience learning. IEL was evaluated using three execution models on ECom-Bench and a cross-category extension, and the comparison results show that IEL achieved the highest average pass^k scores, demonstrating its effectiveness.

Critical strengths

This paper deals with the practical issues of learning from multi-intent dialogues, where even a failed conversation can still have successfully handled parts for a single request. Specifically, it offers a reasonable way to preserve these useful parts by using cross-turn intent instances as the unit of experience. Overall, this gives the method a clear motivation and practical value for service agents.

Critical weaknesses
  1. The construction and evaluation of ECom-Bench-XCat are not sufficiently documented. Appendix A reports task counts, user profiles, and environment contents, but does not explain whether the tasks were manually written, generated by a model, or adapted from existing templates. Besides, pass^k is appropriate for measuring overall task success, but it does not show which parts of a multi-intent task improve.

  2. Although the paper discusses recent procedural memory approaches such as MemP, these closely related methods are not included in the experimental comparisons. The current baselines cover general experience reuse and workflow learning, but may not adequately represent recent advances in structured procedural memory. Consequently, it remains unclear whether IEL's improvements stem specifically from its proposed intent-level credit assignment mechanism or more generally from structured experience organization and retrieval.

Questions for authors
  1. How were the XCat tasks and evaluation annotations created and validated? In particular, how were task goals translated into Action, Search, and Output checks, and how was task feasibility verified?

  2. Could the authors explain why closely related procedural memory methods, particularly MemP, were not included in the experimental comparisons? Additionally, could the authors provide further experimental evidence or analysis to clarify whether the observed improvements are primarily attributable to intent-level credit assignment rather than structured experience organization and retrieval?

3 comments
AuthorsOpenReview ↗

Response to W1 & Q1: How XCat tasks and checks were built and validated

W1 & Q1

"How were the XCat tasks and evaluation annotations created and validated? In particular, how were task goals translated into Action, Search, and Output checks, and how was task feasibility verified?"

Response: Thanks for this valuable question! We will add the following description to Appendix A. We adapted the service scenarios of ECom-Bench to clothing and food. An LLM drafted the user profiles and goal lists, and the authors revised every task by hand. Each task gives the simulated user a profile and two to six ordered goals (252 in total), all grounded in a fixed store state of 36 products, 20 customers, and 60 orders with their logistics, invoice, and after-sales records. We then annotated each goal with the tool calls it requires, following a written guideline and ECom-Bench's annotation schema and scorer. A goal that only asks for information is checked by a Search annotation: the required call, with its exact arguments, must appear among the agent's calls. A goal that changes the store state is checked by an Action annotation: the gold calls are replayed on a copy of the database, and the agent's final database must match the result exactly. An Output check requires the agent's replies to contain predefined key strings, matched as exact substrings. Every task and its checks were then reviewed against the guideline by an author who had not annotated that task, and disagreements were resolved by discussion. Finally, we replayed each task's gold calls in the environment and confirmed that every call succeeds and that the replayed state passes all checks, so every task is solvable. We also confirmed that every identifier a goal needs is either given to the simulated user or reachable through the tools.

AuthorsOpenReview ↗

Response to W1 (cont.): Finer-grained metrics beyond pass^k

W1 (cont.)

"... pass^k is appropriate for measuring overall task success, but it does not show which parts of a multi-intent task improve."

Response: Thanks for this insightful suggestion! We report finer-grained metrics:

  1. Per-check success: IEL and ReAct pass Search checks at almost the same rate (92.74% vs. 91.91%), so IEL's gain does not come from retrieving more information.
  2. Gain vs. number of goals: the pass^1 gain over ReAct is 0 (2 goals), +3.70 (3–4), and +6.25 points (6). The gain grows with the number of intents, where episode-level supervision is most ambiguous.
AuthorsOpenReview ↗

Response to W2 & Q2: MemP, and credit assignment vs. structure

W2 & Q2

"Could the authors explain why closely related procedural memory methods, particularly MemP, were not included ... whether the observed improvements are primarily attributable to intent-level credit assignment rather than structured experience organization and retrieval?"

Response: Thanks for this key question!

MemP. We added it (Table R2, in our response to Reviewer b93v, W1). IEL leads by 18.44 pass^1 and 32.24 pass^3 points. MemP indexes each episode by its opening request, so later intents rarely retrieve a memory (only 14.5% of test episodes do), and it discards partially failed episodes.

Credit assignment vs. structure. All variants below keep IEL's structure (intent reconstruction, taxonomy, intent-indexed retrieval, procedure format) and the same growth and evaluation protocol. They differ only in the outcome labels used for procedure induction:

  • IEL-EpisodeCredit: the verifier is replaced by the oracle benchmark reward. Every intent record in a successful episode is labeled POSITIVE, and every record in a failed episode NEGATIVE.
  • IEL-NoVerify: the verifier is removed, and every attributed intent record is treated as POSITIVE.
  • IEL-NoNeg: the verifier is kept, but NEGATIVE records are dropped, so procedures are induced from POSITIVE records only and contain no error warnings.

Table R1: ECom-Bench, DS-V4.1-Flash, mean over three folds

VariantOrganizationOutcome labelspass^1pass^2pass^3
IEL (full)intent-levelintent-level verified66.8858.0352.94
IEL-EpisodeCreditintent-levelepisode-level (oracle)56.6045.9139.62
IEL-NoVerifyintent-levelall positive52.8744.7039.65
IEL-NoNegintent-levelpositives only59.6950.8445.10
w/o Intent Decompositionepisode-levelepisode-level50.2238.8531.92
MemPstructured proceduralepisode-level48.4428.9420.70
ReAct (no memory)––59.3046.7039.76

With the same structure and even oracle episode rewards, IEL-EpisodeCredit loses 13.32 pass^3 points and only matches ReAct; treating every record as positive (IEL-NoVerify) gives the same result. Starting from w/o Intent Decomposition (31.92 pass^3), intent-level organization adds 7.70 points (→ 39.62), and intent-level credit assignment adds the remaining 13.32 (→ 52.94). Dropping negative records (IEL-NoNeg) costs 7.84 pass^3 points, so verified failures also contribute.

We thank the reviewer again for the positive assessment and are happy to provide further details.

Reviewer_b93vOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IEL to address the difficulty of using overall success or failure for detailed experience attribution in conversations with multiple intents. IEL uses intent instances across turns as its basic unit and follows the principle of “attribution before abstraction.” It reconstructs intent instances, links interaction evidence, and verifies outcomes. It extracts reusable procedures from successful instances and identifies common errors from failed instances. It also expands the intent categories as evidence grows. The paper evaluates the method on downstream tasks.

Critical strengths
  1. The paper highlights the limits of turn level supervision and proposes intent instances as the unit for experience attribution and reuse.
  2. IEL combines intent reconstruction, outcome verification, and procedure extraction. It also supports goals outside the existing intent categories.
  3. The writing is clear and easy to understand, and the paper is well structured.
Critical weaknesses
  1. The baselines are relatively old. The main baselines are methods from 2023, 2024, and 2025. One concern is whether the paper considers newer methods from 2026 as baselines.
  2. The evaluation mainly uses ECOM-BENCH and its extended version. The paper needs to include other benchmarks in this field to demonstrate that the method generalizes across different benchmarks.
  3. The paper lacks explanations for the baseline results. Some memory methods even perform worse than methods without memory. For example, with DS-V4.1-FLASH, AWM performs worse than ReAct. The paper needs to explain the reasons.
  4. The framework uses a large language model as the verifier, but the paper does not report how often its judgments agree with human labels or how often these judgments are incorrect.
Questions for authors

See above.

6 comments
AuthorsOpenReview ↗

Response to W1: Added the 2026 baseline MemP

W1

"The baselines are relatively old ... One concern is whether the paper considers newer methods from 2026 as baselines."

Response: Thanks for this valuable comment! We added MemP (Fang et al., ACL Findings 2026), the closest structured procedural-memory method, under the same growth tasks, executor prompt, and frozen evaluation as all other methods.

Table R2: ECom-Bench, DS-V4.1-Flash, mean ± std over three folds

Methodpass^1pass^2pass^3
ReAct59.30 ± 14.5046.70 ± 16.4339.76 ± 15.26
Trajectory-RAG37.91 ± 9.6820.26 ± 7.2511.44 ± 6.05
ExpeL54.25 ± 8.2439.76 ± 6.3930.28 ± 4.34
AWM34.75 ± 7.3221.53 ± 6.8815.25 ± 7.17
MemP48.44 ± 1.9228.94 ± 1.0120.70 ± 2.64
IEL66.88 ± 14.9558.03 ± 16.6052.94 ± 17.43

IEL leads MemP by 18.44 pass^1 and 32.24 pass^3 points; we explain MemP's result in our response to W3.

AuthorsOpenReview ↗

Response to W2: New benchmark, τ²-bench retail

W2

"The evaluation mainly uses ECOM-BENCH and its extended version. The paper needs to include other benchmarks in this field ..."

Response: Thanks for this helpful suggestion! We added τ²-bench retail v1.0.1 (Barres et al., 2025). It fits IEL's setting: all tasks share one tool interface, and 20 of the 40 test tasks (50%) contain ≥2 intents (12 contain ≥2 write intents). Experience is grown on the 74 official train tasks and evaluated, frozen, on the 40 official test tasks, with DS-V4.1-Flash, 3 trials per task, and τ²'s official scoring.

Table R4: τ²-bench retail v1.0.1, 40 test tasks, DS-V4.1-Flash

Methodpass^1pass^2pass^3
ReAct81.6775.0070.00
Trajectory-RAG88.3383.3380.00
ExpeL89.1783.3377.50
AWM87.5082.5080.00
MemP87.5081.6777.50
IEL91.6786.6782.50

IEL is best on all three metrics: +10.00 / 11.67 / 12.50 over ReAct and +2.50 / 3.34 / 2.50 over the strongest baseline. Mixed outcomes also occur here: during growth, 11 verified successes come from failed episodes, which episode-level selection would discard.

AuthorsOpenReview ↗

Response to W3: Why some memory baselines fall below ReAct

W3

"The paper lacks explanations for the baseline results. Some memory methods even perform worse than methods without memory. For example, with DS-V4.1-FLASH, AWM performs worse than ReAct."

Response: Thanks for pointing this out! This is not specific to AWM: with DS-V4.1-Flash on ECom-Bench, all episode-level baselines fall below ReAct (59.30 pass^1): Trajectory-RAG 37.91, AWM 34.75, ExpeL 54.25, MemP 48.44. The same baselines match or exceed ReAct with DS-V3.2 and GLM-5.3-Flash, and on τ²-bench retail with DS-V4.1-Flash (Table R4). IEL is the only method above ReAct in every setting.

Why MemP gains little. Over its 159 test episodes, two design properties explain the gap:

  1. Episode-level index. Each memory is indexed by the episode's opening request, so intents raised later in a dialogue cannot retrieve it. Only 23 of 159 episodes (14.5%) ever retrieve a memory; the rest effectively run as ReAct.
  2. Episode-level filtering. MemP keeps only fully successful episodes, so correctly handled intents in failed episodes are lost (13–16 memories per fold).

IEL indexes and verifies experience per intent and avoids both.

Reviewer_voTMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduce IEL, which focuses on intent-level experience reusing. IEL treats a cross-turn intent instance as the unit of experience attribution and procedural learning.

Critical strengths

Overall, I believe the paper’s main strength lies in its clear introduction of the research problem, although I remain unconvinced of its significance and have reservations about the proposed approach.

Critical weaknesses
  1. Importance of the problem and significance of the novelty. The paper is motivated by the idea that experience should be reused at the level of specific intents rather than entire trajectories. While this is intuitive, prior work such as Agent Workflow Memory and Generalizing Experience for Language Agents with Hierarchical MetaFlows already reuses experience at the subgoal or subtask level. It is therefore unclear how substantially the proposed notion of “intent” differs from these existing units of experience reuse. In my view, the contribution appears incremental.

  2. Necessity of the elaborate pipeline. Section 3.3 introduces a relatively complex mechanism for intent identification, verification, and resolution, but the paper does not provide experiments establishing the necessity of these design choices. Given the strong capabilities of current LLMs, are these predefined workflow steps necessary? Would this pipeline outperform directly prompting an LLM to sequentially produce intent instances, supporting evidence, outcome labels, and reusable procedures? A controlled comparison against this simpler alternative would help justify the proposed design.

  3. Limited scale and diversity of the evaluation. The in-domain evaluation uses the 53-task ECom-Bench, while ECom-Bench-XCat is a 60-task extension. Neither benchmark appears to be widely used, and both are limited in scale and diversity. In my view, evaluation on these benchmarks alone provides insufficient evidence for the broader effectiveness and generalizability of the proposed method.

Questions for authors

See above.

5 comments
AuthorsOpenReview ↗

Response to W1: Difference from AWM and Hierarchical MetaFlows

W1

"... prior work such as Agent Workflow Memory and Generalizing Experience for Language Agents with Hierarchical MetaFlows already reuses experience at the subgoal or subtask level. It is therefore unclear how substantially the proposed notion of "intent" differs from these existing units of experience reuse ..."

Response: Thanks for raising this important point! AWM and MetaFlow reuse experience at the sub-task level, but both learn it from whole-task success: AWM induces workflows from trajectories judged successful per task, and MetaFlow (Fan et al., NeurIPS 2025) builds its tree from fully solved training tasks and keeps a merged node only if it passes all task-level test cases. A sub-routine from a successful episode can still encode a failed sub-goal, and successful sub-goals in failed episodes are discarded. IEL instead assigns outcomes to cross-turn intent instances before abstraction; in customer service these instances are interleaved and non-contiguous (Fig. 1). IEL also never uses the benchmark reward (Appendix B.1): its verifier judges each intent from the dialogue and tool evidence, whereas MetaFlow requires task-level test cases or golden solutions, which real customer-service logs do not provide. The difference is measurable:

  1. 54.72% of growth episodes are mixed-outcome, and 32.46% of IEL's verified positives come from failed episodes.
  2. With IEL's own structure fixed, replacing intent-level labels with oracle episode labels lowers pass^3 by 13.32 points, to the level of ReAct (Table R1, in our response to Reviewer Mj9Y, Q2). Sub-task representation alone is not enough; supervision granularity matters.

We will make this distinction, and the comparison with MetaFlow, explicit in the Introduction and Related Work.

AuthorsOpenReview ↗

Response to W2: Comparison with a single-prompt alternative

W2

"... Would this pipeline outperform directly prompting an LLM to sequentially produce intent instances, supporting evidence, outcome labels, and reusable procedures? A controlled comparison against this simpler alternative would help justify the proposed design."

Response: Thanks for this constructive suggestion! We implemented IEL-OneShot: one LLM call per episode outputs intent instances, evidence, labels, and procedures, stored and retrieved as in IEL.

Table R3: ECom-Bench, DS-V4.1-Flash, mean ± std over three folds

Variantpass^1pass^2pass^3
ReAct (no memory)59.30 ± 14.5046.70 ± 16.4339.76 ± 15.26
IEL-OneShot44.84 ± 12.0625.85 ± 8.0216.88 ± 9.44
IEL66.88 ± 14.9558.03 ± 16.6052.94 ± 17.43

A single prompt is not enough: IEL-OneShot is 22.04 pass^1 and 36.06 pass^3 points below IEL, and even falls below ReAct. The staged design enforces constraints that a single generation does not reliably respect: exclusive attribution, an explicit UNCERTAIN option, and induction from verified records only. In exchange, IEL's intent pipeline makes 12.07 LLM calls per growth episode on average.

AuthorsOpenReview ↗

Response to W3: Additional benchmark, τ²-bench retail

W3

"... The in-domain evaluation uses the 53-task ECom-Bench, while ECom-Bench-XCat is a 60-task extension. Neither benchmark appears to be widely used, and both are limited in scale and diversity ..."

Response: Thanks for this helpful comment! We added τ²-bench retail (74 train / 40 test tasks; Table R4, in our response to Reviewer b93v, W2). IEL is best among six methods (91.67 / 86.67 / 82.50 pass^1/2/3), +10.00 / 11.67 / 12.50 over ReAct. The evaluation now covers 153 test tasks from two benchmark suites (53 ECom-Bench, 60 XCat, 40 τ²-bench retail), with three executors on ECom-Bench.

We hope these results clarify the contribution and would be grateful if the reviewer would reconsider the assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of New Experiments and Clarifications

We sincerely thank all reviewers for their thorough and constructive reviews! We appreciate the following encouraging comments: "it offers a reasonable way to preserve these useful parts by using cross-turn intent instances as the unit of experience ... this gives the method a clear motivation and practical value for service agents" [Mj9Y]; "The writing is clear and easy to understand, and the paper is well structured" [b93v]; "the paper's main strength lies in its clear introduction of the research problem" [voTM].

The main concerns were: (i) whether the gains come from intent-level credit assignment or from structured organization (Mj9Y, voTM); (ii) recent baselines (Mj9Y, b93v); (iii) evaluation scale and diversity (b93v, voTM); and (iv) verifier reliability and XCat documentation (Mj9Y, b93v). We made the following additions and will include all of them in the final version:

  1. Credit assignment vs. structure (Table R1). Keeping IEL's structure fixed and replacing only its intent-level labels with the oracle episode reward lowers pass^3 from 52.94 to 39.62, the level of ReAct (39.76). The gain therefore requires intent-level credit assignment.
  2. Mixed outcomes are common. 54.72% of growth episodes contain both successful and failed intents, and 32.46% of IEL's verified positives come from failed episodes, which episode-level methods discard.
  3. New baseline: MemP (Table R2). On ECom-Bench with DS-V4.1-Flash, MemP reaches 48.44 pass^1 / 20.70 pass^3, versus 66.88 / 52.94 for IEL.
  4. Single-prompt alternative (Table R3). Producing intents, labels, and procedures with one LLM call per episode (IEL-OneShot) reaches 44.84 pass^1 and 16.88 pass^3, versus 66.88 and 52.94 for IEL, and falls below ReAct (59.30 and 39.76).
  5. New benchmark: τ²-bench retail (Table R4). On the official test tasks, IEL is best of six methods: 91.67 / 86.67 / 82.50 pass^1/2/3, versus 81.67 / 75.00 / 70.00 for ReAct and at most 89.17 / 83.33 / 80.00 for the memory baselines.
  6. Verifier reliability. On 200 randomly sampled intent records, the verifier agrees with human labels on 87% (96% between two annotators), and its POSITIVE precision is 92%.
  7. Finer-grained metrics (Mj9Y, W1). The pass^1 gain of IEL over ReAct grows with the number of goals: 0, 3.70, and 6.25 points for 2, 3–4, and 6 goals.
  8. XCat documentation (Mj9Y, Q1). We describe how XCat tasks and checks were built and validated, and will add this description to Appendix A.

In the following, we provide detailed responses to each reviewer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.