acceptodds
Under review as a conference paper at ICLR 2027

Beyond Successful Experience Reuse: Inducing Revision Knowledge from Evolving Experience for LLM Agents

Abstract

Experience-based learning enables LLM agents to accumulate and reuse knowledge from prior interactions, but previously successful procedures may become inapplicable as contexts evolve. We study revision knowledge induction: recovering structured knowledge about when acquired procedures remain valid or require revision. We introduce Contrastive Evolution Attribution (CEA), which factorizes evolutionary experience through structured contrasts to identify revision-relevant contextual factors and recover the corresponding state-to-procedure mapping. To evaluate this capability, we develop ApplyBoundary, a controlled diagnostic benchmark for revision knowledge recovery and downstream procedure selection under unseen contexts with correlated contextual changes. Experiments show that CEA improves both revision-factor identification and state-to-procedure mapping recovery over experience-based approaches that use the same evolutionary evidence without contrastive factorization. A small-scale validation on documented real-world system changes provides preliminary evidence that similar revision patterns can also appear in practical settings. Further analysis shows that controlled-success evidence helps identify revision-relevant changes, while effectively utilizing recovered knowledge for future decisions remains a separate challenge.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Revision Knowledge Induction, Procedure Applicability, Experience-Based Learning

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

This paper studies how agents can learn when a previously successful procedure remains applicable and when it needs to be replaced as the environment changes. The proposed CEA organizes contrasts between successful, failed, controlled-success, and recovery executions to identify the relevant contextual factor and recover its state-to-procedure mapping. Experiments on APPLYBOUNDARY, three model backbones, and three documented system changes show better knowledge recovery, but no downstream selection improvement over the strongest control.

Critical strengths

The paper addresses an important issue in experience reuse: past success does not establish applicability after the environment changes. Separating factor identification from procedure-mapping recovery provides a useful framework for evaluating what an agent actually learned, and this is the contribution I find most compelling.

Critical weaknesses

Prior work such as AutoGuide already learns state-associated guidelines from successful and failed experiences, so the paper needs to establish more clearly what is added by resolving correlated distractors and recovering explicit mappings. The appendix also lacks the complete paired prompts for CEA and B4, leaving both the main improvement and its source insufficiently transparent.

Questions for authors
  1. Compared with state-conditioned guidelines such as AutoGuide, what capability or operation does CEA add when both approaches receive the same controlled evidence?
  2. Could you provide the complete CEA and B4 inputs for the same task and identify every difference beyond the contrastive relations?
  3. Do the downstream findings hold without the Explanation field, and where do selection errors occur when both the factor and mapping are recovered correctly? Isolating the explanation field and locating selection errors under correctly recovered knowledge would make the claimed distinction between recovery and utilization much more convincing.
Reviewer_5sDEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how experience-based agents can determine whether previously acquired procedures remain applicable after contextual changes. The authors define revision knowledge as two components: the contextual factor governing procedure applicability and the mapping from that factor’s states to the applicable procedures. The proposed Contrastive Evolution Attribution (CEA) organizes evolving experience into validity, controlled, and recovery contrasts. A frozen LLM induces structured revision knowledge from these contrasts, which is then used to select procedures in unseen contexts. The authors also introduce APPLYBOUNDARY, a controlled diagnostic benchmark with known candidate factors, binary states, and two previously acquired procedures. The paper reports that, compared with the same-evidence B4 control, CEA improves factor-identification accuracy from 87.50% to 94.07% and mapping-recovery accuracy from 53.33% to 62.71%. However, downstream selection accuracy does not improve significantly: CEA and B4 achieve 84.38% and 84.72%, respectively. Additional backbone evaluations report similar recovery gains, and three documented real-world system changes provide preliminary validation.

Critical strengths
  1. Experience-based agents need to accumulate successful procedures and understand the conditions under which those procedures remain valid. The paper creates attribution ambiguity through correlated contextual changes and uses boundary–distractor disagreement queries to test whether the agent relies on the actual applicability factor. This makes learning procedure-applicability boundaries an independently evaluable problem and is the paper’s most valuable contribution.
  2. B4 and CEA use the same experience, ordering, task information, and output format, allowing the comparison to focus on contrastive organization. Recovery gains are also reported on additional models. The authors explicitly report the absence of a significant downstream gain and examine the selection interface using oracle knowledge.
Critical weaknesses
  1. CEA’s additional contribution over general contrastive induction methods has not been sufficiently isolated. The current method primarily supplies a frozen LLM with three predefined contrastive relations. At the same time, each task already provides the candidate factors, two procedures, and four observations containing both controlled success and successful recovery; the crucial attribution evidence has therefore already been organized in the input.The CEA–B4 comparison supports the claim that explicitly prompting these contrasts helps recover knowledge. However, the baselines mainly consist of author-designed experience-organization strategies. The paper lacks comparisons that could further isolate the method’s contribution, such as general step-by-step comparison, elimination of mappings inconsistent with execution outcomes, or simple constraint-based induction using the same evidence. It therefore remains unclear how much of the gain comes from CEA-specific organizational principles and whether it extends beyond explicit solution guidance for the controlled reasoning task. The frozen evaluation contains only 24 independent task programs, and the three documented cases use a similar structure, providing limited additional evidence on this question.
  2. The utilization diagnostics do not fully separate induction errors from selection errors. CEA and B4 obtain similar downstream accuracy in Table 5—84.38% and 84.72%, respectively—with a relatively wide confidence interval for the paired difference. Interpreting this result requires distinguishing errors in the induced knowledge from errors introduced by the selection step. Table 6 evaluates LLM-based selection with ground-truth knowledge and direct decoding of the ground-truth mapping. It does not report direct decoding of the knowledge actually induced by CEA. Applying the same decoder to the knowledge induced by CEA and B4 would help determine whether the recovery gains translate into more accurate decisions when LLM-based selection is removed. The explanation field is also passed to the downstream selector but excluded from the recovery metrics. An explanation ablation would clarify whether selection follows the structured factor and mapping or relies on information in the rationale.
Questions for authors
  1. Does CEA retain an advantage over general step-by-step comparison or consistency-constrained induction when evidence, output format, and inference budget are held comparable? Please clarify which results specifically support the contribution of CEA’s organizational principles.
  2. What downstream accuracy is obtained by directly decoding CEA’s induced factor and mapping? How does the difference between CEA and B4 change when the explanation is removed and only the structured fields are retained? Can these results distinguish knowledge-induction errors, context-state identification errors, and procedure-selection errors?
  3. How are contextual differences between e₂ and e₃ constrained in the current tasks? How many candidate factors does each task contain, and what conditions ensure that the four observations uniquely identify the revision factor and mapping? Please describe how these conditions vary across task programs.
Reviewer_U1wsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how LLM agents infer when existing procedures remain applicable as environments change. It proposes Contrastive Evolution Attribution (CEA), which organizes evolutionary experiences into structured contrasts to identify relevant contextual factors and recover state-to-procedure mappings. Experiments on the proposed APPLYBOUNDARY benchmark show improved knowledge recovery over a same-evidence baseline, but no improvement in downstream procedure selection.

Critical strengths
  1. The paper considers an interesting scenario in which previously acquired experience may become inapplicable as environments evolve.
  2. Relevant controls and ablations. The same-evidence baseline and evidence-removal experiments help assess the contribution of contrastive organization.
Critical weaknesses

1. Insufficient dataset detail. The paper lists the task domains but does not provide concrete benchmark cases. Several key details remain unclear, including the number of candidate factors (mm), whether these factors are manually constructed, and whether different tasks share the same candidate factors. Since the main experiments rely on the constructed APPLYBOUNDARY benchmark, the lack of concrete examples makes the evaluation abstract and difficult to assess.

2. Limited practical applicability. The method substantially simplifies real-world conditions by assuming access to comparable successful and failed experiences, predefined candidate factors, and registered procedures. These inputs can be difficult to obtain in practice. The method’s strong dependence on such prepared inputs therefore raises concerns about its applicability to real-world settings.

3. Confounded knowledge-utilization analysis. The downstream selector receives a natural-language explanation in addition to the factor and mapping, whereas the recovery metrics evaluate only the latter two. Without separating these inputs or directly decoding the predicted mappings, the experiments cannot establish whether the observed discrepancy reflects a knowledge-utilization bottleneck or a mismatch between the information evaluated and the information available to the selector.

Questions for authors

Please see the weaknesses above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.