acceptodds
Under review as a conference paper at ICLR 2027

SETA: Amortizing Test-Time Compute with Explicit-Failure Memory for Tool-Using Agents

Abstract

Agents can fail a task and still judge themselves successful. Success-only memory discards this signal, and K-attempt self-consistency multiplies compute. SETA (Selective Explicit-failure Trajectory Amortization) turns externally verified failures into reusable memory. Offline, we collect four exploratory trajectories per task and flag false successes. We deterministically rewrite each: the error path stays, only the outcome changes, with no model call. At deployment, a single attempt retrieves this memory: a success skeleton, known-pitfall exemplars, and one explicit-failure memory. For example, a wrongly approved leave request becomes a note the agent reads when the same request returns. We evaluate 60 requesting-time-off tasks and 50 customer-request-routing tasks from AgentArch, assuming mature memory for recurring requests; cold-start and cross-task transfer are out of scope. Under this assumption, a single Think-mode attempt reaches 85.00% final-outcome success on time-off tasks. On a separate seed set, removing only the explicit-failure memory from otherwise identical retrieval lowers final-outcome success by 16.00 points. A corrected bf16 harness reproduces this effect at 15.67 points; a second stack shows 9.00 points, and retrieval alone gives no gain on either. In Think mode, SETA exceeds a workflow-memory baseline, an LLM-written lesson-memory baseline, and four-attempt majority voting without memory. SETA and the workflow-memory baseline pay the same strict-task-success cost, but only SETA converts it into a gain; under non-thinking (NonThink) deployment, both are indistinguishable from no memory. On the longer routing tasks, memory injected once has no effect. Restating explicit-failure memory each call yields a 19.20-point gain; restating retrieval-only memory yields none. 5

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, tool-using agents, agent memory, test-time compute, failure-aware memory, false success, retrieval-augmented agents, compute amortization

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_j4AFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SETA, a failure-aware memory framework for tool-using agents. It converts externally verified false-success trajectories into explicit-failure memories through deterministic rewriting and reuses them during single-attempt inference. Experiments on two AgentArch workflows show improvements in final-outcome correctness under Think mode and examine the effect of memory placement in long-horizon tasks.

Critical strengths
  1. Practical Motivation. The paper addresses false-success cases in tool-using agents, where agents incorrectly report successful task completion. The ablation studies suggest that task-specific failure diagnostics can improve final-outcome correctness in the evaluated settings.
  2. Memory Placement Analysis. The paper examines different memory injection strategies for long-horizon workflows. The results show that repeated memory injection can be more effective than single injection in the evaluated setting.
Critical weaknesses
  1. Limited Novelty and Generalization. The methodological distinction from existing failure-aware memory approaches is not sufficiently established. Moreover, the main evaluation allows retrieving previously verified correct outcomes for the same tasks, making it difficult to distinguish the benefits of failure-memory construction from direct answer reuse. The evaluation is also limited to two workflows, with weak cross-family and cross-benchmark results. (Presentation issues: The abstract is divided into two paragraphs, which does not comply with the ICLR formatting requirements, and Fig. 1 extends beyond the prescribed text width. The authors are encouraged to further improve the overall presentation of the manuscript.)
  2. Concerns about Evaluation Validity and Task Reliability. The reported improvements mainly concern final-outcome correctness rather than actual task completion. Strict task success decreases substantially on RTO and remains very low on CRR, raising concerns about practical effectiveness. Additionally, the deterministic rewrite may introduce inconsistencies between declared outcomes and preserved tool actions (Appendix G.10). The K=4 majority-voting baseline also aggregates externally judged correctness labels rather than model-generated answers, making its comparison with standard self-consistency questionable.
Questions for authors
  1. How does SETA compare against a simpler baseline that directly retrieves previously verified correct outcomes for the same tasks?
  2. How do the authors explain the decline in strict task success, and how is consistency between rewritten memories and actual tool execution ensured?
  3. Could the authors clarify why the K=4 majority-voting baseline aggregates correctness labels rather than model-generated answers?
Reviewer_3qUAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies cases where an agent fails at a task but judges its own performance to be successful. The authors propose SETA, which records these cases in an explicit-failure (EF) memory for use during subsequent inference. Experiment shows improvements in final-outcome success.

Critical strengths
  1. This paper addresses an interesting question for agent, and proposes framework that may improve performance.
Critical weaknesses
  1. The figures of this paper are not well-made, particularly in Figures 1 and 2, there are many instances where the text physically overlaps.

  2. As this framework uses offline memory construction, the time required for the overall process should be noticed, but missing.

  3. It is still difficult to conclude that the authors' judgment is correct that SETA makes better performance by FS. Authors need to show that using FS is the key more precisely.

Questions for authors
  1. How much time and memory does the method use in the offline and online settings, especially for offline memory construction?

  2. What do the labels L1 and L2 refer to in Section 6?

  3. Could you show the results when items other than FS are stored in memory? (such as FF to clarify that the improvement of this method comes from FS)?

1 comment

Thank you for clarifying questions. However, I still have concerns about using four-attempt voting as the reference point for the resource effieciency claim. Response doesn't explain why four-attempt voting can be a fair comparison with SETA. As author said, SETA itself uses more tokens and calls compare with one attempt. Thus, I will take the clarifications and additional results into account in my final assessment.

Reviewer_q2oXOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a trick to solve the issue where agent gives wrong answer but still think it is correct or what people usually call it as hallucination. The method first do offline inference that collects four exploratory trajectories per task and flags false successes. Then, it rewrites the false positive trajectory to set it as failed and attach the reason why it fails. The rewritten trajectories are stored in a memory block. In online inference time, the system retrieves the memory and add the relevant explicit false trajectory into the system prompt to remind the agent with the past false success experience in order to improve the performance. The experiments are conducted on 60 requesting-time-off tasks and 50 customer-request-routing tasks from AgentArch showing 16 points improvements.

Critical strengths
  1. The paper studies an important research problem "Agent Hallucination". The agent hallucination is of great importance as it can dangerous when it is applied in important system such as company-wide file system, etc. Reducing agent hallucination is one of the most critical problems that the LLM research community needs to focus. This paper falls into this area and helps raise the attention for the researchers to study the problem.
  2. The proposed method is effective which improves 16 points after retrieving the explicit false trajectory to help the agent better learn from the past false success experience.
Critical weaknesses
  1. The experiments are not solid and concerned. The studied benchmark or dataset is too small, which contains only 110 examples in total. Among them, there are only 50
  2. It is concerned whether the improvements are from the method itself or from that the agent already sees the answer. The system needs to retrieve similar pattern trajectories. What if the agent is applied to a new task? The current system needs to rewrite the false success trajectories and provides it to the inference time, which sounds like providing the answer to inference. In view of this, the system design is questionable.
  3. The paper writing is so hard to follow, and many sentences are not professional. For example, "it recovers the Figure 1 task on 4 of 5 seeds, not 5 of 5".
  4. The proposed method does not involve model training and cannot actually improve the agent's capability from the fundamental layer. The proposed method sounds more like a trick which collects some trajectories and provide it to correct agent's behavior. That's good. But for an ICLR research paper, it feels trivial.
Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.