acceptodds
Under review as a conference paper at ICLR 2027

ACE-Trader: Adversarial Co-Evolution for Generator–Critic Trading Systems

Abstract

Generator–critic trading systems pair an LLM-based trader that makes trading decisions with a red team agent that evaluates these decisions and provides feedback for revision. However, improving individual trading decisions does not ensure that both agents learn from experience and improve together over time. This raises an important question: how can the trader and red-Team select useful experience and learn from each other's evolving behavior? On one hand, experience that benefits decisions under one market condition may become ineffective or harmful under another. On the other hand, trading outcomes reflect both agents' actions, making it difficult to derive appropriate experience for each role. In this work, we propose ACE-Trader, an adversarial co-evolution trading framework. Specifically, we first develop separate context-dependent memory selectors that learn to identify useful experience for the current trading decision. Both selectors receive shared feedback from net-return differences between complete memory-assisted and memory-free trading pipelines, with online updates emphasizing recent outcomes. We further design a co-evolution mechanism that links role-specific memories through shared trading records. We use these records to evaluate the Red-Team's risk assessments and the effects of the Trader's revisions against realized market outcomes, then derive experience for their respective memories. By linking this experience to the corresponding records, we enable the Trader to improve its proposals based on prior critiques and the Red-Team to adjust its review focus as the Trader's behavior changes. Experiments on U.S. stock trading show that ACE-Trader outperforms all 22 active baselines in cumulative return and achieves the highest Sharpe ratio and lowest maximum drawdown across both active methods and passive benchmarks. Code is available at https://anonymous.4open.science/r/ACE-Trader-B1BC for anonymous review.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
AI4Trading, Agent system

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_id7SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ACE-Trader, a memory-adaptive trading framework in which an LLM-based Trader and Red-Team interact through a Commit-Challenge-Revise workflow. Its central idea is to improve collaboration across trading episodes through context-dependent memory selection and linked, role-specific experience rather than by fine-tuning the underlying LLMs. Separate lightweight selectors rank candidate memories using semantic, market, portfolio, and historical-outcome features, with an ensemble-disagreement bonus encouraging exploration. The selectors receive shared binary supervision from the net-return difference between complete memory-assisted and memory-free workflows evaluated from the same account state. After settlement, the framework separately evaluates whether the Red-Team’s warnings were correct and whether the Trader's revisions improved returns, then stores role-specific lessons linked through shared interaction records. Experiments cover 33 equities across 11 sectors, using 2021-2023 historical training and warm-up data followed by a January-June 2026 test period. The reported results seem promising.

Critical strengths
  1. The distinction between producing a better revised portfolio and learning useful experience for future interactions is well motivated.

  2. ACE-Trader evaluates pre-specified risk conditions independently from the net-return difference between the original and revised portfolios. This avoids automatically crediting every warning associated with a profitable trade or treating every risk-reducing revision as beneficial.

Critical weaknesses
  1. The shared supervision signal leaves role-specific credit assignment unresolved. Both selectors receive the same binary label derived from the return difference between the complete Memory-On and Memory-Off workflows, which the paper explicitly acknowledges does not measure an individual memory's contribution. A beneficial memory selected by one agent could therefore be penalized because the other agent’s selected memory harms the joint outcome, or vice versa. Although risk-warning correctness and revision benefit are evaluated separately, the paper does not establish whether the shared training signal reliably identifies useful memories for each role. This weakens the evidence for the claimed role-specific memory-selection capability.

  2. Memory management and first-stage retrieval are insufficiently specified. Although the paper specifies a 24-record short-term memory capacity and consolidation after every ten new records, it does not adequately describe eviction policies or how long-term and shared memories are maintained as they grow. The construction of the ten-candidate shortlist is also unclear, particularly how entries from different memory sources are retrieved, combined, and initially ranked.

  3. The framework accumulates memories and updates selectors during historical warm-up, then continues both processes during testing. These test-time updates use only completed settlement outcomes. However, the reported ablations do not distinguish the benefits of historical initialization, newly accumulated experience, and continued selector learning. It remains unclear whether the gains primarily reflect effective historical memories, adaptation through changing memory content, or online optimization of the selectors.

  4. The evaluation covers one six-month test interval and one fixed 33-stock universe, with a single Trader Red-Team backbone configuration. The main results are reported as point estimates without run-to-run variability or confidence intervals. Thus, the stability of the improvements under different selector initializations, any stochastic LLM generation, and alternative market trajectories remains uncertain.

  5. The shared memory stores historical proposals, critiques, revisions, and outcomes, making retrieval of richer past cases a plausible alternative explanation for some of the gains. In the ablation study, disabling co-evolution removes both shared-case retrieval and the construction of memories linking the two roles. The existing results therefore cannot distinguish reciprocal adaptation from the benefit of additional historical information. A matched case-bank-augmented single-agent or financial-LLM baseline, together with a two-agent control that retains shared-case access but removes cross-role memory construction, might be needed to substantiate the specific contribution of co-evolution.

Questions for authors
  1. How are outdated entries evicted, and how are long-term and shared memories managed as they grow? How are candidates retrieved and ranked across different memory sources?

  2. What justifies using the same pipeline-level binary label to supervise both role-specific selectors?

  3. Could the authors include controlled experiments that independently freeze memory updates and selector parameters during testing to disentangle these contributions? How many independent runs underlie the main results, and what uncertainty surrounds the improvements over the strongest active baseline and the key ablation variants?

  4. Does ACE-Trader retain its advantage over a single-agent trader or finance-specific LLM given equivalent historical case information under matched information cutoffs, retrieval budgets, and portfolio interfaces?

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

This paper proposes ACE-Trader, an LLM-based financial trading framework in which a Trader and a Red-Team jointly adapt through memory. The method has two main ideas. First, separate context-dependent memory selectors choose experience for the Trader and Red-Team based on the current decision context. The selectors are updated online using feedback derived from the return difference between memory-assisted and memory-free trading pipelines. Second, a memory co-evolution mechanism records the Trader's proposal, the Red-Team's critique, the Trader's response, and realized outcomes in shared cases, from which role-specific experiences are constructed.

The resulting system follows a Commit–Challenge–Revise workflow: the Trader proposes a portfolio, the Red-Team examines the proposal against current evidence, and the Trader either revises or retains it. The paper evaluates the method on a 33-stock U.S. equity universe and compares against 22 active baselines and two passive benchmarks. ACE-Trader reports the highest cumulative return among active methods and the highest Sharpe ratio and lowest maximum drawdown among all evaluated methods. Ablations are provided for the Trader selector, Red-Team selector, and memory co-evolution mechanism.

Critical strengths
  1. The paper addresses an interesting credit-assignment problem in generator–critic trading systems. In particular, separating whether the Red-Team correctly identified a risk from whether the Trader's subsequent revision actually improved returns is a meaningful design choice. Linking proposals, critiques, revisions, and realized outcomes through shared cases also provides a concrete mechanism through which the two roles can adapt based on their interactions rather than independently reflecting on overall trading outcomes.
Critical weaknesses
  1. The empirical validation is too narrow to establish the robustness of the trading-performance claims. The main evidence is based on a single fixed universe of 33 U.S. equities and a single test period from January to June 2026. For a trading method whose central motivation is adaptation to changing market conditions, it is important to know whether the gains persist across substantially different market regimes, stock universes, and evaluation periods. This is especially consequential because the headline differences are obtained from one realized market trajectory rather than repeated independent test instances. I therefore find it difficult to determine whether the reported advantage is a robust consequence of the proposed memory mechanisms or is sensitive to this particular market period/universe.

  2. The selector supervision provides only workflow-level credit for individual memory selections. Both the Trader and Red-Team selectors receive the same binary label based on whether the complete memory-assisted pipeline outperforms a memory-free pipeline. However, several memory selections can occur within one step, and the paper explicitly notes that this shared label does not measure the separate contribution of an individual memory. A profitable outcome can therefore positively reinforce a selected memory even when another memory or another component of the workflow was responsible for the improvement. This seems particularly important because the paper's core methodological claim is that the selectors learn which experience is useful for the current decision. The current supervision demonstrates whether memory use was beneficial collectively, but does not directly establish the usefulness of each selected memory.

Questions for authors
  1. Robustness of the trading results: How sensitive are ACE-Trader's gains to the particular stock universe and January–June 2026 evaluation period? Do the authors have results for additional non-overlapping test periods, market regimes, or independently constructed stock universes? Evidence that the improvement persists across such settings would materially increase my confidence in the empirical conclusions.
Reviewer_sKiBOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds an LLM trading system. A Trader proposes trades each day and a Red Team challenges them. Small networks pick one past memory for each role. They are trained with a counterfactual label. The label compares the returns of a run with memory and a run without memory. The system also records whether each warning came true and whether each revision earned more. The test uses 33 US stocks in the first half of 2026.

Critical strengths
  1. The counterfactual label is a novel idea to me. Both runs happen on the same day and start from the same account. They also see the same market. So the market factors of that day cancel out. This is the main step beyond Memento.
  2. Judging warnings and revisions apart is a sensible design. Whether a warning came true and whether a revision earned more are checked on their own. So credit or blame does not go to the wrong role.
Critical weaknesses
  1. The way memories are selected is too crude. The label is matched only within one day. But the selector uses it to decide which memory to use on later days. The paper itself says the same memory may help in one market and hurt in another. Yet each selector update looks only at the last five days of labels. It weights them by how recent they are, not by how similar the market is. I suggest making the label finer, for example per sector or per trading technique. It should also be matched with days of similar market conditions.
  2. The claim of coevolution lacks evidence. The paper says both roles improve together over time. But it does not show any change over time. Table 2 only compares the final return of one run. Also, both selectors are trained on the same label, so they share one goal. The adversarial part exists only in the prompts. I suggest showing how the behavior of both roles changes over time.
  3. The results show no statistical significance. The paper has only one run of about six months. It reports no error bars and no significance test. Daily returns swing a lot. In such a short single run, the gap to the baselines may come from luck. I suggest running paper trading experiments over more time periods and with different random seeds.
Questions for authors
  1. Why does training use only the last five days, weighted by recency? Could you match days by market similarity instead, and compute the label per sector or per trading technique?
  2. Could you show how the behavior of both roles changes over time, such as the warning hit rate and how often the Trader accepts a challenge?
  3. Could you add paper trading results over several time periods and several random seeds, with a significance test?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.